RGW: Make RADOS handles in RGW to be a configurable option

Having a configurable number of RADOS handles is resulting
in significant performance boost for all types of workloads.
Each RGW worker thread would now get to pick a RADOS handle
for its lifetime, from the available bunch.

Signed-off-by: Pavan Rallabhandi <pavan.rallabhandi@sandisk.com>

diff --git a/src/common/config_opts.h b/src/common/config_opts.h
index d9b1dfcf1e9f638994ec51e850787805943a29d1..d69a9df799b98a8bc673741018e7ad112570b87f 100644 (file)
--- a/src/common/config_opts.h
+++ b/src/common/config_opts.h
@@ -992,6 +992,7 @@ OPTION(rgw_op_thread_timeout, OPT_INT, 10*60)
 OPTION(rgw_op_thread_suicide_timeout, OPT_INT, 0)
 OPTION(rgw_thread_pool_size, OPT_INT, 100)
 OPTION(rgw_num_control_oids, OPT_INT, 8)
+OPTION(rgw_num_rados_handles, OPT_U32, 1)
 
 OPTION(rgw_zone, OPT_STR, "") // zone name
 OPTION(rgw_zone_root_pool, OPT_STR, ".rgw.root")    // pool where zone specific info is stored

diff --git a/src/rgw/rgw_gc.cc b/src/rgw/rgw_gc.cc
index c536fa96d161358b0da6453c5c0f5ea46c8bbe09..70f5691d6d4410de55871f87950b53e87988a7d8 100644 (file)
--- a/src/rgw/rgw_gc.cc
+++ b/src/rgw/rgw_gc.cc
@@ -186,7 +186,7 @@ int RGWGC::process(int index, int max_secs)
         if (obj.pool != last_pool) {
           delete ctx;
           ctx = new IoCtx;
-         ret = store->rados->ioctx_create(obj.pool.c_str(), *ctx);
+         ret = store->get_rados_handle()->ioctx_create(obj.pool.c_str(), *ctx);
          if (ret < 0) {
            dout(0) << "ERROR: failed to create ioctx pool=" << obj.pool << dendl;
            continue;

diff --git a/src/rgw/rgw_rados.cc b/src/rgw/rgw_rados.cc
index d7c312fce7d0fcc0d0507fef4c576ba7fa175175..e57af444b51d948405c930191e604c33bbfcd100 100644 (file)
--- a/src/rgw/rgw_rados.cc
+++ b/src/rgw/rgw_rados.cc
@@ -1223,7 +1223,7 @@ int RGWRados::unwatch(uint64_t watch_handle)
     ldout(cct, 0) << "ERROR: rados->unwatch2() returned r=" << r << dendl;
     return r;
   }
-  r = rados->watch_flush();
+  r = rados[0]->watch_flush();
   if (r < 0) {
     ldout(cct, 0) << "ERROR: rados->watch_flush() returned r=" << r << dendl;
     return r;
@@ -1431,23 +1431,50 @@ void RGWRados::finalize()
  */
 int RGWRados::init_rados()
 {
-  int ret;
+  int ret = 0;
 
-  rados = new Rados();
-  if (!rados)
-    return -ENOMEM;
+  num_rados_handles = cct->_conf->rgw_num_rados_handles;
 
-  ret = rados->init_with_context(cct);
-  if (ret < 0)
-   return ret;
+  rados = new librados::Rados *[num_rados_handles];
+  if (!rados) {
+    ret = -ENOMEM;
+    return ret;
+  }
 
-  ret = rados->connect();
-  if (ret < 0)
-   return ret;
+  for (uint32_t i=0; i < num_rados_handles; i++) {
+
+    rados[i] = new Rados();
+    if (!rados[i]) {
+      ret = -ENOMEM;
+      goto fail;
+    }
+
+    ret = rados[i]->init_with_context(cct);
+    if (ret < 0) {
+      goto fail;
+    }
+
+    ret = rados[i]->connect();
+    if (ret < 0) {
+      goto fail;
+    }
+  }
 
   meta_mgr = new RGWMetadataManager(cct, this);
   data_log = new RGWDataChangesLog(cct, this);
 
+  return ret;
+
+fail:
+  for (uint32_t i=0; i < num_rados_handles; i++) {
+    if (rados[i]) {
+      delete rados[i];
+    }
+  }
+  if (rados) {
+    delete[] rados;
+  }
+
   return ret;
 }
 
@@ -1654,15 +1681,16 @@ int RGWRados::open_root_pool_ctx()
 {
   const string& pool = zone.domain_root.name;
   const char *pool_str = pool.c_str();
-  int r = rados->ioctx_create(pool_str, root_pool_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(pool_str, root_pool_ctx);
   if (r == -ENOENT) {
-    r = rados->pool_create(pool_str);
+    r = rad->pool_create(pool_str);
     if (r == -EEXIST)
       r = 0;
     if (r < 0)
       return r;
 
-    r = rados->ioctx_create(pool_str, root_pool_ctx);
+    r = rad->ioctx_create(pool_str, root_pool_ctx);
   }
 
   return r;
@@ -1671,15 +1699,16 @@ int RGWRados::open_root_pool_ctx()
 int RGWRados::open_gc_pool_ctx()
 {
   const char *gc_pool = zone.gc_pool.name.c_str();
-  int r = rados->ioctx_create(gc_pool, gc_pool_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(gc_pool, gc_pool_ctx);
   if (r == -ENOENT) {
-    r = rados->pool_create(gc_pool);
+    r = rad->pool_create(gc_pool);
     if (r == -EEXIST)
       r = 0;
     if (r < 0)
       return r;
 
-    r = rados->ioctx_create(gc_pool, gc_pool_ctx);
+    r = rad->ioctx_create(gc_pool, gc_pool_ctx);
   }
 
   return r;
@@ -1688,15 +1717,16 @@ int RGWRados::open_gc_pool_ctx()
 int RGWRados::init_watch()
 {
   const char *control_pool = zone.control_pool.name.c_str();
-  int r = rados->ioctx_create(control_pool, control_pool_ctx);
+  librados::Rados *rad = rados[0];
+  int r = rad->ioctx_create(control_pool, control_pool_ctx);
   if (r == -ENOENT) {
-    r = rados->pool_create(control_pool);
+    r = rad->pool_create(control_pool);
     if (r == -EEXIST)
       r = 0;
     if (r < 0)
       return r;
 
-    r = rados->ioctx_create(control_pool, control_pool_ctx);
+    r = rad->ioctx_create(control_pool, control_pool_ctx);
     if (r < 0)
       return r;
   }
@@ -1752,18 +1782,19 @@ void RGWRados::pick_control_oid(const string& key, string& notify_oid)
 
 int RGWRados::open_bucket_pool_ctx(const string& bucket_name, const string& pool, librados::IoCtx&  io_ctx)
 {
-  int r = rados->ioctx_create(pool.c_str(), io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(pool.c_str(), io_ctx);
   if (r != -ENOENT)
     return r;
 
   if (!pools_initialized)
     return r;
 
-  r = rados->pool_create(pool.c_str());
+  r = rad->pool_create(pool.c_str());
   if (r < 0 && r != -EEXIST)
     return r;
 
-  r = rados->ioctx_create(pool.c_str(), io_ctx);
+  r = rad->ioctx_create(pool.c_str(), io_ctx);
 
   return r;
 }
@@ -1853,7 +1884,8 @@ int RGWRados::log_list_init(const string& prefix, RGWAccessHandle *handle)
 {
   log_list_state *state = new log_list_state;
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, state->io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, state->io_ctx);
   if (r < 0) {
     delete state;
     return r;
@@ -1888,7 +1920,8 @@ int RGWRados::log_remove(const string& name)
 {
   librados::IoCtx io_ctx;
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r < 0)
     return r;
   return io_ctx.remove(name);
@@ -1908,7 +1941,8 @@ int RGWRados::log_show_init(const string& name, RGWAccessHandle *handle)
 {
   log_show_state *state = new log_show_state;
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, state->io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, state->io_ctx);
   if (r < 0) {
     delete state;
     return r;
@@ -2128,7 +2162,8 @@ int RGWRados::time_log_add(const string& oid, const utime_t& ut, const string& s
   librados::IoCtx io_ctx;
 
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r == -ENOENT) {
     rgw_bucket pool(log_pool);
     r = create_pool(pool);
@@ -2136,7 +2171,7 @@ int RGWRados::time_log_add(const string& oid, const utime_t& ut, const string& s
       return r;
  
     // retry
-    r = rados->ioctx_create(log_pool, io_ctx);
+    r = rad->ioctx_create(log_pool, io_ctx);
   }
   if (r < 0)
     return r;
@@ -2153,7 +2188,8 @@ int RGWRados::time_log_add(const string& oid, list<cls_log_entry>& entries)
   librados::IoCtx io_ctx;
 
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r == -ENOENT) {
     rgw_bucket pool(log_pool);
     r = create_pool(pool);
@@ -2161,7 +2197,7 @@ int RGWRados::time_log_add(const string& oid, list<cls_log_entry>& entries)
       return r;
  
     // retry
-    r = rados->ioctx_create(log_pool, io_ctx);
+    r = rad->ioctx_create(log_pool, io_ctx);
   }
   if (r < 0)
     return r;
@@ -2182,7 +2218,8 @@ int RGWRados::time_log_list(const string& oid, utime_t& start_time, utime_t& end
   librados::IoCtx io_ctx;
 
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r < 0)
     return r;
   librados::ObjectReadOperation op;
@@ -2204,7 +2241,8 @@ int RGWRados::time_log_info(const string& oid, cls_log_header *header)
   librados::IoCtx io_ctx;
 
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r < 0)
     return r;
   librados::ObjectReadOperation op;
@@ -2226,7 +2264,8 @@ int RGWRados::time_log_trim(const string& oid, const utime_t& start_time, const
   librados::IoCtx io_ctx;
 
   const char *log_pool = zone.log_pool.name.c_str();
-  int r = rados->ioctx_create(log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(log_pool, io_ctx);
   if (r < 0)
     return r;
 
@@ -2240,7 +2279,8 @@ int RGWRados::lock_exclusive(rgw_bucket& pool, const string& oid, utime_t& durat
 
   const char *pool_name = pool.name.c_str();
   
-  int r = rados->ioctx_create(pool_name, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(pool_name, io_ctx);
   if (r < 0)
     return r;
   
@@ -2258,7 +2298,8 @@ int RGWRados::unlock(rgw_bucket& pool, const string& oid, string& zone_id, strin
 
   const char *pool_name = pool.name.c_str();
 
-  int r = rados->ioctx_create(pool_name, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(pool_name, io_ctx);
   if (r < 0)
     return r;
   
@@ -2479,14 +2520,15 @@ int RGWRados::create_pool(rgw_bucket& bucket)
 
   string pool = bucket.index_pool;
 
-  ret = rados->pool_create(pool.c_str(), 0);
+  librados::Rados *rad = get_rados_handle();
+  ret = rad->pool_create(pool.c_str(), 0);
   if (ret == -EEXIST)
     ret = 0;
   if (ret < 0)
     return ret;
 
   if (bucket.data_pool != pool) {
-    ret = rados->pool_create(bucket.data_pool.c_str(), 0);
+    ret = rad->pool_create(bucket.data_pool.c_str(), 0);
     if (ret == -EEXIST)
       ret = 0;
     if (ret < 0)
@@ -2542,7 +2584,8 @@ int RGWRados::create_bucket(RGWUserInfo& owner, rgw_bucket& bucket,
     const string& pool = zone.domain_root.name;
     const char *pool_str = pool.c_str();
     librados::IoCtx id_io_ctx;
-    int r = rados->ioctx_create(pool_str, id_io_ctx);
+    librados::Rados *rad = get_rados_handle();
+    int r = rad->ioctx_create(pool_str, id_io_ctx);
     if (r < 0)
       return r;
 
@@ -2828,7 +2871,8 @@ int RGWRados::update_placement_map()
 
 int RGWRados::add_bucket_placement(std::string& new_pool)
 {
-  int ret = rados->pool_lookup(new_pool.c_str());
+  librados::Rados *rad = get_rados_handle();
+  int ret = rad->pool_lookup(new_pool.c_str());
   if (ret < 0) // DNE, or something
     return ret;
 
@@ -2878,11 +2922,12 @@ int RGWRados::create_pools(vector<string>& names, vector<int>& retcodes)
   vector<librados::PoolAsyncCompletion *> completions;
   vector<int> rets;
 
+  librados::Rados *rad = get_rados_handle();
   for (iter = names.begin(); iter != names.end(); ++iter) {
     librados::PoolAsyncCompletion *c = librados::Rados::pool_async_create_completion();
     completions.push_back(c);
     string& name = *iter;
-    int ret = rados->pool_create_async(name.c_str(), c);
+    int ret = rad->pool_create_async(name.c_str(), c);
     rets.push_back(ret);
   }
 
@@ -7156,7 +7201,8 @@ int RGWRados::append_async(rgw_obj& obj, size_t size, bufferlist& bl)
   if (r < 0) {
     return r;
   }
-  librados::AioCompletion *completion = rados->aio_create_completion(NULL, NULL, NULL);
+  librados::Rados *rad = get_rados_handle();
+  librados::AioCompletion *completion = rad->aio_create_completion(NULL, NULL, NULL);
 
   r = ref.ioctx.aio_append(ref.oid, completion, bl, size);
   completion->release();
@@ -7708,7 +7754,8 @@ int RGWRados::cls_obj_usage_log_add(const string& oid, rgw_usage_log_info& info)
   librados::IoCtx io_ctx;
 
   const char *usage_log_pool = zone.usage_log_pool.name.c_str();
-  int r = rados->ioctx_create(usage_log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(usage_log_pool, io_ctx);
   if (r == -ENOENT) {
     rgw_bucket pool(usage_log_pool);
     r = create_pool(pool);
@@ -7716,7 +7763,7 @@ int RGWRados::cls_obj_usage_log_add(const string& oid, rgw_usage_log_info& info)
       return r;
  
     // retry
-    r = rados->ioctx_create(usage_log_pool, io_ctx);
+    r = rad->ioctx_create(usage_log_pool, io_ctx);
   }
   if (r < 0)
     return r;
@@ -7736,7 +7783,8 @@ int RGWRados::cls_obj_usage_log_read(string& oid, string& user, uint64_t start_e
   *is_truncated = false;
 
   const char *usage_log_pool = zone.usage_log_pool.name.c_str();
-  int r = rados->ioctx_create(usage_log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(usage_log_pool, io_ctx);
   if (r < 0)
     return r;
 
@@ -7751,7 +7799,8 @@ int RGWRados::cls_obj_usage_log_trim(string& oid, string& user, uint64_t start_e
   librados::IoCtx io_ctx;
 
   const char *usage_log_pool = zone.usage_log_pool.name.c_str();
-  int r = rados->ioctx_create(usage_log_pool, io_ctx);
+  librados::Rados *rad = get_rados_handle();
+  int r = rad->ioctx_create(usage_log_pool, io_ctx);
   if (r < 0)
     return r;
 
@@ -8244,7 +8293,7 @@ string RGWStateLog::get_oid(const string& object) {
 int RGWStateLog::open_ioctx(librados::IoCtx& ioctx) {
   string pool_name;
   store->get_log_pool_name(pool_name);
-  int r = store->rados->ioctx_create(pool_name.c_str(), ioctx);
+  int r = store->get_rados_handle()->ioctx_create(pool_name.c_str(), ioctx);
   if (r < 0) {
     lderr(store->ctx()) << "ERROR: could not open rados pool" << dendl;
     return r;
@@ -8488,7 +8537,7 @@ int RGWOpStateSingleOp::renew_state() {
 
 uint64_t RGWRados::instance_id()
 {
-  return rados->get_instance_id();
+  return get_rados_handle()->get_instance_id();
 }
 
 uint64_t RGWRados::next_bucket_id()
@@ -8540,3 +8589,30 @@ void RGWStoreManager::close_storage(RGWRados *store)
   delete store;
 }
 
+librados::Rados* RGWRados::get_rados_handle()
+{
+  if (num_rados_handles == 1) {
+    return rados[0];
+  } else {
+    handle_lock.get_read();
+    pthread_t id = pthread_self();
+    std::map<pthread_t, int>:: iterator it = rados_map.find(id);
+
+    if (it != rados_map.end()) {
+      handle_lock.put_read();
+      return rados[it->second];
+    } else {
+      handle_lock.put_read();
+      handle_lock.get_write();
+      if (next_rados_handle.read() == num_rados_handles) {
+        next_rados_handle.set(0);
+      }
+      int handle = next_rados_handle.read();
+      rados_map[id] = handle;
+      next_rados_handle.inc();
+      handle_lock.put_write();
+      return rados[handle];
+    }
+  }
+}
+

diff --git a/src/rgw/rgw_rados.h b/src/rgw/rgw_rados.h
index 45b7f61c30fccf3730066f58cfd99e8113151d1f..d72ac1825fb220bd84fb9bc49e05e735fe577e2d 100644 (file)
--- a/src/rgw/rgw_rados.h
+++ b/src/rgw/rgw_rados.h
@@ -1236,7 +1236,13 @@ class RGWRados
   void cls_obj_check_prefix_exist(librados::ObjectOperation& op, const string& prefix, bool fail_if_exist);
 protected:
   CephContext *cct;
-  librados::Rados *rados;
+
+  librados::Rados **rados;
+  atomic_t next_rados_handle;
+  uint32_t num_rados_handles;
+  RWLock handle_lock;
+  std::map<pthread_t, int> rados_map;
+
   librados::IoCtx gc_pool_ctx;        // .rgw.gc
 
   bool pools_initialized;
@@ -1255,8 +1261,9 @@ public:
                watch_initialized(false),
                bucket_id_lock("rados_bucket_id"),
                bucket_index_max_shards(0),
-               max_bucket_id(0),
-               cct(NULL), rados(NULL),
+               max_bucket_id(0), cct(NULL),
+               rados(NULL), next_rados_handle(0),
+               num_rados_handles(0), handle_lock("rados_handle_lock"),
                pools_initialized(false),
                quota_handler(NULL),
                finisher(NULL),
@@ -1288,9 +1295,14 @@ public:
   RGWDataChangesLog *data_log;
 
   virtual ~RGWRados() {
+    for (uint32_t i=0; i < num_rados_handles; i++) {
+      if (rados[i]) {
+        rados[i]->shutdown();
+        delete rados[i];
+      }
+    }
     if (rados) {
-      rados->shutdown();
-      delete rados;
+      delete[] rados;
     }
   }
 
@@ -2135,6 +2147,8 @@ public:
 
   uint64_t instance_id();
   uint64_t next_bucket_id();
+
+  librados::Rados* get_rados_handle();
 };
 
 class RGWStoreManager {

diff --git a/src/rgw/rgw_replica_log.cc b/src/rgw/rgw_replica_log.cc
index 6d8ed096da6704edf8f9d69eeb216d74fb20445d..b56a90b440f7becae20f502c53a46a8cdeef88b8 100644 (file)
--- a/src/rgw/rgw_replica_log.cc
+++ b/src/rgw/rgw_replica_log.cc
@@ -37,7 +37,7 @@ RGWReplicaLogger::RGWReplicaLogger(RGWRados *_store) :
 
 int RGWReplicaLogger::open_ioctx(librados::IoCtx& ctx, const string& pool)
 {
-  int r = store->rados->ioctx_create(pool.c_str(), ctx);
+  int r = store->get_rados_handle()->ioctx_create(pool.c_str(), ctx);
   if (r == -ENOENT) {
     rgw_bucket p(pool.c_str());
     r = store->create_pool(p);
@@ -45,7 +45,7 @@ int RGWReplicaLogger::open_ioctx(librados::IoCtx& ctx, const string& pool)
       return r;
 
     // retry
-    r = store->rados->ioctx_create(pool.c_str(), ctx);
+    r = store->get_rados_handle()->ioctx_create(pool.c_str(), ctx);
   }
   if (r < 0) {
     lderr(cct) << "ERROR: could not open rados pool " << pool << dendl;