From 1b8a87b39a52b10ac3b1bd3c70062d4dcd53ed5e Mon Sep 17 00:00:00 2001
From: Mark Nelson <mnelson@redhat.com>
Date: Tue, 19 Jun 2018 21:59:26 -0500
Subject: [PATCH] os/bluestore: limit OSD memory usage by tuning the cache
 size.

Signed-off-by: Mark Nelson <mnelson@redhat.com>
---
 src/common/options.cc         |  45 +++++++--
 src/os/CMakeLists.txt         |   5 +-
 src/os/bluestore/BlueStore.cc | 166 +++++++++++++++++++++++++---------
 src/os/bluestore/BlueStore.h  |  15 +--
 4 files changed, 173 insertions(+), 58 deletions(-)

diff --git a/src/common/options.cc b/src/common/options.cc
index 52c1ec002458..b3be001a2756 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -3681,6 +3681,31 @@ std::vector<Option> get_global_options() {
     .set_default(true)
     .set_description(""),
 
+    Option("osd_memory_target", Option::TYPE_UINT, Option::LEVEL_BASIC)
+    .set_default(4_G)
+    .add_see_also("bluestore_cache_autotune")
+    .set_description("When tcmalloc and cache autotuning is enabled, try to keep this many bytes mapped in memory."),
+
+    Option("osd_memory_base", Option::TYPE_UINT, Option::LEVEL_DEV)
+    .set_default(768_M)
+    .add_see_also("bluestore_cache_autotune")
+    .set_description("When tcmalloc and cache autotuning is enabled, estimate the minimum amount of memory in bytes the OSD will need."),
+
+    Option("osd_memory_expected_fragmentation", Option::TYPE_FLOAT, Option::LEVEL_DEV)
+    .set_default(0.15)
+    .add_see_also("bluestore_cache_autotune")
+    .set_description("When tcmalloc and cache autotuning is enabled, estimate the percent of memory fragmentation."),
+
+    Option("osd_memory_cache_min", Option::TYPE_UINT, Option::LEVEL_DEV)
+    .set_default(128_M)
+    .add_see_also("bluestore_cache_autotune")
+    .set_description("When tcmalloc and cache autotuning is enabled, set the minimum amount of memory used for caches."),
+
+    Option("osd_memory_cache_resize_interval", Option::TYPE_FLOAT, Option::LEVEL_DEV)
+    .set_default(1)
+    .add_see_also("bluestore_cache_autotune")
+    .set_description("When tcmalloc and cache autotuning is enabled, wait this many seconds between resizing caches."),
+
     Option("memstore_device_bytes", Option::TYPE_SIZE, Option::LEVEL_ADVANCED)
     .set_default(1_G)
     .set_description(""),
@@ -4081,41 +4106,41 @@ std::vector<Option> get_global_options() {
     .set_default(.5)
     .set_description("2Q paper suggests .5"),
 
-    Option("bluestore_cache_size", Option::TYPE_SIZE, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_size", Option::TYPE_SIZE, Option::LEVEL_DEV)
     .set_default(0)
     .set_description("Cache size (in bytes) for BlueStore")
     .set_long_description("This includes data and metadata cached by BlueStore as well as memory devoted to rocksdb's cache(s)."),
 
-    Option("bluestore_cache_size_hdd", Option::TYPE_SIZE, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_size_hdd", Option::TYPE_SIZE, Option::LEVEL_DEV)
     .set_default(1_G)
     .set_description("Default bluestore_cache_size for rotational media"),
 
-    Option("bluestore_cache_size_ssd", Option::TYPE_SIZE, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_size_ssd", Option::TYPE_SIZE, Option::LEVEL_DEV)
     .set_default(3_G)
     .set_description("Default bluestore_cache_size for non-rotational (solid state) media"),
 
-    Option("bluestore_cache_meta_ratio", Option::TYPE_FLOAT, Option::LEVEL_ADVANCED)
-    .set_default(.5)
+    Option("bluestore_cache_meta_ratio", Option::TYPE_FLOAT, Option::LEVEL_DEV)
+    .set_default(.4)
     .add_see_also("bluestore_cache_size")
     .set_description("Ratio of bluestore cache to devote to metadata"),
 
-    Option("bluestore_cache_kv_ratio", Option::TYPE_FLOAT, Option::LEVEL_ADVANCED)
-    .set_default(.5)
+    Option("bluestore_cache_kv_ratio", Option::TYPE_FLOAT, Option::LEVEL_DEV)
+    .set_default(.4)
     .add_see_also("bluestore_cache_size")
     .set_description("Ratio of bluestore cache to devote to kv database (rocksdb)"),
 
-    Option("bluestore_cache_autotune", Option::TYPE_BOOL, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_autotune", Option::TYPE_BOOL, Option::LEVEL_DEV)
     .set_default(true)
     .add_see_also("bluestore_cache_size")
     .add_see_also("bluestore_cache_meta_ratio")
     .set_description("Automatically tune the ratio of caches while respecting min values."),
 
-    Option("bluestore_cache_autotune_chunk_size", Option::TYPE_SIZE, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_autotune_chunk_size", Option::TYPE_SIZE, Option::LEVEL_DEV)
     .set_default(33554432)
     .add_see_also("bluestore_cache_autotune")
     .set_description("The chunk size in bytes to allocate to caches when cache autotune is enabled."),
 
-    Option("bluestore_cache_autotune_interval", Option::TYPE_FLOAT, Option::LEVEL_ADVANCED)
+    Option("bluestore_cache_autotune_interval", Option::TYPE_FLOAT, Option::LEVEL_DEV)
     .set_default(5)
     .add_see_also("bluestore_cache_autotune")
     .set_description("The number of seconds to wait between rebalances when cache autotune is enabled."),
diff --git a/src/os/CMakeLists.txt b/src/os/CMakeLists.txt
index 17f78b71204f..9456dd2c7428 100644
--- a/src/os/CMakeLists.txt
+++ b/src/os/CMakeLists.txt
@@ -71,7 +71,10 @@ if(WITH_SPDK)
     bluestore/NVMEDevice.cc)
 endif()
 
-add_library(os STATIC ${libos_srcs} $<TARGET_OBJECTS:kv_objs>)
+add_library(os STATIC ${libos_srcs}
+  $<TARGET_OBJECTS:kv_objs>)
+
+target_link_libraries(os heap_profiler)
 
 if(WITH_BLUEFS)
   add_library(bluefs SHARED 
diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 42318ab81114..bced06dac8e4 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -35,6 +35,7 @@
 #include "BlueRocksEnv.h"
 #include "auth/Crypto.h"
 #include "common/EventTrace.h"
+#include "perfglue/heap_profiler.h"
 
 #define dout_context cct
 #define dout_subsys ceph_subsys_bluestore
@@ -3429,26 +3430,45 @@ void *BlueStore::MempoolThread::entry()
   caches.push_back(store->db);
   caches.push_back(&meta_cache);
   caches.push_back(&data_cache);
+  autotune_cache_size = store->osd_memory_cache_min;
 
   utime_t next_balance = ceph_clock_now();
+  utime_t next_resize = ceph_clock_now();
+
+  bool interval_stats_trim = false;
+  bool interval_stats_resize = false; 
   while (!stop) {
     _adjust_cache_settings();
 
-    // Before we trim, check and see if it's time to rebalance
-    bool log_stats = false;
+    // Before we trim, check and see if it's time to rebalance/resize.
     double autotune_interval = store->cache_autotune_interval;
+    double resize_interval = store->osd_memory_cache_resize_interval;
+
     if (autotune_interval > 0 && next_balance < ceph_clock_now()) {
+      // Log events at 5 instead of 20 when balance happens.
+      interval_stats_resize = true; 
+      interval_stats_trim = true;
       if (store->cache_autotune) {
         _balance_cache(caches);
       }
+
       next_balance = ceph_clock_now();
       next_balance += autotune_interval;
-      log_stats = true;
+    }
+    if (resize_interval > 0 && next_resize < ceph_clock_now()) {
+      if (ceph_using_tcmalloc() && store->cache_autotune) {
+        _tune_cache_size(interval_stats_resize);
+        interval_stats_resize = false;
+      }
+      next_resize = ceph_clock_now();
+      next_resize += resize_interval;
     }
 
-    _trim_shards(log_stats);
-    store->_update_cache_logger();
+    // Now Trim
+    _trim_shards(interval_stats_trim);
+    interval_stats_trim = false;
 
+    store->_update_cache_logger();
     utime_t wait;
     wait += store->cct->_conf->bluestore_cache_trim_interval;
     cond.WaitInterval(lock, wait);
@@ -3464,59 +3484,116 @@ void BlueStore::MempoolThread::_adjust_cache_settings()
   data_cache.set_cache_ratio(store->cache_data_ratio);
 }
 
-void BlueStore::MempoolThread::_trim_shards(bool log_stats)
+void BlueStore::MempoolThread::_trim_shards(bool interval_stats)
 {
-  uint64_t cache_size = store->cache_size;
+  auto cct = store->cct;
   size_t num_shards = store->cache_shards.size();
-  int64_t kv_alloc_bytes = 0;
-  int64_t meta_alloc_bytes = 0;
-  int64_t data_alloc_bytes = 0;
+
+  int64_t kv_used = store->db->get_cache_usage();
+  int64_t meta_used = meta_cache._get_used_bytes();
+  int64_t data_used = data_cache._get_used_bytes();
+
+  uint64_t cache_size = store->cache_size;
+  int64_t kv_alloc =
+     static_cast<int64_t>(store->db->get_cache_ratio() * cache_size); 
+  int64_t meta_alloc =
+     static_cast<int64_t>(meta_cache.get_cache_ratio() * cache_size);
+  int64_t data_alloc =
+     static_cast<int64_t>(data_cache.get_cache_ratio() * cache_size);
 
   if (store->cache_autotune) {
-    kv_alloc_bytes = store->db->get_cache_bytes();
-    meta_alloc_bytes = meta_cache.get_cache_bytes();
-    data_alloc_bytes = data_cache.get_cache_bytes();
+    cache_size = autotune_cache_size;
+
+    kv_alloc = store->db->get_cache_bytes();
+    meta_alloc = meta_cache.get_cache_bytes();
+    data_alloc = data_cache.get_cache_bytes();
+  }
+  
+  if (interval_stats) {
+    ldout(cct, 5) << __func__  << " cache_size: " << cache_size
+                  << " kv_alloc: " << kv_alloc
+                  << " kv_used: " << kv_used
+                  << " meta_alloc: " << meta_alloc
+                  << " meta_used: " << meta_used
+                  << " data_alloc: " << data_alloc
+                  << " data_used: " << data_used << dendl;
   } else {
-    kv_alloc_bytes = static_cast<int64_t>(
-        store->db->get_cache_ratio() * cache_size);
-    meta_alloc_bytes = static_cast<int64_t>(
-        meta_cache.get_cache_ratio() * cache_size);
-    data_alloc_bytes = static_cast<int64_t>(
-        data_cache.get_cache_ratio() * cache_size);
-  }
-  if (log_stats) {
-    double kv_alloc_ratio = (double) kv_alloc_bytes / cache_size;
-    double meta_alloc_ratio = (double) meta_alloc_bytes / cache_size;
-    double data_alloc_ratio = (double) data_alloc_bytes / cache_size;
-    double kv_used_ratio = (double) store->db->get_cache_usage() / cache_size;
-    double meta_used_ratio = (double) meta_cache._get_used_bytes() / cache_size;
-    double data_used_ratio = (double) data_cache._get_used_bytes() / cache_size;
-
-    ldout(store->cct, 5) << __func__ << " ratios -" << std::fixed << std::setprecision(1)
-                         << " kv_alloc: " << 100*kv_alloc_ratio << "%"
-                         << " kv_used: " << 100*kv_used_ratio << "%"
-                         << " meta_alloc: " << 100*meta_alloc_ratio << "%"
-                         << " meta_used: " << 100*meta_used_ratio << "%"
-                         << " data_alloc: " << 100*data_alloc_ratio << "%"
-                         << " data_used: " << 100*data_used_ratio << "%" << dendl;
+    ldout(cct, 20) << __func__  << " cache_size: " << cache_size
+                   << " kv_alloc: " << kv_alloc
+                   << " kv_used: " << kv_used
+                   << " meta_alloc: " << meta_alloc
+                   << " meta_used: " << meta_used
+                   << " data_alloc: " << data_alloc
+                   << " data_used: " << data_used << dendl;
   }
 
   uint64_t max_shard_onodes = static_cast<uint64_t>(
-      (meta_alloc_bytes / (double) num_shards) / meta_cache.get_bytes_per_onode());
-  uint64_t max_shard_buffer = static_cast<uint64_t>(
-      data_alloc_bytes / num_shards);
-  ldout(store->cct, 30) << __func__ << " max_shard_onodes: " << max_shard_onodes
-                        << " max_shard_buffer: " << max_shard_buffer << dendl;
+      (meta_alloc / (double) num_shards) / meta_cache.get_bytes_per_onode());
+  uint64_t max_shard_buffer = static_cast<uint64_t>(data_alloc / num_shards);
+
+  ldout(cct, 30) << __func__ << " max_shard_onodes: " << max_shard_onodes
+                 << " max_shard_buffer: " << max_shard_buffer << dendl;
 
   for (auto i : store->cache_shards) {
     i->trim(max_shard_onodes, max_shard_buffer);
   }
 }
 
+void BlueStore::MempoolThread::_tune_cache_size(bool interval_stats)
+{
+  auto cct = store->cct;
+  uint64_t target = store->osd_memory_target;
+  uint64_t base = store->osd_memory_base;
+  double fragmentation = store->osd_memory_expected_fragmentation;
+  uint64_t cache_max = ((1.0 - fragmentation) * target) - base;
+  uint64_t cache_min = store->osd_memory_cache_min;
+
+  size_t heap_size = 0;
+  size_t unmapped = 0;
+  uint64_t mapped = 0;
+
+  ceph_heap_release_free_memory();
+  ceph_heap_get_numeric_property("generic.heap_size", &heap_size);
+  ceph_heap_get_numeric_property("tcmalloc.pageheap_unmapped_bytes", &unmapped);
+  mapped = heap_size - unmapped;
+
+  uint64_t new_size = autotune_cache_size;
+  new_size = (new_size < cache_max) ? new_size : cache_max;
+  new_size = (new_size > cache_min) ? new_size : cache_min;
+
+  // Approach the min/max slowly, but bounce away quickly.
+  if ((uint64_t) mapped < target) {
+    double ratio = 1 - ((double) mapped / target);
+    new_size += ratio * (cache_max - new_size); 
+  } else {
+    double ratio = 1 - ((double) target / mapped);
+    new_size -= ratio * (new_size - cache_min);
+  }
+
+  if (interval_stats) {
+    ldout(cct, 5) << __func__
+                  << " target: " << target
+                  << " heap: " << heap_size
+                  << " unmapped: " << unmapped
+                  << " mapped: " << mapped 
+                  << " old cache_size: " << autotune_cache_size
+                  << " new cache size: " << new_size << dendl;
+  } else {
+    ldout(cct, 20) << __func__
+                   << " target: " << target
+                   << " heap: " << heap_size
+                   << " unmapped: " << unmapped
+                   << " mapped: " << mapped        
+                   << " old cache_size: " << autotune_cache_size
+                   << " new cache size: " << new_size << dendl;
+  }
+  autotune_cache_size = new_size;
+}
+
 void BlueStore::MempoolThread::_balance_cache(
     const std::list<PriorityCache::PriCache *>& caches)
 {
-  int64_t mem_avail = store->cache_size;
+  int64_t mem_avail = autotune_cache_size;
 
   // Assign memory for each priority level
   for (int i = 0; i < PriorityCache::Priority::LAST + 1; i++) {
@@ -4012,6 +4089,13 @@ int BlueStore::_set_cache_sizes()
       cct->_conf.get_val<Option::size_t>("bluestore_cache_autotune_chunk_size");
   cache_autotune_interval =
       cct->_conf.get_val<double>("bluestore_cache_autotune_interval");
+  osd_memory_target = cct->_conf.get_val<uint64_t>("osd_memory_target");
+  osd_memory_base = cct->_conf.get_val<uint64_t>("osd_memory_base");
+  osd_memory_expected_fragmentation =
+      cct->_conf.get_val<double>("osd_memory_expected_fragmentation");
+  osd_memory_cache_min = cct->_conf.get_val<uint64_t>("osd_memory_cache_min");
+  osd_memory_cache_resize_interval = 
+      cct->_conf.get_val<double>("osd_memory_cache_resize_interval");
 
   if (cct->_conf->bluestore_cache_size) {
     cache_size = cct->_conf->bluestore_cache_size;
diff --git a/src/os/bluestore/BlueStore.h b/src/os/bluestore/BlueStore.h
index 5e915d3e045b..d45522667397 100644
--- a/src/os/bluestore/BlueStore.h
+++ b/src/os/bluestore/BlueStore.h
@@ -1921,17 +1921,18 @@ private:
   uint64_t kv_throttle_costs = 0;
 
   // cache trim control
-  uint64_t cache_size = 0;      ///< total cache size
+  uint64_t cache_size = 0;       ///< total cache size
   double cache_meta_ratio = 0;   ///< cache ratio dedicated to metadata
   double cache_kv_ratio = 0;     ///< cache ratio dedicated to kv (e.g., rocksdb)
   double cache_data_ratio = 0;   ///< cache ratio dedicated to object data
-  uint64_t cache_meta_min = 0;   ///< cache min dedicated to metadata
-  uint64_t cache_kv_min = 0;     ///< cache min dedicated to kv (e.g., rocksdb)
-  uint64_t cache_data_min = 0;   ///< cache min dedicated to object data
   bool cache_autotune = false;   ///< cache autotune setting
   uint64_t cache_autotune_chunk_size = 0; ///< cache autotune chunk size
   double cache_autotune_interval = 0; ///< time to wait between cache rebalancing
-
+  uint64_t osd_memory_target = 0;   ///< OSD memory target when autotuning cache
+  uint64_t osd_memory_base = 0;     ///< OSD base memory when autotuning cache
+  double osd_memory_expected_fragmentation = 0; ///< expected memory fragmentation
+  uint64_t osd_memory_cache_min = 0; ///< Min memory to assign when autotuning cahce
+  double osd_memory_cache_resize_interval = 0; ///< Time to wait between cache resizing 
   std::mutex vstatfs_lock;
   volatile_statfs vstatfs;
 
@@ -1942,6 +1943,7 @@ private:
     Cond cond;
     Mutex lock;
     bool stop = false;
+    uint64_t autotune_cache_size = 0;
 
     struct MempoolCache : public PriorityCache::PriCache {
       BlueStore *store;
@@ -2060,7 +2062,8 @@ private:
 
   private:
     void _adjust_cache_settings();
-    void _trim_shards(bool log_stats);
+    void _trim_shards(bool interval_stats);
+    void _tune_cache_size(bool interval_stats);
     void _balance_cache(const std::list<PriorityCache::PriCache *>& caches);
     void _balance_cache_pri(int64_t *mem_avail, 
                             const std::list<PriorityCache::PriCache *>& caches, 
-- 
2.47.3