Merge pull request #2 from XiaoMi/master

feat: optimize tcmalloc release memory (XiaoMi#343)

src/dist/replication/common/replication_common.cpp

@@ -100,7 +100,8 @@ replication_options::replication_options()
     config_sync_interval_ms = 30000;
 
     mem_release_enabled = true;
-    mem_release_interval_ms = 86400000;
+    mem_release_check_interval_ms = 3600000;
+    mem_release_max_reserved_mem_percentage = 10;
 
     lb_interval_ms = 10000;
 
@@ -479,11 +480,18 @@ void replication_options::initialize()
                                                     mem_release_enabled,
                                                     "whether to enable periodic memory release");
 
-    mem_release_interval_ms = (int)dsn_config_get_value_uint64(
+    mem_release_check_interval_ms = (int)dsn_config_get_value_uint64(
         "replication",
-        "mem_release_interval_ms",
-        mem_release_interval_ms,
-        "the replica releases its idle memory to the system every this period of time(ms)");
+        "mem_release_check_interval_ms",
+        mem_release_check_interval_ms,
+        "the replica check if should release memory to the system every this period of time(ms)");
+
+    mem_release_max_reserved_mem_percentage = (int)dsn_config_get_value_uint64(
+        "replication",
+        "mem_release_max_reserved_mem_percentage",
+        mem_release_max_reserved_mem_percentage,
+        "if tcmalloc reserved but not-used memory exceed this percentage of application allocated "
+        "memory, replica server will release the exceeding memory back to operating system");
 
     lb_interval_ms = (int)dsn_config_get_value_uint64(
         "replication",

src/dist/replication/common/replication_common.h

@@ -104,7 +104,8 @@ class replication_options
     int32_t config_sync_interval_ms;
 
     bool mem_release_enabled;
-    int32_t mem_release_interval_ms;
+    int32_t mem_release_check_interval_ms;
+    int32_t mem_release_max_reserved_mem_percentage;
 
     int32_t lb_interval_ms;
 

src/dist/replication/lib/replica_stub.cpp

@@ -68,11 +68,13 @@ replica_stub::replica_stub(replica_state_subscriber subscriber /*= nullptr*/,
       _query_compact_command(nullptr),
       _query_app_envs_command(nullptr),
       _useless_dir_reserve_seconds_command(nullptr),
+      _max_reserved_memory_percentage_command(nullptr),
       _deny_client(false),
       _verbose_client_log(false),
       _verbose_commit_log(false),
       _gc_disk_error_replica_interval_seconds(3600),
       _gc_disk_garbage_replica_interval_seconds(3600),
+      _mem_release_max_reserved_mem_percentage(10),
       _learn_app_concurrent_count(0),
       _fs_manager(false)
 {
@@ -317,6 +319,7 @@ void replica_stub::initialize(const replication_options &opts, bool clear /* = f
     _verbose_commit_log = _options.verbose_commit_log_on_start;
     _gc_disk_error_replica_interval_seconds = _options.gc_disk_error_replica_interval_seconds;
     _gc_disk_garbage_replica_interval_seconds = _options.gc_disk_garbage_replica_interval_seconds;
+    _mem_release_max_reserved_mem_percentage = _options.mem_release_max_reserved_mem_percentage;
 
     // clear dirs if need
     if (clear) {
@@ -638,17 +641,13 @@ void replica_stub::initialize_start()
 
 #ifdef DSN_ENABLE_GPERF
     if (_options.mem_release_enabled) {
-        _mem_release_timer_task =
-            tasking::enqueue_timer(LPC_MEM_RELEASE,
-                                   &_tracker,
-                                   []() {
-                                       ddebug("Memory release has started...");
-                                       ::MallocExtension::instance()->ReleaseFreeMemory();
-                                       ddebug("Memory release has ended...");
-                                   },
-                                   std::chrono::milliseconds(_options.mem_release_interval_ms),
-                                   0,
-                                   std::chrono::milliseconds(_options.mem_release_interval_ms));
+        _mem_release_timer_task = tasking::enqueue_timer(
+            LPC_MEM_RELEASE,
+            &_tracker,
+            std::bind(&replica_stub::gc_tcmalloc_memory, this),
+            std::chrono::milliseconds(_options.mem_release_check_interval_ms),
+            0,
+            std::chrono::milliseconds(_options.mem_release_check_interval_ms));
     }
 #endif
 
@@ -2053,6 +2052,35 @@ void replica_stub::open_service()
             }
             return result;
         });
+
+#ifdef DSN_ENABLE_GPERF
+    _max_reserved_memory_percentage_command = dsn::command_manager::instance().register_app_command(
+        {"mem-release-max-reserved-percentage"},
+        "mem-release-max-reserved-percentage [num | DEFAULT]",
+        "control tcmalloc max reserved but not-used memory percentage",
+        [this](const std::vector<std::string> &args) {
+            std::string result("OK");
+            if (args.empty()) {
+                // show current value
+                result = "mem-release-max-reserved-percentage = " +
+                         std::to_string(_mem_release_max_reserved_mem_percentage);
+                return result;
+            }
+            if (args[0] == "DEFAULT") {
+                // set to default value
+                _mem_release_max_reserved_mem_percentage =
+                    _options.mem_release_max_reserved_mem_percentage;
+                return result;
+            }
+            int32_t percentage = 0;
+            if (!dsn::buf2int32(args[0], percentage) || percentage <= 0 || percentage >= 100) {
+                result = std::string("ERR: invalid arguments");
+            } else {
+                _mem_release_max_reserved_mem_percentage = percentage;
+            }
+            return result;
+        });
+#endif
 }
 
 std::string
@@ -2178,6 +2206,9 @@ void replica_stub::close()
     dsn::command_manager::instance().deregister_command(_query_compact_command);
     dsn::command_manager::instance().deregister_command(_query_app_envs_command);
     dsn::command_manager::instance().deregister_command(_useless_dir_reserve_seconds_command);
+#ifdef DSN_ENABLE_GPERF
+    dsn::command_manager::instance().deregister_command(_max_reserved_memory_percentage_command);
+#endif
 
     _kill_partition_command = nullptr;
     _deny_client_command = nullptr;
@@ -2187,6 +2218,7 @@ void replica_stub::close()
     _query_compact_command = nullptr;
     _query_app_envs_command = nullptr;
     _useless_dir_reserve_seconds_command = nullptr;
+    _max_reserved_memory_percentage_command = nullptr;
 
     if (_config_sync_timer_task != nullptr) {
         _config_sync_timer_task->cancel(true);
@@ -2304,6 +2336,44 @@ replica_stub::get_child_dir(const char *app_type, gpid child_pid, const std::str
     return child_dir;
 }
 
+#ifdef DSN_ENABLE_GPERF
+// Get tcmalloc numeric property (name is "prop") value.
+// Return -1 if get property failed (property we used will be greater than zero)
+// Properties can be found in 'gperftools/malloc_extension.h'
+static int64_t get_tcmalloc_numeric_property(const char *prop)
+{
+    size_t value;
+    if (!::MallocExtension::instance()->GetNumericProperty(prop, &value)) {
+        derror_f("Failed to get tcmalloc property {}", prop);
+        return -1;
+    }
+    return value;
+}
+
+void replica_stub::gc_tcmalloc_memory()
+{
+    int64_t total_allocated_bytes =
+        get_tcmalloc_numeric_property("generic.current_allocated_bytes");
+    int64_t reserved_bytes = get_tcmalloc_numeric_property("tcmalloc.pageheap_free_bytes");
+    if (total_allocated_bytes == -1 || reserved_bytes == -1) {
+        return;
+    }
+
+    int64_t max_reserved_bytes =
+        total_allocated_bytes * _mem_release_max_reserved_mem_percentage / 100.0;
+    if (reserved_bytes > max_reserved_bytes) {
+        int64_t release_bytes = reserved_bytes - max_reserved_bytes;
+        ddebug_f("Memory release started, almost {} bytes will be released", release_bytes);
+        while (release_bytes > 0) {
+            // tcmalloc releasing memory will lock page heap, release 1MB at a time to avoid locking
+            // page heap for long time
+            ::MallocExtension::instance()->ReleaseToSystem(1024 * 1024);
+            release_bytes -= 1024 * 1024;
+        }
+    }
+}
+#endif
+
 //
 // partition split
 //

src/dist/replication/lib/replica_stub.h

@@ -246,6 +246,11 @@ class replica_stub : public serverlet<replica_stub>, public ref_counter
                          partition_status::type status,
                          error_code error);
 
+#ifdef DSN_ENABLE_GPERF
+    // Try to release tcmalloc memory back to operating system
+    void gc_tcmalloc_memory();
+#endif
+
 private:
     friend class ::dsn::replication::replication_checker;
     friend class ::dsn::replication::test::test_checker;
@@ -305,12 +310,14 @@ class replica_stub : public serverlet<replica_stub>, public ref_counter
     dsn_handle_t _query_compact_command;
     dsn_handle_t _query_app_envs_command;
     dsn_handle_t _useless_dir_reserve_seconds_command;
+    dsn_handle_t _max_reserved_memory_percentage_command;
 
     bool _deny_client;
     bool _verbose_client_log;
     bool _verbose_commit_log;
     int32_t _gc_disk_error_replica_interval_seconds;
     int32_t _gc_disk_garbage_replica_interval_seconds;
+    int32_t _mem_release_max_reserved_mem_percentage;
 
     // we limit LT_APP max concurrent count, because nfs service implementation is
     // too simple, it do not support priority.