]> git.apps.os.sepia.ceph.com Git - ceph.git/commitdiff
Merge commit '5c612d7e32a07a01b3b97f9cf58f2a3cef0e2e73'
authorJ. Eric Ivancich <ivancich@redhat.com>
Tue, 6 Jun 2017 19:01:03 +0000 (15:01 -0400)
committerJ. Eric Ivancich <ivancich@redhat.com>
Tue, 6 Jun 2017 19:01:03 +0000 (15:01 -0400)
Signed-off-by: J. Eric Ivancich <ivancich@redhat.com>
13 files changed:
1  2 
src/dmclock/CMakeLists.txt
src/dmclock/sim/CMakeLists.txt
src/dmclock/sim/src/sim_server.h
src/dmclock/sim/src/simulate.h
src/dmclock/sim/src/test_dmclock_main.cc
src/dmclock/src/CMakeLists.txt
src/dmclock/src/dmclock_server.h
src/dmclock/support/src/indirect_intrusive_heap.h
src/dmclock/support/src/run_every.cc
src/dmclock/support/src/run_every.h
src/dmclock/support/test/CMakeLists.txt
src/dmclock/test/CMakeLists.txt
src/dmclock/test/test_test_client.cc

index 428863dc496df44f3adb733b75e5af564a25dd43,0000000000000000000000000000000000000000..9444ea1e96afd0a27250d6d989a0b41f797cdcd2
mode 100644,000000..100644
--- /dev/null
@@@ -1,32 -1,0 +1,25 @@@
- set(CMAKE_CXX_FLAGS "-std=c++11 -Wno-write-strings ${CMAKE_CXX_FLAGS}")
 +cmake_minimum_required(VERSION 2.8.11)
 +
- set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake/modules/")
- if(DO_NOT_DELAY_TAG_CALC)
-   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDO_NOT_DELAY_TAG_CALC")
- endif()
- if(K_WAY_HEAP)
-   if(K_WAY_HEAP LESS 2)
-     message(FATAL_ERROR "K_WAY_HEAP value should be at least 2")
-   else()
-     set(CMAKE_CXX_SIM_FLAGS "-DK_WAY_HEAP=${K_WAY_HEAP}")
++list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake/modules")
 +
- if (NOT(TARGET gtest AND TARGET gtest_main))
-   find_package(gtest REQUIRED)
-   include_directories(${GTEST_INCLUDE_DIRS})
++if (NOT(TARGET gtest AND TARGET gtest_main))
++  if(NOT(GTEST_FOUND))
++    find_package(GTest REQUIRED)
 +  endif()
 +endif()
 +
- find_package(Boost REQUIRED)
- include_directories(${Boost_INCLUDE_DIRS})
++if (NOT(BOOST_FOUND))
++  find_package(Boost REQUIRED)
 +endif()
 +
- add_subdirectory(support)
++# add_subdirectory(support/src)
 +add_subdirectory(src)
 +add_subdirectory(sim)
 +
 +enable_testing()
 +add_subdirectory(test)
++add_subdirectory(support/test)
++add_test(NAME dmclock-tests
++  COMMAND $<TARGET_FILE:dmclock-tests>)
++add_test(NAME dmclock-data-struct-tests
++  COMMAND $<TARGET_FILE:dmclock-data-struct-tests>)
index febd4f0ab6f826fc669a9047b2c86fd7dc8c351d,0000000000000000000000000000000000000000..c088d2f1cd7c96205a7183354f84e86df65b04be
mode 100644,000000..100644
--- /dev/null
@@@ -1,1 -1,0 +1,9 @@@
++if(K_WAY_HEAP)
++  if(K_WAY_HEAP LESS 2)
++    message(FATAL_ERROR "K_WAY_HEAP value should be at least 2")
++  else()
++    set(CMAKE_CXX_SIM_FLAGS "-DK_WAY_HEAP=${K_WAY_HEAP}")
++  endif()
++endif()
++
 +add_subdirectory(src)
index a61cc3204e4e96dd5fae5c8f5051ff7e0540e0fd,0000000000000000000000000000000000000000..e318d6e90ac328e6810aa4b7a1fc4f628003ee37
mode 100644,000000..100644
--- /dev/null
@@@ -1,225 -1,0 +1,227 @@@
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +
 +/*
 + * Copyright (C) 2016 Red Hat Inc.
 + */
 +
 +
 +#pragma once
 +
 +
 +#include <thread>
 +#include <mutex>
 +#include <condition_variable>
 +#include <chrono>
 +#include <deque>
 +
 +#include "sim_recs.h"
 +
 +
 +namespace crimson {
 +  namespace qos_simulation {
 +
 +    template<typename Q, typename ReqPm, typename RespPm, typename Accum>
 +    class SimulatedServer {
 +
 +      struct QueueItem {
 +      ClientId                     client;
 +      std::unique_ptr<TestRequest> request;
 +      RespPm                       additional;
 +
 +      QueueItem(const ClientId&                _client,
 +                std::unique_ptr<TestRequest>&& _request,
 +                const RespPm&                  _additional) :
 +        client(_client),
 +        request(std::move(_request)),
 +        additional(_additional)
 +      {
 +        // empty
 +      }
 +      }; // QueueItem
 +
 +    public:
 +
 +      struct InternalStats {
 +      std::mutex mtx;
 +      std::chrono::nanoseconds add_request_time;
 +      std::chrono::nanoseconds request_complete_time;
 +      uint32_t add_request_count;
 +      uint32_t request_complete_count;
 +
 +      InternalStats() :
 +        add_request_time(0),
 +        request_complete_time(0),
 +        add_request_count(0),
 +        request_complete_count(0)
 +      {
 +        // empty
 +      }
 +      };
 +
 +      using ClientRespFunc = std::function<void(ClientId,
 +                                              const TestResponse&,
 +                                              const ServerId&,
 +                                              const RespPm&)>;
 +
 +      using ServerAccumFunc = std::function<void(Accum& accumulator,
 +                                               const RespPm& additional)>;
 +
 +    protected:
 +
 +      const ServerId                 id;
 +      Q*                             priority_queue;
 +      ClientRespFunc                 client_resp_f;
 +      int                            iops;
 +      size_t                         thread_pool_size;
 +
 +      bool                           finishing;
 +      std::chrono::microseconds      op_time;
 +
 +      std::mutex                     inner_queue_mtx;
 +      std::condition_variable        inner_queue_cv;
 +      std::deque<QueueItem>          inner_queue;
 +
 +      std::thread*                   threads;
 +
 +      using InnerQGuard = std::lock_guard<decltype(inner_queue_mtx)>;
 +      using Lock = std::unique_lock<std::mutex>;
 +
 +      // data collection
 +
 +      ServerAccumFunc accum_f;
 +      Accum accumulator;
 +
 +      InternalStats internal_stats;
 +
 +    public:
 +
 +      using CanHandleRequestFunc = std::function<bool(void)>;
 +      using HandleRequestFunc =
 +      std::function<void(const ClientId&,std::unique_ptr<TestRequest>,const RespPm&)>;
 +      using CreateQueueF = std::function<Q*(CanHandleRequestFunc,HandleRequestFunc)>;
 +                                      
 +
 +      SimulatedServer(ServerId _id,
 +                    int _iops,
 +                    size_t _thread_pool_size,
 +                    const ClientRespFunc& _client_resp_f,
 +                    const ServerAccumFunc& _accum_f,
 +                    CreateQueueF _create_queue_f) :
 +      id(_id),
 +      priority_queue(_create_queue_f(std::bind(&SimulatedServer::has_avail_thread,
 +                                               this),
 +                                     std::bind(&SimulatedServer::inner_post,
 +                                               this,
 +                                               std::placeholders::_1,
 +                                               std::placeholders::_2,
 +                                               std::placeholders::_3))),
 +      client_resp_f(_client_resp_f),
 +      iops(_iops),
 +      thread_pool_size(_thread_pool_size),
 +      finishing(false),
 +      accum_f(_accum_f)
 +      {
 +      op_time =
 +        std::chrono::microseconds((int) (0.5 +
 +                                         thread_pool_size * 1000000.0 / iops));
 +      std::chrono::milliseconds delay(1000);
 +      threads = new std::thread[thread_pool_size];
 +      for (size_t i = 0; i < thread_pool_size; ++i) {
 +        threads[i] = std::thread(&SimulatedServer::run, this, delay);
 +      }
 +      }
 +
 +      virtual ~SimulatedServer() {
 +      Lock l(inner_queue_mtx);
 +      finishing = true;
 +      inner_queue_cv.notify_all();
 +      l.unlock();
 +
 +      for (size_t i = 0; i < thread_pool_size; ++i) {
 +        threads[i].join();
 +      }
 +
 +      delete[] threads;
++
++      delete priority_queue;
 +      }
 +
 +      void post(const TestRequest& request,
 +              const ClientId& client_id,
 +              const ReqPm& req_params)
 +      {
 +      time_stats(internal_stats.mtx,
 +                 internal_stats.add_request_time,
 +                 [&](){
 +                   priority_queue->add_request(request, client_id, req_params);
 +                 });
 +      count_stats(internal_stats.mtx,
 +                  internal_stats.add_request_count);
 +      }
 +
 +      bool has_avail_thread() {
 +      InnerQGuard g(inner_queue_mtx);
 +      return inner_queue.size() <= thread_pool_size;
 +      }
 +
 +      const Accum& get_accumulator() const { return accumulator; }
 +      const Q& get_priority_queue() const { return *priority_queue; }
 +      const InternalStats& get_internal_stats() const { return internal_stats; }
 +
 +    protected:
 +
 +      void inner_post(const ClientId& client,
 +                    std::unique_ptr<TestRequest> request,
 +                    const RespPm& additional) {
 +      Lock l(inner_queue_mtx);
 +      assert(!finishing);
 +      accum_f(accumulator, additional);
 +      inner_queue.emplace_back(QueueItem(client,
 +                                         std::move(request),
 +                                         additional));
 +      inner_queue_cv.notify_one();
 +      }
 +
 +      void run(std::chrono::milliseconds check_period) {
 +      Lock l(inner_queue_mtx);
 +      while(true) {
 +        while(inner_queue.empty() && !finishing) {
 +          inner_queue_cv.wait_for(l, check_period);
 +        }
 +        if (!inner_queue.empty()) {
 +          auto& front = inner_queue.front();
 +          auto client = front.client;
 +          auto req = std::move(front.request);
 +          auto additional = front.additional;
 +          inner_queue.pop_front();
 +
 +          l.unlock();
 +
 +          // simulation operation by sleeping; then call function to
 +          // notify server of completion
 +          std::this_thread::sleep_for(op_time);
 +
 +          TestResponse resp(req->epoch);
 +          // TODO: rather than assuming this constructor exists, perhaps
 +          // pass in a function that does this mapping?
 +          client_resp_f(client, resp, id, additional);
 +
 +          time_stats(internal_stats.mtx,
 +                     internal_stats.request_complete_time,
 +                     [&](){
 +                       priority_queue->request_completed();
 +                     });
 +          count_stats(internal_stats.mtx,
 +                      internal_stats.request_complete_count);
 +
 +          l.lock(); // in prep for next iteration of loop
 +        } else {
 +          break;
 +        }
 +      }
 +      }
 +    }; // class SimulatedServer
 +
 +  }; // namespace qos_simulation
 +}; // namespace crimson
index 18e752d8a354f513dcfeb58a34e8f70ee07e9ff5,0000000000000000000000000000000000000000..a967d004b6dcfcb0658edc2aa7740a6e8ffb5a21
mode 100644,000000..100644
--- /dev/null
@@@ -1,430 -1,0 +1,441 @@@
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +
 +/*
 + * Copyright (C) 2016 Red Hat Inc.
 + */
 +
 +
 +#pragma once
 +
 +
 +#include <assert.h>
 +
 +#include <memory>
 +#include <chrono>
 +#include <map>
 +#include <random>
 +#include <iostream>
 +#include <iomanip>
 +#include <string>
 +
 +
 +namespace crimson {
 +  namespace qos_simulation {
 +
 +    template<typename ServerId, typename ClientId, typename TS, typename TC>
 +    class Simulation {
 +  
 +    public:
 +
 +      using TimePoint = std::chrono::time_point<std::chrono::steady_clock>;
 +
 +    protected:
 +
 +      using ClientMap = std::map<ClientId,TC*>;
 +      using ServerMap = std::map<ServerId,TS*>;
 +
 +      uint server_count = 0;
 +      uint client_count = 0;
 +
 +      ServerMap servers;
 +      ClientMap clients;
 +      std::vector<ServerId> server_ids;
 +
 +      TimePoint early_time;
 +      TimePoint servers_created_time;
 +      TimePoint clients_created_time;
 +      TimePoint clients_finished_time;
 +      TimePoint late_time;
 +
 +      std::default_random_engine prng;
 +
 +      bool has_run = false;
 +
 +
 +    public:
 +
 +      double fmt_tp(const TimePoint& t) {
 +      auto c = t.time_since_epoch().count();
 +      return uint64_t(c / 1000000.0 + 0.5) % 100000 / 1000.0;
 +      }
 +
 +      TimePoint now() {
 +      return std::chrono::steady_clock::now();
 +      }
 +
 +      using ClientBasedServerSelectFunc =
 +      std::function<const ServerId&(uint64_t, uint16_t)>;
 +
 +      using ClientFilter = std::function<bool(const ClientId&)>;
 +
 +      using ServerFilter = std::function<bool(const ServerId&)>;
 +
 +      using ServerDataOutF =
 +      std::function<void(std::ostream& out,
 +                         Simulation* sim, ServerFilter,
 +                         int header_w, int data_w, int data_prec)>;
 +
 +      using ClientDataOutF =
 +      std::function<void(std::ostream& out,
 +                         Simulation* sim, ClientFilter,
 +                         int header_w, int data_w, int data_prec)>;
 +
 +      Simulation() :
 +      early_time(now()),
 +      prng(std::chrono::system_clock::now().time_since_epoch().count())
 +      {
 +      // empty
 +      }
 +
++      ~Simulation() {
++      for (auto c : clients) {
++        TC* cp = c.second;
++        delete cp;
++      }
++
++      for (auto s : servers) {
++        delete s.second;
++      }
++      }
++
 +      uint get_client_count() const { return client_count; }
 +      uint get_server_count() const { return server_count; }
 +      TC& get_client(ClientId id) { return *clients[id]; }
 +      TS& get_server(ServerId id) { return *servers[id]; }
 +      const ServerId& get_server_id(uint index) const {
 +      return server_ids[index];
 +      }
 +
 +
 +      void add_servers(uint count,
 +                     std::function<TS*(ServerId)> create_server_f) {
 +      uint i = server_count;
 +
 +      // increment server_count before creating servers since they
 +      // will start running immediately and may use the server_count
 +      // value; NB: this could still be an issue if servers are
 +      // added with multiple add_servers calls; consider using a
 +      // separate start function after all servers (and clients?)
 +      // have been added
 +      server_count += count;
 +
 +      for (; i < server_count; ++i) {
 +        server_ids.push_back(i);
 +        servers[i] = create_server_f(i);
 +      }
 +
 +      servers_created_time = now();
 +      }
 +
 +
 +      void add_clients(uint count,
 +                     std::function<TC*(ClientId)> create_client_f) {
 +      uint i = client_count;
 +
 +      // increment client_count before creating clients since they
 +      // will start running immediately and may use the client_count
 +      // value (e.g., in the server selection function); NB: this could
 +      // still be an issue if clients are added with multiple
 +      // add_clients calls; consider using a separate start function
 +      // after all clients have been added
 +      client_count += count;
 +
 +      for (; i < client_count; ++i) {
 +        clients[i] = create_client_f(i);
 +      }
 +
 +      clients_created_time = now();
 +      }
 +
 +
 +      void run() {
 +      assert(server_count > 0);
 +      assert(client_count > 0);
 +
 +      std::cout << "simulation started" << std::endl;
 +
 +      // clients are now running; wait for all to finish
 +
 +      for (auto const &i : clients) {
 +        i.second->wait_until_done();
 +      }
 +
 +      late_time = clients_finished_time = now();
 +
 +      std::cout << "simulation completed in " <<
 +        std::chrono::duration_cast<std::chrono::milliseconds>(clients_finished_time - servers_created_time).count() <<
 +        " millisecs" << std::endl;
 +
 +      has_run = true;
 +      } // run
 +
 +
 +      void display_stats(std::ostream& out,
 +                       ServerDataOutF server_out_f, ClientDataOutF client_out_f,
 +                       ServerFilter server_filter =
 +                       [] (const ServerId&) { return true; },
 +                       ClientFilter client_filter =
 +                       [] (const ClientId&) { return true; },
 +                       int head_w = 12, int data_w = 7, int data_prec = 2) {
 +      assert(has_run);
 +
 +      // skip first 2 secondsd of data
 +      const std::chrono::seconds skip_amount(0);
 +      // calculate in groups of 5 seconds
 +      const std::chrono::seconds measure_unit(2);
 +      // unit to output reports in
 +      const std::chrono::seconds report_unit(1);
 +
 +      // compute and display stats
 +
 +      TimePoint earliest_start = late_time;
 +      TimePoint latest_start = early_time;
 +      TimePoint earliest_finish = late_time;
 +      TimePoint latest_finish = early_time;
 +
 +      for (auto const &c : clients) {
 +        auto start = c.second->get_op_times().front();
 +        auto end = c.second->get_op_times().back();
 +
 +        if (start < earliest_start) { earliest_start = start; }
 +        if (start > latest_start) { latest_start = start; }
 +        if (end < earliest_finish) { earliest_finish = end; }
 +        if (end > latest_finish) { latest_finish = end; }
 +      }
 +
 +      double ops_factor =
 +        std::chrono::duration_cast<std::chrono::duration<double>>(measure_unit) /
 +        std::chrono::duration_cast<std::chrono::duration<double>>(report_unit);
 +
 +      const auto start_edge = clients_created_time + skip_amount;
 +
 +      std::map<ClientId,std::vector<double>> ops_data;
 +
 +      for (auto const &c : clients) {
 +        auto it = c.second->get_op_times().begin();
 +        const auto end = c.second->get_op_times().end();
 +        while (it != end && *it < start_edge) { ++it; }
 +
 +        for (auto time_edge = start_edge + measure_unit;
 +             time_edge <= latest_finish + measure_unit;
 +             time_edge += measure_unit) {
 +          int count = 0;
 +          for (; it != end && *it < time_edge; ++count, ++it) { /* empty */ }
 +          double ops_per_second = double(count) / ops_factor;
 +          ops_data[c.first].push_back(ops_per_second);
 +        }
 +      }
 +
 +      out << "==== Client Data ====" << std::endl;
 +
 +      out << std::setw(head_w) << "client:";
 +      for (auto const &c : clients) {
 +        if (!client_filter(c.first)) continue;
 +        out << " " << std::setw(data_w) << c.first;
 +      }
 +      out << std::setw(data_w) << "total" << std::endl;
 +
 +      {
 +        bool has_data;
 +        size_t i = 0;
 +        do {
 +          std::string line_header = "t_" + std::to_string(i) + ":";
 +          out << std::setw(head_w) << line_header;
 +          has_data = false;
 +          double total = 0.0;
 +          for (auto const &c : clients) {
 +            double data = 0.0;
 +            if (i < ops_data[c.first].size()) {
 +              data = ops_data[c.first][i];
 +              has_data = true;
 +            }
 +            total += data;
 +
 +            if (!client_filter(c.first)) continue;
 +
 +            out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +              std::fixed << data;
 +          }
 +          out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +            std::fixed << total << std::endl;
 +          ++i;
 +        } while(has_data);
 +      }
 +
 +      client_out_f(out, this, client_filter, head_w, data_w, data_prec);
 +
 +      display_client_internal_stats<std::chrono::nanoseconds>(out,
 +                                                              "nanoseconds");
 +
 +      out << std::endl << "==== Server Data ====" << std::endl;
 +
 +      out << std::setw(head_w) << "server:";
 +      for (auto const &s : servers) {
 +        if (!server_filter(s.first)) continue;
 +        out << " " << std::setw(data_w) << s.first;
 +      }
 +      out << " " << std::setw(data_w) << "total" << std::endl;
 +
 +      server_out_f(out, this, server_filter, head_w, data_w, data_prec);
 +
 +      display_server_internal_stats<std::chrono::nanoseconds>(out,
 +                                                              "nanoseconds");
 +
 +      // clean up clients then servers
 +
 +      for (auto i = clients.begin(); i != clients.end(); ++i) {
 +        delete i->second;
 +        i->second = nullptr;
 +      }
 +
 +      for (auto i = servers.begin(); i != servers.end(); ++i) {
 +        delete i->second;
 +        i->second = nullptr;
 +      }
 +      } // display_stats
 +
 +
 +      template<typename T>
 +      void display_server_internal_stats(std::ostream& out,
 +                                       std::string time_unit) {
 +      T add_request_time(0);
 +      T request_complete_time(0);
 +      uint32_t add_request_count = 0;
 +      uint32_t request_complete_count = 0;
 +
 +      for (uint i = 0; i < get_server_count(); ++i) {
 +        const auto& server = get_server(i);
 +        const auto& is = server.get_internal_stats();
 +        add_request_time +=
 +          std::chrono::duration_cast<T>(is.add_request_time);
 +        request_complete_time +=
 +          std::chrono::duration_cast<T>(is.request_complete_time);
 +        add_request_count += is.add_request_count;
 +        request_complete_count += is.request_complete_count;
 +      }
 +
 +      double add_request_time_per_unit =
 +        double(add_request_time.count()) / add_request_count ;
 +      out << "total time to add requests: " <<
 +        std::fixed << add_request_time.count() << " " << time_unit <<
 +        ";" << std::endl <<
 +        "    count: " << add_request_count << ";" << std::endl <<
 +        "    average: " << add_request_time_per_unit <<
 +        " " << time_unit << " per request/response" << std::endl;
 +
 +      double request_complete_time_unit =
 +        double(request_complete_time.count()) / request_complete_count ;
 +      out << "total time to note requests complete: " << std::fixed <<
 +        request_complete_time.count() << " " << time_unit << ";" <<
 +        std::endl << 
 +        "    count: " << request_complete_count << ";" << std::endl <<
 +        "    average: " << request_complete_time_unit <<
 +        " " << time_unit << " per request/response" << std::endl;
 +
 +      out << std::endl;
 +
 +      assert(add_request_count == request_complete_count);
 +      out << "server timing for QOS algorithm: " <<
 +        add_request_time_per_unit + request_complete_time_unit <<
 +        " " << time_unit << " per request/response" << std::endl;
 +      }
 +
 +
 +      template<typename T>
 +      void display_client_internal_stats(std::ostream& out,
 +                                       std::string time_unit) {
 +      T track_resp_time(0);
 +      T get_req_params_time(0);
 +      uint32_t track_resp_count = 0;
 +      uint32_t get_req_params_count = 0;
 +
 +      for (uint i = 0; i < get_client_count(); ++i) {
 +        const auto& client = get_client(i);
 +        const auto& is = client.get_internal_stats();
 +        track_resp_time +=
 +          std::chrono::duration_cast<T>(is.track_resp_time);
 +        get_req_params_time +=
 +          std::chrono::duration_cast<T>(is.get_req_params_time);
 +        track_resp_count += is.track_resp_count;
 +        get_req_params_count += is.get_req_params_count;
 +      }
 +
 +      double track_resp_time_unit =
 +        double(track_resp_time.count()) / track_resp_count;
 +      out << "total time to track responses: " <<
 +        std::fixed << track_resp_time.count() << " " << time_unit << ";" <<
 +        std::endl <<
 +        "    count: " << track_resp_count << ";" << std::endl <<
 +        "    average: " << track_resp_time_unit << " " << time_unit <<
 +        " per request/response" << std::endl;
 +
 +      double get_req_params_time_unit =
 +        double(get_req_params_time.count()) / get_req_params_count;
 +      out << "total time to get request parameters: " <<
 +        std::fixed << get_req_params_time.count() << " " << time_unit <<
 +        ";" << std::endl <<
 +        "    count: " << get_req_params_count << ";" << std::endl <<
 +        "    average: " << get_req_params_time_unit << " " << time_unit <<
 +        " per request/response" << std::endl;
 +
 +      out << std::endl;
 +
 +      assert(track_resp_count == get_req_params_count);
 +      out << "client timing for QOS algorithm: " <<
 +        track_resp_time_unit + get_req_params_time_unit << " " <<
 +        time_unit << " per request/response" << std::endl;
 +      }
 +
 +
 +      // **** server selection functions ****
 +
 +
 +      const ServerId& server_select_alternate(uint64_t seed,
 +                                            uint16_t client_idx) {
 +      uint index = (client_idx + seed) % server_count;
 +      return server_ids[index];
 +      }
 +
 +
 +      // returns a lambda using the range specified as servers_per (client)
 +      ClientBasedServerSelectFunc
 +      make_server_select_alt_range(uint16_t servers_per) {
 +      return [servers_per,this](uint64_t seed, uint16_t client_idx)
 +        -> const ServerId& {
 +        double factor = double(server_count) / client_count;
 +        uint offset = seed % servers_per;
 +        uint index = (uint(0.5 + client_idx * factor) + offset) % server_count;
 +        return server_ids[index];
 +      };
 +      }
 +
 +
 +      // function to choose a server randomly
 +      const ServerId& server_select_random(uint64_t seed, uint16_t client_idx) {
 +      uint index = prng() % server_count;
 +      return server_ids[index];
 +      }
 +
 +  
 +      // function to choose a server randomly
 +      ClientBasedServerSelectFunc
 +      make_server_select_ran_range(uint16_t servers_per) {
 +      return [servers_per,this](uint64_t seed, uint16_t client_idx)
 +        -> const ServerId& {
 +        double factor = double(server_count) / client_count;
 +        uint offset = prng() % servers_per;
 +        uint index = (uint(0.5 + client_idx * factor) + offset) % server_count;
 +        return server_ids[index];
 +      };
 +      }
 +
 +
 +      // function to always choose the first server
 +      const ServerId& server_select_0(uint64_t seed, uint16_t client_idx) {
 +      return server_ids[0];
 +      }
 +    }; // class Simulation
 +
 +  }; // namespace qos_simulation
 +}; // namespace crimson
index c3ba1e18fbd6d1a23c155ea26e1de6602aa2b3dd,0000000000000000000000000000000000000000..668869b8d72fb040d3b74997db89a071e0f9a72a
mode 100644,000000..100644
--- /dev/null
@@@ -1,322 -1,0 +1,324 @@@
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +
 +/*
 + * Copyright (C) 2016 Red Hat Inc.
 + */
 +
 +
 +#include "test_dmclock.h"
 +#include "config.h"
 +
 +#ifdef PROFILE
 +#include "profile.h"
 +#endif
 +
 +
 +namespace dmc = crimson::dmclock;
 +namespace test = crimson::test_dmc;
 +namespace sim = crimson::qos_simulation;
 +
 +using namespace std::placeholders;
 +
 +
 +namespace crimson {
 +    namespace test_dmc {
 +        void server_data(std::ostream& out,
 +                         test::MySim* sim,
 +                         test::MySim::ServerFilter server_disp_filter,
 +                         int head_w, int data_w, int data_prec);
 +
 +        void client_data(std::ostream& out,
 +                         test::MySim* sim,
 +                         test::MySim::ClientFilter client_disp_filter,
 +                         int head_w, int data_w, int data_prec);
 +    }
 +}
 +
 +
 +int main(int argc, char* argv[]) {
 +    std::vector<const char*> args;
 +    for (int i = 1; i < argc; ++i) {
 +      args.push_back(argv[i]);
 +    }
 +
 +    std::string conf_file_list;
 +    sim::ceph_argparse_early_args(args, &conf_file_list);
 +
 +    sim::sim_config_t g_conf;
 +    std::vector<sim::cli_group_t> &cli_group = g_conf.cli_group;
 +    std::vector<sim::srv_group_t> &srv_group = g_conf.srv_group;
 +
 +    if (!conf_file_list.empty()) {
 +      int ret;
 +      ret = sim::parse_config_file(conf_file_list, g_conf);
 +      if (ret) {
 +      // error
 +      _exit(1);
 +      }
 +    } else {
 +      // default simulation parameter
 +      g_conf.client_groups = 2;
 +
 +      sim::srv_group_t st;
 +      srv_group.push_back(st);
 +
 +      sim::cli_group_t ct1(99, 0);
 +      cli_group.push_back(ct1);
 +
 +      sim::cli_group_t ct2(1, 10);
 +      cli_group.push_back(ct2);
 +    }
 +
 +    const uint server_groups = g_conf.server_groups;
 +    const uint client_groups = g_conf.client_groups;
 +    const bool server_random_selection = g_conf.server_random_selection;
 +    const bool server_soft_limit = g_conf.server_soft_limit;
 +    uint server_total_count = 0;
 +    uint client_total_count = 0;
 +
 +    for (uint i = 0; i < client_groups; ++i) {
 +      client_total_count += cli_group[i].client_count;
 +    }
 +
 +    for (uint i = 0; i < server_groups; ++i) {
 +      server_total_count += srv_group[i].server_count;
 +    }
 +
 +    std::vector<test::dmc::ClientInfo> client_info;
 +    for (uint i = 0; i < client_groups; ++i) {
 +      client_info.push_back(test::dmc::ClientInfo 
 +                        { cli_group[i].client_reservation,
 +                          cli_group[i].client_weight,
 +                          cli_group[i].client_limit } );
 +    }
 +
 +    auto ret_client_group_f = [&](const ClientId& c) -> uint {
 +      uint group_max = 0;
 +      uint i = 0;
 +      for (; i < client_groups; ++i) {
 +      group_max += cli_group[i].client_count;
 +      if (c < group_max) {
 +        break;
 +      }
 +      }
 +      return i;
 +    };
 +
 +    auto ret_server_group_f = [&](const ServerId& s) -> uint {
 +      uint group_max = 0;
 +      uint i = 0;
 +      for (; i < server_groups; ++i) {
 +      group_max += srv_group[i].server_count;
 +      if (s < group_max) {
 +        break;
 +      }
 +      }
 +      return i;
 +    };
 +
 +    auto client_info_f = [=](const ClientId& c) -> test::dmc::ClientInfo {
 +      return client_info[ret_client_group_f(c)];
 +    };
 +
 +    auto client_disp_filter = [=] (const ClientId& i) -> bool {
 +        return i < 3 || i >= (client_total_count - 3);
 +    };
 +
 +    auto server_disp_filter = [=] (const ServerId& i) -> bool {
 +        return i < 3 || i >= (server_total_count - 3);
 +    };
 +
 +
 +    test::MySim *simulation;
 +  
 +
 +    // lambda to post a request to the identified server; called by client
 +    test::SubmitFunc server_post_f =
 +        [&simulation](const ServerId& server,
 +                      const sim::TestRequest& request,
 +                      const ClientId& client_id,
 +                      const test::dmc::ReqParams& req_params) {
 +        test::DmcServer& s = simulation->get_server(server);
 +        s.post(request, client_id, req_params);
 +    };
 +
 +    std::vector<std::vector<sim::CliInst>> cli_inst;
 +    for (uint i = 0; i < client_groups; ++i) {
 +      if (cli_group[i].client_wait == std::chrono::seconds(0)) {
 +      cli_inst.push_back(
 +          { { sim::req_op, 
 +              (uint32_t)cli_group[i].client_total_ops,
 +              (double)cli_group[i].client_iops_goal, 
 +              (uint16_t)cli_group[i].client_outstanding_ops } } );
 +      } else {
 +      cli_inst.push_back(
 +          { { sim::wait_op, cli_group[i].client_wait },
 +            { sim::req_op, 
 +              (uint32_t)cli_group[i].client_total_ops,
 +              (double)cli_group[i].client_iops_goal, 
 +              (uint16_t)cli_group[i].client_outstanding_ops } } );
 +      }
 +    }
 +
 +    simulation = new test::MySim();
 +
 +    test::DmcServer::ClientRespFunc client_response_f =
 +        [&simulation](ClientId client_id,
 +                      const sim::TestResponse& resp,
 +                      const ServerId& server_id,
 +                      const dmc::PhaseType& phase) {
 +        simulation->get_client(client_id).receive_response(resp,
 +                                                           server_id,
 +                                                           phase);
 +    };
 +
 +    test::CreateQueueF create_queue_f =
 +        [&](test::DmcQueue::CanHandleRequestFunc can_f,
 +            test::DmcQueue::HandleRequestFunc handle_f) -> test::DmcQueue* {
 +        return new test::DmcQueue(client_info_f, can_f, handle_f, server_soft_limit);
 +    };
 +
 + 
 +    auto create_server_f = [&](ServerId id) -> test::DmcServer* {
 +      uint i = ret_server_group_f(id);
 +      return new test::DmcServer(id,
 +                                 srv_group[i].server_iops,
 +                               srv_group[i].server_threads,
 +                               client_response_f,
 +                               test::dmc_server_accumulate_f,
 +                               create_queue_f);
 +    };
 +
 +    auto create_client_f = [&](ClientId id) -> test::DmcClient* {
 +      uint i = ret_client_group_f(id);
 +      test::MySim::ClientBasedServerSelectFunc server_select_f;
 +      uint client_server_select_range = cli_group[i].client_server_select_range;
 +      if (!server_random_selection) {
 +      server_select_f = simulation->make_server_select_alt_range(client_server_select_range);
 +      } else {
 +      server_select_f = simulation->make_server_select_ran_range(client_server_select_range);
 +      }
 +      return new test::DmcClient(id,
 +                               server_post_f,
 +                               std::bind(server_select_f, _1, id),
 +                               test::dmc_client_accumulate_f,
 +                               cli_inst[i]);
 +    };
 +
 +#if 1
 +    std::cout << "[global]" << std::endl << g_conf << std::endl;
 +    for (uint i = 0; i < client_groups; ++i) {
 +      std::cout << std::endl << "[client." << i << "]" << std::endl;
 +      std::cout << cli_group[i] << std::endl;
 +    }
 +    for (uint i = 0; i < server_groups; ++i) {
 +      std::cout << std::endl << "[server." << i << "]" << std::endl;
 +      std::cout << srv_group[i] << std::endl;
 +    }
 +    std::cout << std::endl;
 +#endif
 +
 +    simulation->add_servers(server_total_count, create_server_f);
 +    simulation->add_clients(client_total_count, create_client_f);
 +
 +    simulation->run();
 +    simulation->display_stats(std::cout,
 +                              &test::server_data, &test::client_data,
 +                              server_disp_filter, client_disp_filter);
++
++    delete simulation;
 +} // main
 +
 +
 +void test::client_data(std::ostream& out,
 +               test::MySim* sim,
 +               test::MySim::ClientFilter client_disp_filter,
 +               int head_w, int data_w, int data_prec) {
 +    // report how many ops were done by reservation and proportion for
 +    // each client
 +
 +    int total_r = 0;
 +    out << std::setw(head_w) << "res_ops:";
 +    for (uint i = 0; i < sim->get_client_count(); ++i) {
 +        const auto& client = sim->get_client(i);
 +        auto r = client.get_accumulator().reservation_count;
 +        total_r += r;
 +        if (!client_disp_filter(i)) continue;
 +        out << " " << std::setw(data_w) << r;
 +    }
 +    out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +        std::fixed << total_r << std::endl;
 +
 +    int total_p = 0;
 +    out << std::setw(head_w) << "prop_ops:";
 +    for (uint i = 0; i < sim->get_client_count(); ++i) {
 +        const auto& client = sim->get_client(i);
 +        auto p = client.get_accumulator().proportion_count;
 +        total_p += p;
 +        if (!client_disp_filter(i)) continue;
 +        out << " " << std::setw(data_w) << p;
 +    }
 +    out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +        std::fixed << total_p << std::endl;
 +}
 +
 +
 +void test::server_data(std::ostream& out,
 +               test::MySim* sim,
 +               test::MySim::ServerFilter server_disp_filter,
 +               int head_w, int data_w, int data_prec) {
 +    out << std::setw(head_w) << "res_ops:";
 +    int total_r = 0;
 +    for (uint i = 0; i < sim->get_server_count(); ++i) {
 +        const auto& server = sim->get_server(i);
 +        auto rc = server.get_accumulator().reservation_count;
 +        total_r += rc;
 +        if (!server_disp_filter(i)) continue;
 +        out << " " << std::setw(data_w) << rc;
 +    }
 +    out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +        std::fixed << total_r << std::endl;
 +
 +    out << std::setw(head_w) << "prop_ops:";
 +    int total_p = 0;
 +    for (uint i = 0; i < sim->get_server_count(); ++i) {
 +        const auto& server = sim->get_server(i);
 +        auto pc = server.get_accumulator().proportion_count;
 +        total_p += pc;
 +        if (!server_disp_filter(i)) continue;
 +        out << " " << std::setw(data_w) << pc;
 +    }
 +    out << " " << std::setw(data_w) << std::setprecision(data_prec) <<
 +        std::fixed << total_p << std::endl;
 +
 +    const auto& q = sim->get_server(0).get_priority_queue();
 +    out << std::endl <<
 +      " k-way heap: " << q.get_heap_branching_factor() << std::endl
 +      << std::endl;
 +
 +#ifdef PROFILE
 +    crimson::ProfileCombiner<std::chrono::nanoseconds> art_combiner;
 +    crimson::ProfileCombiner<std::chrono::nanoseconds> rct_combiner;
 +    for (uint i = 0; i < sim->get_server_count(); ++i) {
 +      const auto& q = sim->get_server(i).get_priority_queue();
 +      const auto& art = q.add_request_timer;
 +      art_combiner.combine(art);
 +      const auto& rct = q.request_complete_timer;
 +      rct_combiner.combine(rct);
 +    }
 +    out << "Server add_request_timer: count:" << art_combiner.get_count() <<
 +      ", mean:" << art_combiner.get_mean() <<
 +      ", std_dev:" << art_combiner.get_std_dev() <<
 +      ", low:" << art_combiner.get_low() <<
 +      ", high:" << art_combiner.get_high() << std::endl;
 +    out << "Server request_complete_timer: count:" << rct_combiner.get_count() <<
 +      ", mean:" << rct_combiner.get_mean() <<
 +      ", std_dev:" << rct_combiner.get_std_dev() <<
 +      ", low:" << rct_combiner.get_low() <<
 +      ", high:" << rct_combiner.get_high() << std::endl;
 +    out << "Server combined mean: " <<
 +      (art_combiner.get_mean() + rct_combiner.get_mean()) <<
 +      std::endl;
 +#endif
 +}
index 691e64cce4389f914047ad49e45d49519984f2eb,0000000000000000000000000000000000000000..7178f266e2a41d13276ad1fbe3a92bc49135cd37
mode 100644,000000..100644
--- /dev/null
@@@ -1,19 -1,0 +1,24 @@@
- include_directories(${BOOST_INCLUDE_DIR})
++include_directories(${GTEST_INCLUDE_DIRS})
++include_directories(${Boost_INCLUDE_DIRS})
 +include_directories(../support/src)
- set(local_flags "-Wall -pthread")
 +
++set(local_flags "-std=c++11 -Wno-write-strings -Wall -pthread")
++
++if(DO_NOT_DELAY_TAG_CALC)
++  set(local_flags "${local_flags} -DDO_NOT_DELAY_TAG_CALC")
++endif()
 +
 +set(dmc_srcs dmclock_util.cc ../support/src/run_every.cc)
 +
 +set_source_files_properties(${dmc_srcs}
 +  PROPERTIES
 +  COMPILE_FLAGS "${local_flags}"
 +  )
 +
 +if ("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
 +  set(warnings_off " -Wno-unused-variable -Wno-unused-function")
 +elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
 +  set(warnings_off " -Wno-unused-but-set-variable -Wno-unused-function")
 +endif()
 +
 +add_library(dmclock STATIC ${dmc_srcs})
index 65013063fa7156ea0384bceb8258fe5bab45c44c,0000000000000000000000000000000000000000..8aaad3977264847c0c3351d76eff05da7d629171
mode 100644,000000..100644
--- /dev/null
@@@ -1,1588 -1,0 +1,1602 @@@
-                const ReqParams& req_params,
-                const Time& time,
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +
 +/*
 + * Copyright (C) 2017 Red Hat Inc.
 + */
 +
 +
 +#pragma once
 +
 +/* COMPILATION OPTIONS
 + *
 + * By default we include an optimization over the originally published
 + * dmclock algorithm using not the values of rho and delta that were
 + * sent in with a request but instead the most recent rho and delta
 + * values from the requests's client. To restore the algorithm's
 + * original behavior, define DO_NOT_DELAY_TAG_CALC (i.e., compiler
 + * argument -DDO_NOT_DELAY_TAG_CALC).
 + *
 + * The prop_heap does not seem to be necessary. The only thing it
 + * would help with is quickly finding the mininum proportion/prioity
 + * when an idle client became active. To have the code maintain the
 + * proportional heap, define USE_PROP_HEAP (i.e., compiler argument
 + * -DUSE_PROP_HEAP).
 + */
 +
 +#include <assert.h>
 +
 +#include <cmath>
 +#include <memory>
 +#include <map>
 +#include <deque>
 +#include <queue>
 +#include <atomic>
 +#include <mutex>
 +#include <condition_variable>
 +#include <thread>
 +#include <iostream>
 +#include <sstream>
 +#include <limits>
 +
 +#include <boost/variant.hpp>
 +
 +#include "indirect_intrusive_heap.h"
 +#include "run_every.h"
 +#include "dmclock_util.h"
 +#include "dmclock_recs.h"
 +
 +#ifdef PROFILE
 +#include "profile.h"
 +#endif
 +
 +#include "gtest/gtest_prod.h"
 +
 +
 +namespace crimson {
 +
 +  namespace dmclock {
 +
 +    namespace c = crimson;
 +
 +    constexpr double max_tag = std::numeric_limits<double>::is_iec559 ?
 +      std::numeric_limits<double>::infinity() :
 +      std::numeric_limits<double>::max();
 +    constexpr double min_tag = std::numeric_limits<double>::is_iec559 ?
 +      -std::numeric_limits<double>::infinity() :
 +      std::numeric_limits<double>::lowest();
 +    constexpr uint tag_modulo = 1000000;
 +
 +    struct ClientInfo {
 +      const double reservation;  // minimum
 +      const double weight;       // proportional
 +      const double limit;        // maximum
 +
 +      // multiplicative inverses of above, which we use in calculations
 +      // and don't want to recalculate repeatedly
 +      const double reservation_inv;
 +      const double weight_inv;
 +      const double limit_inv;
 +
 +      // order parameters -- min, "normal", max
 +      ClientInfo(double _reservation, double _weight, double _limit) :
 +      reservation(_reservation),
 +      weight(_weight),
 +      limit(_limit),
 +      reservation_inv(0.0 == reservation ? 0.0 : 1.0 / reservation),
 +      weight_inv(     0.0 == weight      ? 0.0 : 1.0 / weight),
 +      limit_inv(      0.0 == limit       ? 0.0 : 1.0 / limit)
 +      {
 +      // empty
 +      }
 +
 +
 +      friend std::ostream& operator<<(std::ostream& out,
 +                                    const ClientInfo& client) {
 +      out <<
 +        "{ ClientInfo:: r:" << client.reservation <<
 +        " w:" << std::fixed << client.weight <<
 +        " l:" << std::fixed << client.limit <<
 +        " 1/r:" << std::fixed << client.reservation_inv <<
 +        " 1/w:" << std::fixed << client.weight_inv <<
 +        " 1/l:" << std::fixed << client.limit_inv <<
 +        " }";
 +      return out;
 +      }
 +    }; // class ClientInfo
 +
 +
 +    struct RequestTag {
 +      double reservation;
 +      double proportion;
 +      double limit;
 +      bool   ready; // true when within limit
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      Time   arrival;
 +#endif
 +
 +      RequestTag(const RequestTag& prev_tag,
 +               const ClientInfo& client,
-                                   req_params.rho,
++               const uint32_t delta,
++               const uint32_t rho,
++               const Time time,
 +               const double cost = 0.0) :
 +      reservation(cost + tag_calc(time,
 +                                  prev_tag.reservation,
 +                                  client.reservation_inv,
-                           req_params.delta,
++                                  rho,
 +                                  true)),
 +      proportion(tag_calc(time,
 +                          prev_tag.proportion,
 +                          client.weight_inv,
-                      req_params.delta,
++                          delta,
 +                          true)),
 +      limit(tag_calc(time,
 +                     prev_tag.limit,
 +                     client.limit_inv,
-       RequestTag(double _res, double _prop, double _lim, const Time& _arrival) :
++                     delta,
 +                     false)),
 +      ready(false)
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      , arrival(time)
 +#endif
 +      {
 +      assert(reservation < max_tag || proportion < max_tag);
 +      }
 +
-       static double tag_calc(const Time& time,
++      RequestTag(const RequestTag& prev_tag,
++               const ClientInfo& client,
++               const ReqParams req_params,
++               const Time time,
++               const double cost = 0.0) :
++      RequestTag(prev_tag, client, req_params.delta, req_params.rho, time, cost)
++      { /* empty */ }
++
++      RequestTag(double _res, double _prop, double _lim, const Time _arrival) :
 +      reservation(_res),
 +      proportion(_prop),
 +      limit(_lim),
 +      ready(false)
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      , arrival(_arrival)
 +#endif
 +      {
 +      assert(reservation < max_tag || proportion < max_tag);
 +      }
 +
 +      RequestTag(const RequestTag& other) :
 +      reservation(other.reservation),
 +      proportion(other.proportion),
 +      limit(other.limit),
 +      ready(other.ready)
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      , arrival(other.arrival)
 +#endif
 +      {
 +      // empty
 +      }
 +
 +      static std::string format_tag_change(double before, double after) {
 +      if (before == after) {
 +        return std::string("same");
 +      } else {
 +        std::stringstream ss;
 +        ss << format_tag(before) << "=>" << format_tag(after);
 +        return ss.str();
 +      }
 +      }
 +
 +      static std::string format_tag(double value) {
 +      if (max_tag == value) {
 +        return std::string("max");
 +      } else if (min_tag == value) {
 +        return std::string("min");
 +      } else {
 +        return format_time(value, tag_modulo);
 +      }
 +      }
 +
 +    private:
 +
-       void do_add_request(RequestRef&&     request,
-                         const C&         client_id,
++      static double tag_calc(const Time time,
 +                           double prev,
 +                           double increment,
 +                           uint32_t dist_req_val,
 +                           bool extreme_is_high) {
 +      if (0.0 == increment) {
 +        return extreme_is_high ? max_tag : min_tag;
 +      } else {
 +        if (0 != dist_req_val) {
 +          increment *= dist_req_val;
 +        }
 +        return std::max(time, prev + increment);
 +      }
 +      }
 +
 +      friend std::ostream& operator<<(std::ostream& out,
 +                                    const RequestTag& tag) {
 +      out <<
 +        "{ RequestTag:: ready:" << (tag.ready ? "true" : "false") <<
 +        " r:" << format_tag(tag.reservation) <<
 +        " p:" << format_tag(tag.proportion) <<
 +        " l:" << format_tag(tag.limit) <<
 +#if 0 // try to resolve this to make sure Time is operator<<'able.
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +        " arrival:" << tag.arrival <<
 +#endif
 +#endif
 +        " }";
 +      return out;
 +      }
 +    }; // class RequestTag
 +
 +
 +    // C is client identifier type, R is request type, B is heap
 +    // branching factor
 +    template<typename C, typename R, uint B>
 +    class PriorityQueueBase {
 +      FRIEND_TEST(dmclock_server, client_idle_erase);
 +
 +    public:
 +
 +      using RequestRef = std::unique_ptr<R>;
 +
 +    protected:
 +
 +      using TimePoint = decltype(std::chrono::steady_clock::now());
 +      using Duration = std::chrono::milliseconds;
 +      using MarkPoint = std::pair<TimePoint,Counter>;
 +
 +      enum class ReadyOption {ignore, lowers, raises};
 +
 +      // forward decl for friend decls
 +      template<double RequestTag::*, ReadyOption, bool>
 +      struct ClientCompare;
 +
 +      class ClientReq {
 +      friend PriorityQueueBase;
 +
 +      RequestTag tag;
 +      C          client_id;
 +      RequestRef request;
 +
 +      public:
 +
 +      ClientReq(const RequestTag& _tag,
 +                const C&          _client_id,
 +                RequestRef&&      _request) :
 +        tag(_tag),
 +        client_id(_client_id),
 +        request(std::move(_request))
 +      {
 +        // empty
 +      }
 +
 +      friend std::ostream& operator<<(std::ostream& out, const ClientReq& c) {
 +        out << "{ ClientReq:: tag:" << c.tag << " client:" <<
 +          c.client_id << " }";
 +        return out;
 +      }
 +      }; // class ClientReq
 +
 +    public:
 +
 +      // NOTE: ClientRec is in the "public" section for compatibility
 +      // with g++ 4.8.4, which complains if it's not. By g++ 6.3.1
 +      // ClientRec could be "protected" with no issue. [See comments
 +      // associated with function submit_top_request.]
 +      class ClientRec {
 +      friend PriorityQueueBase<C,R,B>;
 +
 +      C                     client;
 +      RequestTag            prev_tag;
 +      std::deque<ClientReq> requests;
 +
 +      // amount added from the proportion tag as a result of
 +      // an idle client becoming unidle
 +      double                prop_delta = 0.0;
 +
 +      c::IndIntruHeapData   reserv_heap_data;
 +      c::IndIntruHeapData   lim_heap_data;
 +      c::IndIntruHeapData   ready_heap_data;
 +#if USE_PROP_HEAP
 +      c::IndIntruHeapData   prop_heap_data;
 +#endif
 +
 +      public:
 +
 +      ClientInfo            info;
 +      bool                  idle;
 +      Counter               last_tick;
 +      uint32_t              cur_rho;
 +      uint32_t              cur_delta;
 +
 +      ClientRec(C _client,
 +                const ClientInfo& _info,
 +                Counter current_tick) :
 +        client(_client),
 +        prev_tag(0.0, 0.0, 0.0, TimeZero),
 +        info(_info),
 +        idle(true),
 +        last_tick(current_tick),
 +        cur_rho(1),
 +        cur_delta(1)
 +      {
 +        // empty
 +      }
 +
 +      inline const RequestTag& get_req_tag() const {
 +        return prev_tag;
 +      }
 +
 +      static inline void assign_unpinned_tag(double& lhs, const double rhs) {
 +        if (rhs != max_tag && rhs != min_tag) {
 +          lhs = rhs;
 +        }
 +      }
 +
 +      inline void update_req_tag(const RequestTag& _prev,
 +                                 const Counter& _tick) {
 +        assign_unpinned_tag(prev_tag.reservation, _prev.reservation);
 +        assign_unpinned_tag(prev_tag.limit, _prev.limit);
 +        assign_unpinned_tag(prev_tag.proportion, _prev.proportion);
 +        last_tick = _tick;
 +      }
 +
 +      inline void add_request(const RequestTag& tag,
 +                              const C&          client_id,
 +                              RequestRef&&      request) {
 +        requests.emplace_back(ClientReq(tag, client_id, std::move(request)));
 +      }
 +
 +      inline const ClientReq& next_request() const {
 +        return requests.front();
 +      }
 +
 +      inline ClientReq& next_request() {
 +        return requests.front();
 +      }
 +
 +      inline void pop_request() {
 +        requests.pop_front();
 +      }
 +
 +      inline bool has_request() const {
 +        return !requests.empty();
 +      }
 +
 +      inline size_t request_count() const {
 +        return requests.size();
 +      }
 +
 +      // NB: because a deque is the underlying structure, this
 +      // operation might be expensive
 +      bool remove_by_req_filter_fw(std::function<bool(const R&)> filter_accum) {
 +        bool any_removed = false;
 +        for (auto i = requests.begin();
 +             i != requests.end();
 +             /* no inc */) {
 +          if (filter_accum(*i->request)) {
 +            any_removed = true;
 +            i = requests.erase(i);
 +          } else {
 +            ++i;
 +          }
 +        }
 +        return any_removed;
 +      }
 +
 +      // NB: because a deque is the underlying structure, this
 +      // operation might be expensive
 +      bool remove_by_req_filter_bw(std::function<bool(const R&)> filter_accum) {
 +        bool any_removed = false;
 +        for (auto i = requests.rbegin();
 +             i != requests.rend();
 +             /* no inc */) {
 +          if (filter_accum(*i->request)) {
 +            any_removed = true;
 +            i = decltype(i){ requests.erase(std::next(i).base()) };
 +          } else {
 +            ++i;
 +          }
 +        }
 +        return any_removed;
 +      }
 +
 +      inline bool
 +      remove_by_req_filter(std::function<bool(const R&)> filter_accum,
 +                           bool visit_backwards) {
 +        if (visit_backwards) {
 +          return remove_by_req_filter_bw(filter_accum);
 +        } else {
 +          return remove_by_req_filter_fw(filter_accum);
 +        }
 +      }
 +
 +      friend std::ostream&
 +      operator<<(std::ostream& out,
 +                 const typename PriorityQueueBase<C,R,B>::ClientRec& e) {
 +        out << "{ ClientRec::" <<
 +          " client:" << e.client <<
 +          " prev_tag:" << e.prev_tag <<
 +          " req_count:" << e.requests.size() <<
 +          " top_req:";
 +        if (e.has_request()) {
 +          out << e.next_request();
 +        } else {
 +          out << "none";
 +        }
 +        out << " }";
 +
 +        return out;
 +      }
 +      }; // class ClientRec
 +
 +      using ClientRecRef = std::shared_ptr<ClientRec>;
 +
 +      // when we try to get the next request, we'll be in one of three
 +      // situations -- we'll have one to return, have one that can
 +      // fire in the future, or not have any
 +      enum class NextReqType { returning, future, none };
 +
 +      // specifies which queue next request will get popped from
 +      enum class HeapId { reservation, ready };
 +
 +      // this is returned from next_req to tell the caller the situation
 +      struct NextReq {
 +      NextReqType type;
 +      union {
 +        HeapId    heap_id;
 +        Time      when_ready;
 +      };
 +      };
 +
 +
 +      // a function that can be called to look up client information
 +      using ClientInfoFunc = std::function<ClientInfo(const C&)>;
 +
 +
 +      bool empty() const {
 +      DataGuard g(data_mtx);
 +      return (resv_heap.empty() || ! resv_heap.top().has_request());
 +      }
 +
 +
 +      size_t client_count() const {
 +      DataGuard g(data_mtx);
 +      return resv_heap.size();
 +      }
 +
 +
 +      size_t request_count() const {
 +      DataGuard g(data_mtx);
 +      size_t total = 0;
 +      for (auto i = resv_heap.cbegin(); i != resv_heap.cend(); ++i) {
 +        total += i->request_count();
 +      }
 +      return total;
 +      }
 +
 +
 +      bool remove_by_req_filter(std::function<bool(const R&)> filter_accum,
 +                              bool visit_backwards = false) {
 +      bool any_removed = false;
 +      DataGuard g(data_mtx);
 +      for (auto i : client_map) {
 +        bool modified =
 +          i.second->remove_by_req_filter(filter_accum, visit_backwards);
 +        if (modified) {
 +          resv_heap.adjust(*i.second);
 +          limit_heap.adjust(*i.second);
 +          ready_heap.adjust(*i.second);
 +#if USE_PROP_HEAP
 +          prop_heap.adjust(*i.second);
 +#endif
 +          any_removed = true;
 +        }
 +      }
 +      return any_removed;
 +      }
 +
 +
 +      // use as a default value when no accumulator is provide
 +      static void request_sink(const R& req) {
 +      // do nothing
 +      }
 +
 +
 +      void remove_by_client(const C& client,
 +                          bool reverse = false,
 +                          std::function<void (const R&)> accum = request_sink) {
 +      DataGuard g(data_mtx);
 +
 +      auto i = client_map.find(client);
 +
 +      if (i == client_map.end()) return;
 +
 +      if (reverse) {
 +        for (auto j = i->second->requests.rbegin();
 +             j != i->second->requests.rend();
 +             ++j) {
 +          accum(*j->request);
 +        }
 +      } else {
 +        for (auto j = i->second->requests.begin();
 +             j != i->second->requests.end();
 +             ++j) {
 +          accum(*j->request);
 +        }
 +      }
 +
 +      i->second->requests.clear();
 +
 +      resv_heap.adjust(*i->second);
 +      limit_heap.adjust(*i->second);
 +      ready_heap.adjust(*i->second);
 +#if USE_PROP_HEAP
 +      prop_heap.adjust(*i->second);
 +#endif
 +      }
 +
 +
 +      uint get_heap_branching_factor() const {
 +      return B;
 +      }
 +
 +
 +      friend std::ostream& operator<<(std::ostream& out,
 +                                    const PriorityQueueBase& q) {
 +      std::lock_guard<decltype(q.data_mtx)> guard(q.data_mtx);
 +
 +      out << "{ PriorityQueue::";
 +      for (const auto& c : q.client_map) {
 +        out << "  { client:" << c.first << ", record:" << *c.second <<
 +          " }";
 +      }
 +      if (!q.resv_heap.empty()) {
 +        const auto& resv = q.resv_heap.top();
 +        out << " { reservation_top:" << resv << " }";
 +        const auto& ready = q.ready_heap.top();
 +        out << " { ready_top:" << ready << " }";
 +        const auto& limit = q.limit_heap.top();
 +        out << " { limit_top:" << limit << " }";
 +      } else {
 +        out << " HEAPS-EMPTY";
 +      }
 +      out << " }";
 +
 +      return out;
 +      }
 +
 +      // for debugging
 +      void display_queues(std::ostream& out,
 +                        bool show_res = true,
 +                        bool show_lim = true,
 +                        bool show_ready = true,
 +                        bool show_prop = true) const {
 +      auto filter = [](const ClientRec& e)->bool { return true; };
 +      DataGuard g(data_mtx);
 +      if (show_res) {
 +        resv_heap.display_sorted(out << "RESER:", filter);
 +      }
 +      if (show_lim) {
 +        limit_heap.display_sorted(out << "LIMIT:", filter);
 +      }
 +      if (show_ready) {
 +        ready_heap.display_sorted(out << "READY:", filter);
 +      }
 +#if USE_PROP_HEAP
 +      if (show_prop) {
 +        prop_heap.display_sorted(out << "PROPO:", filter);
 +      }
 +#endif
 +      } // display_queues
 +
 +
 +    protected:
 +
 +      // The ClientCompare functor is essentially doing a precedes?
 +      // operator, returning true if and only if the first parameter
 +      // must precede the second parameter. If the second must precede
 +      // the first, or if they are equivalent, false should be
 +      // returned. The reason for this behavior is that it will be
 +      // called to test if two items are out of order and if true is
 +      // returned it will reverse the items. Therefore false is the
 +      // default return when it doesn't matter to prevent unnecessary
 +      // re-ordering.
 +      //
 +      // The template is supporting variations in sorting based on the
 +      // heap in question and allowing these variations to be handled
 +      // at compile-time.
 +      //
 +      // tag_field determines which tag is being used for comparison
 +      //
 +      // ready_opt determines how the ready flag influences the sort
 +      //
 +      // use_prop_delta determines whether the proportional delta is
 +      // added in for comparison
 +      template<double RequestTag::*tag_field,
 +             ReadyOption ready_opt,
 +             bool use_prop_delta>
 +      struct ClientCompare {
 +      bool operator()(const ClientRec& n1, const ClientRec& n2) const {
 +        if (n1.has_request()) {
 +          if (n2.has_request()) {
 +            const auto& t1 = n1.next_request().tag;
 +            const auto& t2 = n2.next_request().tag;
 +            if (ReadyOption::ignore == ready_opt || t1.ready == t2.ready) {
 +              // if we don't care about ready or the ready values are the same
 +              if (use_prop_delta) {
 +                return (t1.*tag_field + n1.prop_delta) <
 +                  (t2.*tag_field + n2.prop_delta);
 +              } else {
 +                return t1.*tag_field < t2.*tag_field;
 +              }
 +            } else if (ReadyOption::raises == ready_opt) {
 +              // use_ready == true && the ready fields are different
 +              return t1.ready;
 +            } else {
 +              return t2.ready;
 +            }
 +          } else {
 +            // n1 has request but n2 does not
 +            return true;
 +          }
 +        } else if (n2.has_request()) {
 +          // n2 has request but n1 does not
 +          return false;
 +        } else {
 +          // both have none; keep stable w false
 +          return false;
 +        }
 +      }
 +      };
 +
 +      ClientInfoFunc       client_info_f;
 +
 +      mutable std::mutex data_mtx;
 +      using DataGuard = std::lock_guard<decltype(data_mtx)>;
 +
 +      // stable mapping between client ids and client queues
 +      std::map<C,ClientRecRef> client_map;
 +
 +      c::IndIntruHeap<ClientRecRef,
 +                    ClientRec,
 +                    &ClientRec::reserv_heap_data,
 +                    ClientCompare<&RequestTag::reservation,
 +                                  ReadyOption::ignore,
 +                                  false>,
 +                    B> resv_heap;
 +#if USE_PROP_HEAP
 +      c::IndIntruHeap<ClientRecRef,
 +                    ClientRec,
 +                    &ClientRec::prop_heap_data,
 +                    ClientCompare<&RequestTag::proportion,
 +                                  ReadyOption::ignore,
 +                                  true>,
 +                    B> prop_heap;
 +#endif
 +      c::IndIntruHeap<ClientRecRef,
 +                    ClientRec,
 +                    &ClientRec::lim_heap_data,
 +                    ClientCompare<&RequestTag::limit,
 +                                  ReadyOption::lowers,
 +                                  false>,
 +                    B> limit_heap;
 +      c::IndIntruHeap<ClientRecRef,
 +                    ClientRec,
 +                    &ClientRec::ready_heap_data,
 +                    ClientCompare<&RequestTag::proportion,
 +                                  ReadyOption::raises,
 +                                  true>,
 +                    B> ready_heap;
 +
 +      // if all reservations are met and all other requestes are under
 +      // limit, this will allow the request next in terms of
 +      // proportion to still get issued
 +      bool             allow_limit_break;
 +
 +      std::atomic_bool finishing;
 +
 +      // every request creates a tick
 +      Counter tick = 0;
 +
 +      // performance data collection
 +      size_t reserv_sched_count = 0;
 +      size_t prop_sched_count = 0;
 +      size_t limit_break_sched_count = 0;
 +
 +      Duration                  idle_age;
 +      Duration                  erase_age;
 +      Duration                  check_time;
 +      std::deque<MarkPoint>     clean_mark_points;
 +
 +      // NB: All threads declared at end, so they're destructed first!
 +
 +      std::unique_ptr<RunEvery> cleaning_job;
 +
 +
 +      // COMMON constructor that others feed into; we can accept three
 +      // different variations of durations
 +      template<typename Rep, typename Per>
 +      PriorityQueueBase(ClientInfoFunc _client_info_f,
 +                      std::chrono::duration<Rep,Per> _idle_age,
 +                      std::chrono::duration<Rep,Per> _erase_age,
 +                      std::chrono::duration<Rep,Per> _check_time,
 +                      bool _allow_limit_break) :
 +      client_info_f(_client_info_f),
 +      allow_limit_break(_allow_limit_break),
 +      finishing(false),
 +      idle_age(std::chrono::duration_cast<Duration>(_idle_age)),
 +      erase_age(std::chrono::duration_cast<Duration>(_erase_age)),
 +      check_time(std::chrono::duration_cast<Duration>(_check_time))
 +      {
 +      assert(_erase_age >= _idle_age);
 +      assert(_check_time < _idle_age);
 +      cleaning_job =
 +        std::unique_ptr<RunEvery>(
 +          new RunEvery(check_time,
 +                       std::bind(&PriorityQueueBase::do_clean, this)));
 +      }
 +
 +
 +      ~PriorityQueueBase() {
 +      finishing = true;
 +      }
 +
 +
 +      // data_mtx must be held by caller
-                         const Time       time,
-                         const double     cost = 0.0) {
++      void do_add_request(RequestRef&& request,
++                        const C& client_id,
 +                        const ReqParams& req_params,
-         tag = RequestTag(client.get_req_tag(), client.info,
-                          req_params, time, cost);
++                        const Time time,
++                        const double cost = 0.0) {
 +      ++tick;
 +
 +      // this pointer will help us create a reference to a shared
 +      // pointer, no matter which of two codepaths we take
 +      ClientRec* temp_client;
 +
 +      auto client_it = client_map.find(client_id);
 +      if (client_map.end() != client_it) {
 +        temp_client = &(*client_it->second); // address of obj of shared_ptr
 +      } else {
 +        ClientInfo info = client_info_f(client_id);
 +        ClientRecRef client_rec =
 +          std::make_shared<ClientRec>(client_id, info, tick);
 +        resv_heap.push(client_rec);
 +#if USE_PROP_HEAP
 +        prop_heap.push(client_rec);
 +#endif
 +        limit_heap.push(client_rec);
 +        ready_heap.push(client_rec);
 +        client_map[client_id] = client_rec;
 +        temp_client = &(*client_rec); // address of obj of shared_ptr
 +      }
 +
 +      // for convenience, we'll create a reference to the shared pointer
 +      ClientRec& client = *temp_client;
 +
 +      if (client.idle) {
 +        // We need to do an adjustment so that idle clients compete
 +        // fairly on proportional tags since those tags may have
 +        // drifted from real-time. Either use the lowest existing
 +        // proportion tag -- O(1) -- or the client with the lowest
 +        // previous proportion tag -- O(n) where n = # clients.
 +        //
 +        // So we don't have to maintain a propotional queue that
 +        // keeps the minimum on proportional tag alone (we're
 +        // instead using a ready queue), we'll have to check each
 +        // client.
 +        //
 +        // The alternative would be to maintain a proportional queue
 +        // (define USE_PROP_TAG) and do an O(1) operation here.
 +
 +        // Was unable to confirm whether equality testing on
 +        // std::numeric_limits<double>::max() is guaranteed, so
 +        // we'll use a compile-time calculated trigger that is one
 +        // third the max, which should be much larger than any
 +        // expected organic value.
 +        constexpr double lowest_prop_tag_trigger =
 +          std::numeric_limits<double>::max() / 3.0;
 +
 +        double lowest_prop_tag = std::numeric_limits<double>::max();
 +        for (auto const &c : client_map) {
 +          // don't use ourselves (or anything else that might be
 +          // listed as idle) since we're now in the map
 +          if (!c.second->idle) {
 +            double p;
 +            // use either lowest proportion tag or previous proportion tag
 +            if (c.second->has_request()) {
 +              p = c.second->next_request().tag.proportion +
 +                c.second->prop_delta;
 +            } else {
 +              p = c.second->get_req_tag().proportion + c.second->prop_delta;
 +            }
 +
 +            if (p < lowest_prop_tag) {
 +              lowest_prop_tag = p;
 +            }
 +          }
 +        }
 +
 +        // if this conditional does not fire, it
 +        if (lowest_prop_tag < lowest_prop_tag_trigger) {
 +          client.prop_delta = lowest_prop_tag - time;
 +        }
 +        client.idle = false;
 +      } // if this client was idle
 +
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      RequestTag tag(0, 0, 0, time);
 +
 +      if (!client.has_request()) {
-       ClientReq& first = top.next_request();
-       RequestRef request = std::move(first.request);
++        tag = RequestTag(client.get_req_tag(),
++                         client.info,
++                         req_params,
++                         time,
++                         cost);
 +
 +        // copy tag to previous tag for client
 +        client.update_req_tag(tag, tick);
 +      }
 +#else
 +      RequestTag tag(client.get_req_tag(), client.info, req_params, time, cost);
 +      // copy tag to previous tag for client
 +      client.update_req_tag(tag, tick);
 +#endif
 +
 +      client.add_request(tag, client.client, std::move(request));
 +      if (1 == client.requests.size()) {
 +        // NB: can the following 4 calls to adjust be changed
 +        // promote? Can adding a request ever demote a client in the
 +        // heaps?
 +        resv_heap.adjust(client);
 +        limit_heap.adjust(client);
 +        ready_heap.adjust(client);
 +#if USE_PROP_HEAP
 +        prop_heap.adjust(client);
 +#endif
 +      }
 +
 +      client.cur_rho = req_params.rho;
 +      client.cur_delta = req_params.delta;
 +
 +      resv_heap.adjust(client);
 +      limit_heap.adjust(client);
 +      ready_heap.adjust(client);
 +#if USE_PROP_HEAP
 +      prop_heap.adjust(client);
 +#endif
 +      } // add_request
 +
 +
 +      // data_mtx should be held when called; top of heap should have
 +      // a ready request
 +      template<typename C1, IndIntruHeapData ClientRec::*C2, typename C3>
 +      void pop_process_request(IndIntruHeap<C1, ClientRec, C2, C3, B>& heap,
 +                             std::function<void(const C& client,
 +                                                RequestRef& request)> process) {
 +      // gain access to data
 +      ClientRec& top = heap.top();
-         next_first.tag = RequestTag(first.tag, top.info,
-                                     ReqParams(top.cur_delta, top.cur_rho),
++
++      RequestRef request = std::move(top.next_request().request);
++      RequestTag tag = top.next_request().tag;
 +
 +      // pop request and adjust heaps
 +      top.pop_request();
 +
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +      if (top.has_request()) {
 +        ClientReq& next_first = top.next_request();
-                      const C&         client_id,
++        next_first.tag = RequestTag(tag, top.info,
++                                    top.cur_delta, top.cur_rho,
 +                                    next_first.tag.arrival);
 +
 +        // copy tag to previous tag for client
 +        top.update_req_tag(next_first.tag, tick);
 +      }
 +#endif
 +
 +      resv_heap.demote(top);
 +      limit_heap.adjust(top);
 +#if USE_PROP_HEAP
 +      prop_heap.demote(top);
 +#endif
 +      ready_heap.demote(top);
 +
 +      // process
 +      process(top.client, request);
 +      } // pop_process_request
 +
 +
 +      // data_mtx should be held when called
 +      void reduce_reservation_tags(ClientRec& client) {
 +      for (auto& r : client.requests) {
 +        r.tag.reservation -= client.info.reservation_inv;
 +
 +#ifndef DO_NOT_DELAY_TAG_CALC
 +        // reduce only for front tag. because next tags' value are invalid
 +        break;
 +#endif
 +      }
 +      // don't forget to update previous tag
 +      client.prev_tag.reservation -= client.info.reservation_inv;
 +      resv_heap.promote(client);
 +      }
 +
 +
 +      // data_mtx should be held when called
 +      void reduce_reservation_tags(const C& client_id) {
 +      auto client_it = client_map.find(client_id);
 +
 +      // means the client was cleaned from map; should never happen
 +      // as long as cleaning times are long enough
 +      assert(client_map.end() != client_it);
 +      reduce_reservation_tags(*client_it->second);
 +      }
 +
 +
 +      // data_mtx should be held when called
 +      NextReq do_next_request(Time now) {
 +      NextReq result;
 +
 +      // if reservation queue is empty, all are empty (i.e., no active clients)
 +      if(resv_heap.empty()) {
 +        result.type = NextReqType::none;
 +        return result;
 +      }
 +
 +      // try constraint (reservation) based scheduling
 +
 +      auto& reserv = resv_heap.top();
 +      if (reserv.has_request() &&
 +          reserv.next_request().tag.reservation <= now) {
 +        result.type = NextReqType::returning;
 +        result.heap_id = HeapId::reservation;
 +        return result;
 +      }
 +
 +      // no existing reservations before now, so try weight-based
 +      // scheduling
 +
 +      // all items that are within limit are eligible based on
 +      // priority
 +      auto limits = &limit_heap.top();
 +      while (limits->has_request() &&
 +             !limits->next_request().tag.ready &&
 +             limits->next_request().tag.limit <= now) {
 +        limits->next_request().tag.ready = true;
 +        ready_heap.promote(*limits);
 +        limit_heap.demote(*limits);
 +
 +        limits = &limit_heap.top();
 +      }
 +
 +      auto& readys = ready_heap.top();
 +      if (readys.has_request() &&
 +          readys.next_request().tag.ready &&
 +          readys.next_request().tag.proportion < max_tag) {
 +        result.type = NextReqType::returning;
 +        result.heap_id = HeapId::ready;
 +        return result;
 +      }
 +
 +      // if nothing is schedulable by reservation or
 +      // proportion/weight, and if we allow limit break, try to
 +      // schedule something with the lowest proportion tag or
 +      // alternatively lowest reservation tag.
 +      if (allow_limit_break) {
 +        if (readys.has_request() &&
 +            readys.next_request().tag.proportion < max_tag) {
 +          result.type = NextReqType::returning;
 +          result.heap_id = HeapId::ready;
 +          return result;
 +        } else if (reserv.has_request() &&
 +                   reserv.next_request().tag.reservation < max_tag) {
 +          result.type = NextReqType::returning;
 +          result.heap_id = HeapId::reservation;
 +          return result;
 +        }
 +      }
 +
 +      // nothing scheduled; make sure we re-run when next
 +      // reservation item or next limited item comes up
 +
 +      Time next_call = TimeMax;
 +      if (resv_heap.top().has_request()) {
 +        next_call =
 +          min_not_0_time(next_call,
 +                         resv_heap.top().next_request().tag.reservation);
 +      }
 +      if (limit_heap.top().has_request()) {
 +        const auto& next = limit_heap.top().next_request();
 +        assert(!next.tag.ready || max_tag == next.tag.proportion);
 +        next_call = min_not_0_time(next_call, next.tag.limit);
 +      }
 +      if (next_call < TimeMax) {
 +        result.type = NextReqType::future;
 +        result.when_ready = next_call;
 +        return result;
 +      } else {
 +        result.type = NextReqType::none;
 +        return result;
 +      }
 +      } // do_next_request
 +
 +
 +      // if possible is not zero and less than current then return it;
 +      // otherwise return current; the idea is we're trying to find
 +      // the minimal time but ignoring zero
 +      static inline const Time& min_not_0_time(const Time& current,
 +                                             const Time& possible) {
 +      return TimeZero == possible ? current : std::min(current, possible);
 +      }
 +
 +
 +      /*
 +       * This is being called regularly by RunEvery. Every time it's
 +       * called it notes the time and delta counter (mark point) in a
 +       * deque. It also looks at the deque to find the most recent
 +       * mark point that is older than clean_age. It then walks the
 +       * map and delete all server entries that were last used before
 +       * that mark point.
 +       */
 +      void do_clean() {
 +      TimePoint now = std::chrono::steady_clock::now();
 +      DataGuard g(data_mtx);
 +      clean_mark_points.emplace_back(MarkPoint(now, tick));
 +
 +      // first erase the super-old client records
 +
 +      Counter erase_point = 0;
 +      auto point = clean_mark_points.front();
 +      while (point.first <= now - erase_age) {
 +        erase_point = point.second;
 +        clean_mark_points.pop_front();
 +        point = clean_mark_points.front();
 +      }
 +
 +      Counter idle_point = 0;
 +      for (auto i : clean_mark_points) {
 +        if (i.first <= now - idle_age) {
 +          idle_point = i.second;
 +        } else {
 +          break;
 +        }
 +      }
 +
 +      if (erase_point > 0 || idle_point > 0) {
 +        for (auto i = client_map.begin(); i != client_map.end(); /* empty */) {
 +          auto i2 = i++;
 +          if (erase_point && i2->second->last_tick <= erase_point) {
 +            delete_from_heaps(i2->second);
 +            client_map.erase(i2);
 +          } else if (idle_point && i2->second->last_tick <= idle_point) {
 +            i2->second->idle = true;
 +          }
 +        } // for
 +      } // if
 +      } // do_clean
 +
 +
 +      // data_mtx must be held by caller
 +      template<IndIntruHeapData ClientRec::*C1,typename C2>
 +      void delete_from_heap(ClientRecRef& client,
 +                          c::IndIntruHeap<ClientRecRef,ClientRec,C1,C2,B>& heap) {
 +      auto i = heap.rfind(client);
 +      heap.remove(i);
 +      }
 +
 +
 +      // data_mtx must be held by caller
 +      void delete_from_heaps(ClientRecRef& client) {
 +      delete_from_heap(client, resv_heap);
 +#if USE_PROP_HEAP
 +      delete_from_heap(client, prop_heap);
 +#endif
 +      delete_from_heap(client, limit_heap);
 +      delete_from_heap(client, ready_heap);
 +      }
 +    }; // class PriorityQueueBase
 +
 +
 +    template<typename C, typename R, uint B=2>
 +    class PullPriorityQueue : public PriorityQueueBase<C,R,B> {
 +      using super = PriorityQueueBase<C,R,B>;
 +
 +    public:
 +
 +      // When a request is pulled, this is the return type.
 +      struct PullReq {
 +      struct Retn {
 +        C                           client;
 +        typename super::RequestRef  request;
 +        PhaseType                   phase;
 +      };
 +
 +      typename super::NextReqType   type;
 +      boost::variant<Retn,Time>     data;
 +
 +      bool is_none() const { return type == super::NextReqType::none; }
 +
 +      bool is_retn() const { return type == super::NextReqType::returning; }
 +      Retn& get_retn() {
 +        return boost::get<Retn>(data);
 +      }
 +
 +      bool is_future() const { return type == super::NextReqType::future; }
 +      Time getTime() const { return boost::get<Time>(data); }
 +      };
 +
 +
 +#ifdef PROFILE
 +      ProfileTimer<std::chrono::nanoseconds> pull_request_timer;
 +      ProfileTimer<std::chrono::nanoseconds> add_request_timer;
 +#endif
 +
 +      template<typename Rep, typename Per>
 +      PullPriorityQueue(typename super::ClientInfoFunc _client_info_f,
 +                      std::chrono::duration<Rep,Per> _idle_age,
 +                      std::chrono::duration<Rep,Per> _erase_age,
 +                      std::chrono::duration<Rep,Per> _check_time,
 +                      bool _allow_limit_break = false) :
 +      super(_client_info_f,
 +            _idle_age, _erase_age, _check_time,
 +            _allow_limit_break)
 +      {
 +      // empty
 +      }
 +
 +
 +      // pull convenience constructor
 +      PullPriorityQueue(typename super::ClientInfoFunc _client_info_f,
 +                      bool _allow_limit_break = false) :
 +      PullPriorityQueue(_client_info_f,
 +                        std::chrono::minutes(10),
 +                        std::chrono::minutes(15),
 +                        std::chrono::minutes(6),
 +                        _allow_limit_break)
 +      {
 +      // empty
 +      }
 +
 +
 +      inline void add_request(const R& request,
 +                            const C& client_id,
 +                            const ReqParams& req_params,
 +                            double addl_cost = 0.0) {
 +      add_request(typename super::RequestRef(new R(request)),
 +                  client_id,
 +                  req_params,
 +                  get_time(),
 +                  addl_cost);
 +      }
 +
 +
 +      inline void add_request(const R& request,
 +                            const C& client_id,
 +                            double addl_cost = 0.0) {
 +      static const ReqParams null_req_params;
 +      add_request(typename super::RequestRef(new R(request)),
 +                  client_id,
 +                  null_req_params,
 +                  get_time(),
 +                  addl_cost);
 +      }
 +
 +
 +
 +      inline void add_request_time(const R& request,
 +                                 const C& client_id,
 +                                 const ReqParams& req_params,
 +                                 const Time time,
 +                                 double addl_cost = 0.0) {
 +      add_request(typename super::RequestRef(new R(request)),
 +                  client_id,
 +                  req_params,
 +                  time,
 +                  addl_cost);
 +      }
 +
 +
 +      inline void add_request(typename super::RequestRef&& request,
 +                            const C& client_id,
 +                            const ReqParams& req_params,
 +                            double addl_cost = 0.0) {
 +      add_request(request, req_params, client_id, get_time(), addl_cost);
 +      }
 +
 +
 +      inline void add_request(typename super::RequestRef&& request,
 +                            const C& client_id,
 +                            double addl_cost = 0.0) {
 +      static const ReqParams null_req_params;
 +      add_request(request, null_req_params, client_id, get_time(), addl_cost);
 +      }
 +
 +
 +      // this does the work; the versions above provide alternate interfaces
 +      void add_request(typename super::RequestRef&& request,
 +                     const C&                     client_id,
 +                     const ReqParams&             req_params,
 +                     const Time                   time,
 +                     double                       addl_cost = 0.0) {
 +      typename super::DataGuard g(this->data_mtx);
 +#ifdef PROFILE
 +      add_request_timer.start();
 +#endif
 +      super::do_add_request(std::move(request),
 +                            client_id,
 +                            req_params,
 +                            time,
 +                            addl_cost);
 +      // no call to schedule_request for pull version
 +#ifdef PROFILE
 +      add_request_timer.stop();
 +#endif
 +      }
 +
 +
 +      inline PullReq pull_request() {
 +      return pull_request(get_time());
 +      }
 +
 +
 +      PullReq pull_request(Time now) {
 +      PullReq result;
 +      typename super::DataGuard g(this->data_mtx);
 +#ifdef PROFILE
 +      pull_request_timer.start();
 +#endif
 +
 +      typename super::NextReq next = super::do_next_request(now);
 +      result.type = next.type;
 +      switch(next.type) {
 +      case super::NextReqType::none:
 +        return result;
 +        break;
 +      case super::NextReqType::future:
 +        result.data = next.when_ready;
 +        return result;
 +        break;
 +      case super::NextReqType::returning:
 +        // to avoid nesting, break out and let code below handle this case
 +        break;
 +      default:
 +        assert(false);
 +      }
 +
 +      // we'll only get here if we're returning an entry
 +
 +      auto process_f =
 +        [&] (PullReq& pull_result, PhaseType phase) ->
 +        std::function<void(const C&,
 +                           typename super::RequestRef&)> {
 +        return [&pull_result, phase](const C& client,
 +                                     typename super::RequestRef& request) {
 +          pull_result.data =
 +          typename PullReq::Retn{client, std::move(request), phase};
 +        };
 +      };
 +
 +      switch(next.heap_id) {
 +      case super::HeapId::reservation:
 +        super::pop_process_request(this->resv_heap,
 +                                   process_f(result, PhaseType::reservation));
 +        ++this->reserv_sched_count;
 +        break;
 +      case super::HeapId::ready:
 +        super::pop_process_request(this->ready_heap,
 +                                   process_f(result, PhaseType::priority));
 +        { // need to use retn temporarily
 +          auto& retn = boost::get<typename PullReq::Retn>(result.data);
 +          super::reduce_reservation_tags(retn.client);
 +        }
 +        ++this->prop_sched_count;
 +        break;
 +      default:
 +        assert(false);
 +      }
 +
 +#ifdef PROFILE
 +      pull_request_timer.stop();
 +#endif
 +      return result;
 +      } // pull_request
 +
 +
 +    protected:
 +
 +
 +      // data_mtx should be held when called; unfortunately this
 +      // function has to be repeated in both push & pull
 +      // specializations
 +      typename super::NextReq next_request() {
 +      return next_request(get_time());
 +      }
 +    }; // class PullPriorityQueue
 +
 +
 +    // PUSH version
 +    template<typename C, typename R, uint B=2>
 +    class PushPriorityQueue : public PriorityQueueBase<C,R,B> {
 +
 +    protected:
 +
 +      using super = PriorityQueueBase<C,R,B>;
 +
 +    public:
 +
 +      // a function to see whether the server can handle another request
 +      using CanHandleRequestFunc = std::function<bool(void)>;
 +
 +      // a function to submit a request to the server; the second
 +      // parameter is a callback when it's completed
 +      using HandleRequestFunc =
 +      std::function<void(const C&,typename super::RequestRef,PhaseType)>;
 +
 +    protected:
 +
 +      CanHandleRequestFunc can_handle_f;
 +      HandleRequestFunc    handle_f;
 +      // for handling timed scheduling
 +      std::mutex  sched_ahead_mtx;
 +      std::condition_variable sched_ahead_cv;
 +      Time sched_ahead_when = TimeZero;
 +
 +#ifdef PROFILE
 +    public:
 +      ProfileTimer<std::chrono::nanoseconds> add_request_timer;
 +      ProfileTimer<std::chrono::nanoseconds> request_complete_timer;
 +    protected:
 +#endif
 +
 +      // NB: threads declared last, so constructed last and destructed first
 +
 +      std::thread sched_ahead_thd;
 +
 +    public:
 +
 +      // push full constructor
 +      template<typename Rep, typename Per>
 +      PushPriorityQueue(typename super::ClientInfoFunc _client_info_f,
 +                      CanHandleRequestFunc _can_handle_f,
 +                      HandleRequestFunc _handle_f,
 +                      std::chrono::duration<Rep,Per> _idle_age,
 +                      std::chrono::duration<Rep,Per> _erase_age,
 +                      std::chrono::duration<Rep,Per> _check_time,
 +                      bool _allow_limit_break = false) :
 +      super(_client_info_f,
 +            _idle_age, _erase_age, _check_time,
 +            _allow_limit_break)
 +      {
 +      can_handle_f = _can_handle_f;
 +      handle_f = _handle_f;
 +      sched_ahead_thd = std::thread(&PushPriorityQueue::run_sched_ahead, this);
 +      }
 +
 +
 +      // push convenience constructor
 +      PushPriorityQueue(typename super::ClientInfoFunc _client_info_f,
 +                      CanHandleRequestFunc _can_handle_f,
 +                      HandleRequestFunc _handle_f,
 +                      bool _allow_limit_break = false) :
 +      PushPriorityQueue(_client_info_f,
 +                        _can_handle_f,
 +                        _handle_f,
 +                        std::chrono::minutes(10),
 +                        std::chrono::minutes(15),
 +                        std::chrono::minutes(6),
 +                        _allow_limit_break)
 +      {
 +      // empty
 +      }
 +
 +
 +      ~PushPriorityQueue() {
 +      this->finishing = true;
 +      sched_ahead_cv.notify_one();
 +      sched_ahead_thd.join();
 +      }
 +
 +    public:
 +
 +      inline void add_request(const R& request,
 +                            const C& client_id,
 +                            const ReqParams& req_params,
 +                            double addl_cost = 0.0) {
 +      add_request(typename super::RequestRef(new R(request)),
 +                  client_id,
 +                  req_params,
 +                  get_time(),
 +                  addl_cost);
 +      }
 +
 +
 +      inline void add_request(typename super::RequestRef&& request,
 +                            const C& client_id,
 +                            const ReqParams& req_params,
 +                            double addl_cost = 0.0) {
 +      add_request(request, req_params, client_id, get_time(), addl_cost);
 +      }
 +
 +
 +      inline void add_request_time(const R& request,
 +                                 const C& client_id,
 +                                 const ReqParams& req_params,
 +                                 const Time time,
 +                                 double addl_cost = 0.0) {
 +      add_request(typename super::RequestRef(new R(request)),
 +                  client_id,
 +                  req_params,
 +                  time,
 +                  addl_cost);
 +      }
 +
 +
 +      void add_request(typename super::RequestRef&& request,
-                      const Time       time,
-                      double           addl_cost = 0.0) {
++                     const C& client_id,
 +                     const ReqParams& req_params,
++                     const Time time,
++                     double addl_cost = 0.0) {
 +      typename super::DataGuard g(this->data_mtx);
 +#ifdef PROFILE
 +      add_request_timer.start();
 +#endif
 +      super::do_add_request(std::move(request),
 +                            client_id,
 +                            req_params,
 +                            time,
 +                            addl_cost);
 +      schedule_request();
 +#ifdef PROFILE
 +      add_request_timer.stop();
 +#endif
 +      }
 +
 +
 +      void request_completed() {
 +      typename super::DataGuard g(this->data_mtx);
 +#ifdef PROFILE
 +      request_complete_timer.start();
 +#endif
 +      schedule_request();
 +#ifdef PROFILE
 +      request_complete_timer.stop();
 +#endif
 +      }
 +
 +    protected:
 +
 +      // data_mtx should be held when called; furthermore, the heap
 +      // should not be empty and the top element of the heap should
 +      // not be already handled
 +      //
 +      // NOTE: the use of "super::ClientRec" in either the template
 +      // construct or as a parameter to submit_top_request generated
 +      // a compiler error in g++ 4.8.4, when ClientRec was
 +      // "protected" rather than "public". By g++ 6.3.1 this was not
 +      // an issue. But for backwards compatibility
 +      // PriorityQueueBase::ClientRec is public.
 +      template<typename C1,
 +             IndIntruHeapData super::ClientRec::*C2,
 +             typename C3,
 +             uint B4>
 +      C submit_top_request(IndIntruHeap<C1,typename super::ClientRec,C2,C3,B4>& heap,
 +                         PhaseType phase) {
 +      C client_result;
 +      super::pop_process_request(heap,
 +                                 [this, phase, &client_result]
 +                                 (const C& client,
 +                                  typename super::RequestRef& request) {
 +                                   client_result = client;
 +                                   handle_f(client, std::move(request), phase);
 +                                 });
 +      return client_result;
 +      }
 +
 +
 +      // data_mtx should be held when called
 +      void submit_request(typename super::HeapId heap_id) {
 +      C client;
 +      switch(heap_id) {
 +      case super::HeapId::reservation:
 +        // don't need to note client
 +        (void) submit_top_request(this->resv_heap, PhaseType::reservation);
 +        // unlike the other two cases, we do not reduce reservation
 +        // tags here
 +        ++this->reserv_sched_count;
 +        break;
 +      case super::HeapId::ready:
 +        client = submit_top_request(this->ready_heap, PhaseType::priority);
 +        super::reduce_reservation_tags(client);
 +        ++this->prop_sched_count;
 +        break;
 +      default:
 +        assert(false);
 +      }
 +      } // submit_request
 +
 +
 +      // data_mtx should be held when called; unfortunately this
 +      // function has to be repeated in both push & pull
 +      // specializations
 +      typename super::NextReq next_request() {
 +      return next_request(get_time());
 +      }
 +
 +
 +      // data_mtx should be held when called; overrides member
 +      // function in base class to add check for whether a request can
 +      // be pushed to the server
 +      typename super::NextReq next_request(Time now) {
 +      if (!can_handle_f()) {
 +        typename super::NextReq result;
 +        result.type = super::NextReqType::none;
 +        return result;
 +      } else {
 +        return super::do_next_request(now);
 +      }
 +      } // next_request
 +
 +
 +      // data_mtx should be held when called
 +      void schedule_request() {
 +      typename super::NextReq next_req = next_request();
 +      switch (next_req.type) {
 +      case super::NextReqType::none:
 +        return;
 +      case super::NextReqType::future:
 +        sched_at(next_req.when_ready);
 +        break;
 +      case super::NextReqType::returning:
 +        submit_request(next_req.heap_id);
 +        break;
 +      default:
 +        assert(false);
 +      }
 +      }
 +
 +
 +      // this is the thread that handles running schedule_request at
 +      // future times when nothing can be scheduled immediately
 +      void run_sched_ahead() {
 +      std::unique_lock<std::mutex> l(sched_ahead_mtx);
 +
 +      while (!this->finishing) {
 +        if (TimeZero == sched_ahead_when) {
 +          sched_ahead_cv.wait(l);
 +        } else {
 +          Time now;
 +          while (!this->finishing && (now = get_time()) < sched_ahead_when) {
 +            long microseconds_l = long(1 + 1000000 * (sched_ahead_when - now));
 +            auto microseconds = std::chrono::microseconds(microseconds_l);
 +            sched_ahead_cv.wait_for(l, microseconds);
 +          }
 +          sched_ahead_when = TimeZero;
 +          if (this->finishing) return;
 +
 +          l.unlock();
 +          if (!this->finishing) {
 +            typename super::DataGuard g(this->data_mtx);
 +            schedule_request();
 +          }
 +          l.lock();
 +        }
 +      }
 +      }
 +
 +
 +      void sched_at(Time when) {
 +      std::lock_guard<std::mutex> l(sched_ahead_mtx);
++      if (this->finishing) return;
 +      if (TimeZero == sched_ahead_when || when < sched_ahead_when) {
 +        sched_ahead_when = when;
 +        sched_ahead_cv.notify_one();
 +      }
 +      }
 +    }; // class PushPriorityQueue
 +
 +  } // namespace dmclock
 +} // namespace crimson
index b6075bda22f92be2f5080b6a78fd0f7f4c741b74,0000000000000000000000000000000000000000..6342e29b16370dc1a5508a360b9fa1c31581428c
mode 100644,000000..100644
--- /dev/null
@@@ -1,549 -1,0 +1,549 @@@
-       remove(0);
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +
 +/*
 + * Copyright (C) 2016 Red Hat Inc.
 + */
 +
 +
 +#pragma once
 +
 +
 +#include <memory>
 +#include <vector>
 +#include <string>
 +#include <iostream>
 +#include <functional>
 +#include <algorithm>
 +
 +#include "assert.h"
 +
 +
 +namespace crimson {
 +  using IndIntruHeapData = size_t;
 +
 +  /* T is the ultimate data that's being stored in the heap, although
 +   *   through indirection.
 +   *
 +   * I is the indirect type that will actually be stored in the heap
 +   *   and that must allow dereferencing (via operator*) to yield a
 +   *   T&.
 +   *
 +   * C is a functor when given two T&'s will return true if the first
 +   *   must precede the second.
 +   *
 +   * heap_info is a data member pointer as to where the heap data in T
 +   * is stored.
 +   *
 +   * K is the branching factor of the heap, default is 2 (binary heap).
 +   */
 +  template<typename I,
 +         typename T,
 +         IndIntruHeapData T::*heap_info,
 +         typename C,
 +         uint K = 2>
 +  class IndIntruHeap {
 +
 +    // shorthand
 +    using HeapIndex = IndIntruHeapData;
 +
 +    static_assert(
 +      std::is_same<T,typename std::pointer_traits<I>::element_type>::value,
 +      "class I must resolve to class T by indirection (pointer dereference)");
 +
 +    static_assert(
 +      std::is_same<bool,
 +      typename std::result_of<C(const T&,const T&)>::type>::value,
 +      "class C must define operator() to take two const T& and return a bool");
 +
 +    static_assert(K >= 2, "K (degree of branching) must be at least 2");
 +
 +    class Iterator {
 +      friend IndIntruHeap<I, T, heap_info, C, K>;
 +
 +      IndIntruHeap<I, T, heap_info, C, K>& heap;
 +      HeapIndex                            index;
 +
 +      Iterator(IndIntruHeap<I, T, heap_info, C, K>& _heap, HeapIndex _index) :
 +      heap(_heap),
 +      index(_index)
 +      {
 +      // empty
 +      }
 +
 +    public:
 +
 +      Iterator(Iterator&& other) :
 +      heap(other.heap),
 +      index(other.index)
 +      {
 +      // empty
 +      }
 +
 +      Iterator(const Iterator& other) :
 +      heap(other.heap),
 +      index(other.index)
 +      {
 +      // empty
 +      }
 +
 +      Iterator& operator=(Iterator&& other) {
 +      std::swap(heap, other.heap);
 +      std::swap(index, other.index);
 +      return *this;
 +      }
 +
 +      Iterator& operator=(const Iterator& other) {
 +      heap = other.heap;
 +      index = other.index;
 +      }
 +
 +      Iterator& operator++() {
 +      if (index <= heap.count) {
 +        ++index;
 +      }
 +      return *this;
 +      }
 +
 +      bool operator==(const Iterator& other) const {
 +      return &heap == &other.heap && index == other.index;
 +      }
 +
 +      bool operator!=(const Iterator& other) const {
 +      return !(*this == other);
 +      }
 +
 +      T& operator*() {
 +      return *heap.data[index];
 +      }
 +
 +      T* operator->() {
 +      return &(*heap.data[index]);
 +      }
 +
 +#if 0
 +      // the item this iterator refers to
 +      void increase() {
 +      heap.sift_up(index);
 +      }
 +#endif
 +    }; // class Iterator
 +
 +
 +    class ConstIterator {
 +      friend IndIntruHeap<I, T, heap_info, C, K>;
 +
 +      const IndIntruHeap<I, T, heap_info, C, K>& heap;
 +      HeapIndex                                  index;
 +
 +      ConstIterator(const IndIntruHeap<I, T, heap_info, C, K>& _heap,
 +                  HeapIndex _index) :
 +      heap(_heap),
 +      index(_index)
 +      {
 +      // empty
 +      }
 +
 +    public:
 +
 +      ConstIterator(ConstIterator&& other) :
 +      heap(other.heap),
 +      index(other.index)
 +      {
 +      // empty
 +      }
 +
 +      ConstIterator(const ConstIterator& other) :
 +      heap(other.heap),
 +      index(other.index)
 +      {
 +      // empty
 +      }
 +
 +      ConstIterator& operator=(ConstIterator&& other) {
 +      std::swap(heap, other.heap);
 +      std::swap(index, other.index);
 +      return *this;
 +      }
 +
 +      ConstIterator& operator=(const ConstIterator& other) {
 +      heap = other.heap;
 +      index = other.index;
 +      }
 +
 +      ConstIterator& operator++() {
 +      if (index <= heap.count) {
 +        ++index;
 +      }
 +      return *this;
 +      }
 +
 +      bool operator==(const ConstIterator& other) const {
 +      return &heap == &other.heap && index == other.index;
 +      }
 +
 +      bool operator!=(const ConstIterator& other) const {
 +      return !(*this == other);
 +      }
 +
 +      const T& operator*() {
 +      return *heap.data[index];
 +      }
 +
 +      const T* operator->() {
 +      return &(*heap.data[index]);
 +      }
 +    }; // class ConstIterator
 +
 +
 +  protected:
 +
 +    std::vector<I> data;
 +    HeapIndex      count;
 +    C              comparator;
 +
 +  public:
 +
 +    IndIntruHeap() :
 +      count(0)
 +    {
 +      // empty
 +    }
 +
 +    IndIntruHeap(const IndIntruHeap<I,T,heap_info,C,K>& other) :
 +      count(other.count)
 +    {
 +      for (HeapIndex i = 0; i < other.count; ++i) {
 +      data.push_back(other.data[i]);
 +      }
 +    }
 +
 +    bool empty() const { return 0 == count; }
 +
 +    size_t size() const { return (size_t) count; }
 +
 +    T& top() { return *data[0]; }
 +
 +    const T& top() const { return *data[0]; }
 +
 +    I& top_ind() { return data[0]; }
 +
 +    const I& top_ind() const { return data[0]; }
 +
 +    void push(I&& item) {
 +      HeapIndex i = count++;
 +      intru_data_of(item) = i;
 +      data.emplace_back(std::move(item));
 +      sift_up(i);
 +    }
 +
 +    void push(const I& item) {
 +      I copy(item);
 +      push(std::move(copy));
 +    }
 +
 +    void pop() {
++      remove(HeapIndex(0));
 +    }
 +
 +    void remove(Iterator& i) {
 +      remove(i.index);
 +      i = end();
 +    }
 +
 +    Iterator find(const I& ind_item) {
 +      for (HeapIndex i = 0; i < count; ++i) {
 +      if (data[i] == ind_item) {
 +        return Iterator(*this, i);
 +      }
 +      }
 +      return end();
 +    }
 +
 +    // when passing in value we do a comparison via operator==
 +    Iterator find(const T& item) {
 +      for (HeapIndex i = 0; i < count; ++i) {
 +      if (*data[i] == item) {
 +        return Iterator(*this, i);
 +      }
 +      }
 +      return end();
 +    }
 +
 +    // reverse find -- start looking from bottom of heap
 +    Iterator rfind(const I& ind_item) {
 +      // HeapIndex is unsigned, so we can't allow to go negative; so
 +      // we'll keep it one more than actual index
 +      for (HeapIndex i = count; i > 0; --i) {
 +      if (data[i-1] == ind_item) {
 +        return Iterator(*this, i-1);
 +      }
 +      }
 +      return end();
 +    }
 +
 +    // reverse find -- start looking from bottom of heap
 +    Iterator rfind(const T& item) {
 +      // HeapIndex is unsigned, so we can't allow to go negative; so
 +      // we'll keep it one more than actual index
 +      for (HeapIndex i = count; i > 0; --i) {
 +      if (*data[i-1] == item) {
 +        return Iterator(*this, i-1);
 +      }
 +      }
 +      return end();
 +    }
 +
 +    ConstIterator find(const I& ind_item) const {
 +      for (HeapIndex i = 0; i < count; ++i) {
 +      if (data[i] == ind_item) {
 +        return ConstIterator(*this, i);
 +      }
 +      }
 +      return cend();
 +    }
 +
 +    // when passing in value we do a comparison via operator==
 +    ConstIterator find(const T& item) const {
 +      for (HeapIndex i = 0; i < count; ++i) {
 +      if (*data[i] == item) {
 +        return ConstIterator(*this, i);
 +      }
 +      }
 +      return cend();
 +    }
 +
 +    // reverse find -- start looking from bottom of heap
 +    ConstIterator rfind(const I& ind_item) const {
 +      // HeapIndex is unsigned, so we can't allow to go negative; so
 +      // we'll keep it one more than actual index
 +      for (HeapIndex i = count; i > 0; --i) {
 +      if (data[i-1] == ind_item) {
 +        return ConstIterator(*this, i-1);
 +      }
 +      }
 +      return cend();
 +    }
 +
 +    // reverse find -- start looking from bottom of heap
 +    ConstIterator rfind(const T& item) const {
 +      // HeapIndex is unsigned, so we can't allow to go negative; so
 +      // we'll keep it one more than actual index
 +      for (HeapIndex i = count; i > 0; --i) {
 +      if (*data[i-1] == item) {
 +        return ConstIterator(*this, i-1);
 +      }
 +      }
 +      return cend();
 +    }
 +
 +    void promote(T& item) {
 +      sift_up(item.*heap_info);
 +    }
 +
 +    void demote(T& item) {
 +      sift_down(item.*heap_info);
 +    }
 +
 +    void adjust(T& item) {
 +      sift(item.*heap_info);
 +    }
 +
 +    Iterator begin() {
 +      return Iterator(*this, 0);
 +    }
 +
 +    Iterator end() {
 +      return Iterator(*this, count);
 +    }
 +
 +    ConstIterator cbegin() const {
 +      return ConstIterator(*this, 0);
 +    }
 +
 +    ConstIterator cend() const {
 +      return ConstIterator(*this, count);
 +    }
 +
 +    friend std::ostream& operator<<(std::ostream& out, const IndIntruHeap& h) {
 +      auto i = h.data.cbegin();
 +      if (i != h.data.cend()) {
 +      out << **i;
 +      ++i;
 +      while (i != h.data.cend()) {
 +        out << ", " << **i;
 +      }
 +      }
 +      return out;
 +    }
 +
 +    // can only be called if I is copyable; copies heap into a vector
 +    // and sorts it before displaying it
 +    std::ostream&
 +    display_sorted(std::ostream& out,
 +                 std::function<bool(const T&)> filter = all_filter) const {
 +      static_assert(std::is_copy_constructible<I>::value,
 +                  "cannot call display_sorted when class I is not copy"
 +                  " constructible");
 +      auto compare = [this] (const I first, const I second) -> bool {
 +      return this->comparator(*first, *second);
 +      };
 +      std::vector<I> copy(data);
 +      std::sort(copy.begin(), copy.end(), compare);
 +
 +      bool first = true;
 +      for (auto c = copy.begin(); c != copy.end(); ++c) {
 +      if (filter(**c)) {
 +        if (!first) {
 +          out << ", ";
 +        } else {
 +          first = false;
 +        }
 +        out << **c;
 +      }
 +      }
 +
 +      return out;
 +    }
 +
 +
 +  protected:
 +
 +    static IndIntruHeapData& intru_data_of(I& item) {
 +      return (*item).*heap_info;
 +    }
 +
 +    void remove(HeapIndex i) {
 +      std::swap(data[i], data[--count]);
 +      intru_data_of(data[i]) = i;
 +      data.pop_back();
 +
 +      // the following needs to be sift (and not sift_down) as it can
 +      // go up or down the heap; imagine the heap vector contains 0,
 +      // 10, 100, 20, 30, 200, 300, 40; then 200 is removed, and 40
 +      // would have to be sifted upwards
 +      // sift(i);
 +      sift(i);
 +    }
 +
 +    // default value of filter parameter to display_sorted
 +    static bool all_filter(const T& data) { return true; }
 +
 +    // when i is negative?
 +    static inline HeapIndex parent(HeapIndex i) {
 +      assert(0 != i);
 +      return (i - 1) / K;
 +    }
 +
 +    // index of left child when K==2, index of left-most child when K>2
 +    static inline HeapIndex lhs(HeapIndex i) { return K*i + 1; }
 +
 +    // index of right child when K==2, index of right-most child when K>2
 +    static inline HeapIndex rhs(HeapIndex i) { return K*i + K; }
 +
 +    void sift_up(HeapIndex i) {
 +      while (i > 0) {
 +      HeapIndex pi = parent(i);
 +      if (!comparator(*data[i], *data[pi])) {
 +        break;
 +      }
 +
 +      std::swap(data[i], data[pi]);
 +      intru_data_of(data[i]) = i;
 +      intru_data_of(data[pi]) = pi;
 +      i = pi;
 +      }
 +    } // sift_up
 +
 +    // use this sift_down definition when K>2; it's more general and
 +    // uses a loop; EnableBool insures template uses a template
 +    // parameter
 +    template<bool EnableBool=true>
 +    typename std::enable_if<(K>2)&&EnableBool,void>::type sift_down(HeapIndex i) {
 +      if (i >= count) return;
 +      while (true) {
 +      HeapIndex li = lhs(i);
 +
 +      if (li < count) {
 +        HeapIndex ri = std::min(rhs(i), count - 1);
 +
 +        // find the index of min. child
 +        HeapIndex min_i = li;
 +        for (HeapIndex k = li + 1; k <= ri; ++k) {
 +          if (comparator(*data[k], *data[min_i])) {
 +            min_i = k;
 +          }
 +        }
 +
 +        if (comparator(*data[min_i], *data[i])) {
 +          std::swap(data[i], data[min_i]);
 +          intru_data_of(data[i]) = i;
 +          intru_data_of(data[min_i]) = min_i;
 +          i = min_i;
 +        } else {
 +          // no child is smaller
 +          break;
 +        }
 +      } else {
 +        // no children
 +        break;
 +      }
 +      }
 +    } // sift_down
 +
 +    // use this sift_down definition when K==2; EnableBool insures
 +    // template uses a template parameter
 +    template<bool EnableBool=true>
 +    typename std::enable_if<K==2&&EnableBool,void>::type sift_down(HeapIndex i) {
 +      if (i >= count) return;
 +      while (true) {
 +      const HeapIndex li = lhs(i);
 +      const HeapIndex ri = 1 + li;
 +
 +        if (li < count) {
 +        if (comparator(*data[li], *data[i])) {
 +          if (ri < count && comparator(*data[ri], *data[li])) {
 +            std::swap(data[i], data[ri]);
 +            intru_data_of(data[i]) = i;
 +            intru_data_of(data[ri]) = ri;
 +            i = ri;
 +          } else {
 +            std::swap(data[i], data[li]);
 +            intru_data_of(data[i]) = i;
 +            intru_data_of(data[li]) = li;
 +            i = li;
 +            }
 +        } else if (ri < count && comparator(*data[ri], *data[i])) {
 +          std::swap(data[i], data[ri]);
 +            intru_data_of(data[i]) = i;
 +          intru_data_of(data[ri]) = ri;
 +          i = ri;
 +          } else {
 +          // no child is smaller
 +            break;
 +          }
 +        } else {
 +        // no children
 +          break;
 +        }
 +      } // while
 +    } // sift_down
 +
 +    void sift(HeapIndex i) {
 +      if (i == 0) {
 +      // if we're at top, can only go down
 +      sift_down(i);
 +      } else {
 +      HeapIndex pi = parent(i);
 +      if (comparator(*data[i], *data[pi])) {
 +        // if we can go up, we will
 +        sift_up(i);
 +      } else {
 +        // otherwise we'll try to go down
 +        sift_down(i);
 +      }
 +      }
 +    } // sift
 +  }; // class IndIntruHeap
 +
 +} // namespace crimson
index 258baaa74c041bf3b6d0159b85c1135a82009fa5,0000000000000000000000000000000000000000..cab414109a9dc6fcc648fff7873d432961d29cdb
mode 100644,000000..100644
--- /dev/null
@@@ -1,73 -1,0 +1,80 @@@
-  * Copyright (C) 2016 Red Hat Inc.
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +/*
- #include <iostream>
++ * Copyright (C) 2017 Red Hat Inc.
 + */
 +
 +
-   finishing = true;
-   cv.notify_all();
 +#include "run_every.h"
 +
 +
 +// can define ADD_MOVE_SEMANTICS, although not fully debugged and tested
 +
 +
 +namespace chrono = std::chrono;
 +
 +
 +#ifdef ADD_MOVE_SEMANTICS
 +crimson::RunEvery::RunEvery()
 +{
 +  // empty
 +}
 +
 +
 +crimson::RunEvery& crimson::RunEvery::operator=(crimson::RunEvery&& other)
 +{
 +  // finish run every thread
 +  {
 +    Guard g(mtx);
 +    finishing = true;
 +    cv.notify_one();
 +  }
 +  if (thd.joinable()) {
 +    thd.join();
 +  }
 +
 +  // transfer info over from previous thread
 +  finishing.store(other.finishing);
 +  wait_period = other.wait_period;
 +  body = other.body;
 +
 +  // finish other thread
 +  other.finishing.store(true);
 +  other.cv.notify_one();
 +
 +  // start this thread
 +  thd = std::thread(&RunEvery::run, this);
 +
 +  return *this;
 +}
 +#endif
 +
 +
 +crimson::RunEvery::~RunEvery() {
++  join();
++}
++
++
++void crimson::RunEvery::join() {
++  {
++    Guard l(mtx);
++    if (finishing) return;
++    finishing = true;
++    cv.notify_all();
++  }
 +  thd.join();
 +}
 +
 +
 +void crimson::RunEvery::run() {
 +  Lock l(mtx);
 +  while(!finishing) {
 +    TimePoint until = chrono::steady_clock::now() + wait_period;
 +    while (!finishing && chrono::steady_clock::now() < until) {
 +      cv.wait_until(l, until);
 +    }
 +    if (!finishing) {
 +      body();
 +    }
 +  }
 +}
index 58b85563e7eb0a4a3753987a450b04ee9b7fd527,0000000000000000000000000000000000000000..0d4c538577971f032f5d5575c9521fa963429b59
mode 100644,000000..100644
--- /dev/null
@@@ -1,68 -1,0 +1,70 @@@
-  * Copyright (C) 2016 Red Hat Inc.
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +/*
++ * Copyright (C) 2017 Red Hat Inc.
 + */
 +
 +
 +#pragma once
 +
 +#include <chrono>
 +#include <mutex>
 +#include <condition_variable>
 +#include <thread>
 +#include <functional>
 +
 +namespace crimson {
 +  using std::chrono::duration_cast;
 +  using std::chrono::milliseconds;
 +
 +  // runs a given simple function object waiting wait_period
 +  // milliseconds between; the destructor stops the other thread
 +  // immediately
 +  class RunEvery {
 +    using Lock      = std::unique_lock<std::mutex>;
 +    using Guard     = std::lock_guard<std::mutex>;
 +    using TimePoint = std::chrono::steady_clock::time_point;
 +
 +    bool                      finishing = false;
 +    std::chrono::milliseconds wait_period;
 +    std::function<void()>     body;
 +    std::mutex                mtx;
 +    std::condition_variable   cv;
 +
 +    // put threads last so all other variables are initialized first
 +
 +    std::thread               thd;
 +
 +  public:
 +
 +#ifdef ADD_MOVE_SEMANTICS
 +    RunEvery();
 +#endif
 +
 +    template<typename D>
 +    RunEvery(D                     _wait_period,
 +           std::function<void()> _body) :
 +      wait_period(duration_cast<milliseconds>(_wait_period)),
 +      body(_body)
 +    {
 +      thd = std::thread(&RunEvery::run, this);
 +    }
 +
 +    RunEvery(const RunEvery& other) = delete;
 +    RunEvery& operator=(const RunEvery& other) = delete;
 +    RunEvery(RunEvery&& other) = delete;
 +#ifdef ADD_MOVE_SEMANTICS
 +    RunEvery& operator=(RunEvery&& other);
 +#else
 +    RunEvery& operator=(RunEvery&& other) = delete;
 +#endif
 +
 +    ~RunEvery();
 +
++    void join();
++
 +  protected:
 +
 +    void run();
 +  };
 +}
index addea6c96a996c607345c6139674705142084bb6,0000000000000000000000000000000000000000..1e1ea25cc14e3dcaac1c2bca421a82e8eab025c2
mode 100644,000000..100644
--- /dev/null
@@@ -1,29 -1,0 +1,26 @@@
- add_test(NAME dmclock-data-struct-tests
-   COMMAND $<TARGET_FILE:dmclock-data-struct-tests>)
 +include_directories(../src)
 +
 +set(local_flags "-Wall -pthread")
 +
 +# dmclock does not use intrusive heap (but it does use indirect
 +# intrusive heap), so we won't use this code
 +if(false)
 +  set(srcs
 +    test_intrusive_heap.cc)
 +  add_executable(test_intru_heap test_intrusive_heap.cc)
 +  set_source_files_properties(${srcs}
 +    PROPERTIES
 +    COMPILE_FLAGS "${local_flags}")
 +endif(false)
 +
 +set(test_srcs test_indirect_intrusive_heap.cc)
 +
 +set_source_files_properties(${test_srcs}
 +  PROPERTIES
 +  COMPILE_FLAGS "${local_flags}"
 +  )
 +
 +add_executable(dmclock-data-struct-tests ${test_srcs})
 +
 +target_link_libraries(dmclock-data-struct-tests
 +  LINK_PRIVATE gtest gtest_main pthread)
index e72810b56aa73feeaaf738922f50fba44d853e13,0000000000000000000000000000000000000000..24da2e7320a5d02343355ee442b825a424bdbf17
mode 100644,000000..100644
--- /dev/null
@@@ -1,35 -1,0 +1,31 @@@
- include_directories(${BOOST_INCLUDE_DIR})
 +include_directories(../src)
 +include_directories(../support/src)
 +include_directories(../sim/src)
-     LINK_PRIVATE $<TARGET_FILE:dmclock> pthread ${GTEST_LIBRARY} ${GTEST_MAIN_LIBRARY})
 +
 +set(support_srcs ../sim/src/test_dmclock.cc)
 +set(test_srcs
 +  test_test_client.cc
 +  test_dmclock_server.cc
 +  test_dmclock_client.cc
 +  )
 +
 +set_source_files_properties(${core_srcs} ${test_srcs}
 +  PROPERTIES
 +  COMPILE_FLAGS "${local_flags}"
 +  )
 +
 +add_executable(dmclock-tests ${test_srcs} ${support_srcs})
 +
 +if (TARGET gtest AND TARGET gtest_main)
 +  add_dependencies(dmclock-tests gtest gtest_main)
 +  target_link_libraries(dmclock-tests
 +    LINK_PRIVATE $<TARGET_FILE:dmclock>
 +    pthread
 +    $<TARGET_FILE:gtest>
 +    $<TARGET_FILE:gtest_main>)
 +else()
 +  target_link_libraries(dmclock-tests
- add_test(NAME dmclock-tests
-   COMMAND $<TARGET_FILE:dmclock-tests>)
++    LINK_PRIVATE $<TARGET_FILE:dmclock> pthread ${GTEST_LIBRARIES} ${GTEST_MAIN_LIBRARIES})
 +endif()
 +  
 +add_dependencies(dmclock-tests dmclock)
index 6015cb9bf7bb020a004eaa8022266d1878fd50e7,0000000000000000000000000000000000000000..dd04fd179af8f5a325454447f1eef66be0ab4227
mode 100644,000000..100644
--- /dev/null
@@@ -1,123 -1,0 +1,127 @@@
 +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
 +// vim: ts=8 sw=2 smarttab
 +/*
 + * Copyright (C) 2016 Red Hat Inc.
 + */
 +
 +#include <atomic>
 +#include <thread>
 +#include <chrono>
 +#include <iostream>
 +
 +#include "gtest/gtest.h"
 +
 +#include "sim_recs.h"
 +#include "sim_client.h"
 +
 +#include "test_dmclock.h"
 +
 +
 +using namespace std::placeholders;
 +
 +namespace dmc = crimson::dmclock;
 +namespace test = crimson::test_dmc;
 +namespace sim = crimson::qos_simulation;
 +
 +using TimePoint = std::chrono::time_point<std::chrono::system_clock>;
 +
 +static TimePoint now() { return std::chrono::system_clock::now(); }
 +
 +
 +TEST(test_client, full_bore_timing) {
 +  std::atomic_ulong count(0);
 +
 +  ServerId server_id = 3;
 +
 +  sim::TestResponse resp(0);
 +  dmc::PhaseType resp_params = dmc::PhaseType::priority;
 +  test::DmcClient* client;
 +
 +  auto start = now();
 +  client =
 +    new test::DmcClient(ClientId(0),
 +                      [&] (const ServerId& server,
 +                           const sim::TestRequest& req,
 +                           const ClientId& client_id,
 +                           const dmc::ReqParams& req_params) {
 +                        ++count;
 +                        client->receive_response(resp, client_id, resp_params);
 +                      },
 +                      [&] (const uint64_t seed) -> ServerId& {
 +                        return server_id;
 +                      },
 +                      test::dmc_client_accumulate_f,
 +                      1000, // ops to run
 +                      100, // iops goal
 +                      5); // outstanding ops allowed
 +  client->wait_until_done();
 +  auto end = now();
 +  EXPECT_EQ(1000u, count) << "didn't get right number of ops";
 +
 +  int milliseconds = (end - start) / std::chrono::milliseconds(1);
 +  EXPECT_LT(10000, milliseconds) << "timing too fast to be correct";
 +  EXPECT_GT(12000, milliseconds) << "timing suspiciously slow";
++
++  delete client;
 +}
 +
 +
 +TEST(test_client, paused_timing) {
 +  std::atomic_ulong count(0);
 +  std::atomic_ulong unresponded_count(0);
 +  std::atomic_bool auto_respond(false);
 +
 +  ClientId my_client_id = 0;
 +  ServerId server_id = 3;
 +
 +  sim::TestResponse resp(0);
 +  dmc::PhaseType resp_params = dmc::PhaseType::priority;
 +  test::DmcClient* client;
 +
 +  auto start = now();
 +  client =
 +    new test::DmcClient(my_client_id,
 +                      [&] (const ServerId& server,
 +                           const sim::TestRequest& req,
 +                           const ClientId& client_id,
 +                           const dmc::ReqParams& req_params) {
 +                        ++count;
 +                        if (auto_respond.load()) {
 +                          client->receive_response(resp, client_id, resp_params);
 +                        } else {
 +                          ++unresponded_count;
 +                        }
 +                      },
 +                      [&] (const uint64_t seed) -> ServerId& {
 +                        return server_id;
 +                      },
 +                      test::dmc_client_accumulate_f,
 +
 +                      1000, // ops to run
 +                      100, // iops goal
 +                      50); // outstanding ops allowed
 +  std::thread t([&]() {
 +      std::this_thread::sleep_for(std::chrono::seconds(5));
 +      EXPECT_EQ(50u, unresponded_count.load()) <<
 +      "should have 50 unresponded calls";
 +      auto_respond = true;
 +      // respond to those 50 calls
 +      for(int i = 0; i < 50; ++i) {
 +      client->receive_response(resp, my_client_id, resp_params);
 +      --unresponded_count;
 +      }
 +    });
 +
 +  client->wait_until_done();
 +  auto end = now();
 +  int milliseconds = (end - start) / std::chrono::milliseconds(1);
 +
 +  // the 50 outstanding ops allowed means the first half-second of
 +  // requests get responded to during the 5 second pause. So we have
 +  // to adjust our expectations by a half-second.
 +  EXPECT_LT(15000 - 500, milliseconds) << "timing too fast to be correct";
 +  EXPECT_GT(17000 - 500, milliseconds) << "timing suspiciously slow";
 +  t.join();
++
++  delete client;
 +}