Improve Multi Batch Write (tikv#154)

* perform like normal pipelined write

Signed-off-by: Little-Wallace <bupt2013211450@gmail.com>
Signed-off-by: tabokie <xy.tao@outlook.com>

db/column_family.cc

@@ -1389,6 +1389,12 @@ Status ColumnFamilyData::ValidateOptions(
         "FIFO compaction only supported with max_open_files = -1.");
   }
 
+  if (db_options.enable_multi_thread_write &&
+      cf_options.max_successive_merges > 0) {
+    return Status::NotSupported(
+        "Multi thread write is only supported with no successive merges");
+  }
+
   return s;
 }
 

db/db_impl/db_impl.h

@@ -168,10 +168,9 @@ class DBImpl : public DB {
   virtual Status Write(const WriteOptions& options,
                        WriteBatch* updates) override;
 
-  using DB::MultiThreadWrite;
-  virtual Status MultiThreadWrite(
-      const WriteOptions& options,
-      const std::vector<WriteBatch*>& updates) override;
+  using DB::MultiBatchWrite;
+  virtual Status MultiBatchWrite(const WriteOptions& options,
+                                 std::vector<WriteBatch*>&& updates) override;
 
   using DB::Get;
   virtual Status Get(const ReadOptions& options,
@@ -1296,12 +1295,11 @@ class DBImpl : public DB {
                    size_t batch_cnt = 0,
                    PreReleaseCallback* pre_release_callback = nullptr);
 
-  Status MultiThreadWriteImpl(const WriteOptions& write_options,
-                              const autovector<WriteBatch*>& my_batch,
-                              WriteCallback* callback,
-                              uint64_t* log_used = nullptr,
-                              uint64_t log_ref = 0,
-                              uint64_t* seq_used = nullptr);
+  Status MultiBatchWriteImpl(const WriteOptions& write_options,
+                             std::vector<WriteBatch*>&& my_batch,
+                             WriteCallback* callback,
+                             uint64_t* log_used = nullptr, uint64_t log_ref = 0,
+                             uint64_t* seq_used = nullptr);
 
   Status PipelinedWriteImpl(const WriteOptions& options, WriteBatch* updates,
                             WriteCallback* callback = nullptr,

db/db_impl/db_impl_open.cc

@@ -153,7 +153,7 @@ DBOptions SanitizeOptions(const std::string& dbname, const DBOptions& src,
     result.avoid_flush_during_recovery = false;
   }
 
-  // multi thread write do not support two-write-que or write in pipeline
+  // multi thread write do not support two-write-que or write in 2PC
   if (result.two_write_queues || result.allow_2pc) {
     result.enable_multi_thread_write = false;
   }

db/db_impl/db_impl_write.cc

@@ -62,30 +62,24 @@ Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
 }
 #endif  // ROCKSDB_LITE
 
-Status DBImpl::MultiThreadWrite(const WriteOptions& options,
-                                const std::vector<WriteBatch*>& updates) {
+Status DBImpl::MultiBatchWrite(const WriteOptions& options,
+                               std::vector<WriteBatch*>&& updates) {
   if (immutable_db_options_.enable_multi_thread_write) {
-    if (UNLIKELY(updates.empty())) {
-      return Status::OK();
-    }
-    autovector<WriteBatch*> batches;
-    for (auto w : updates) {
-      batches.push_back(w);
-    }
-    return MultiThreadWriteImpl(options, batches, nullptr, nullptr);
+    return MultiBatchWriteImpl(options, std::move(updates), nullptr, nullptr);
   } else {
     return Status::NotSupported();
   }
 }
 
-Status DBImpl::MultiThreadWriteImpl(const WriteOptions& write_options,
-                                    const autovector<WriteBatch*>& updates,
-                                    WriteCallback* callback, uint64_t* log_used,
-                                    uint64_t log_ref, uint64_t* seq_used) {
+Status DBImpl::MultiBatchWriteImpl(const WriteOptions& write_options,
+                                   std::vector<WriteBatch*>&& my_batch,
+                                   WriteCallback* callback, uint64_t* log_used,
+                                   uint64_t log_ref, uint64_t* seq_used) {
   PERF_TIMER_GUARD(write_pre_and_post_process_time);
   StopWatch write_sw(immutable_db_options_.clock,
                      immutable_db_options_.statistics.get(), DB_WRITE);
-  WriteThread::Writer writer(write_options, updates, callback, log_ref);
+  WriteThread::Writer writer(write_options, std::move(my_batch), callback,
+                             log_ref);
   write_thread_.JoinBatchGroup(&writer);
 
   WriteContext write_context;
@@ -170,36 +164,72 @@ Status DBImpl::MultiThreadWriteImpl(const WriteOptions& write_options,
     }
     write_thread_.ExitAsBatchGroupLeader(wal_write_group, writer.status);
   }
+  bool is_leader_thread = false;
   if (writer.state == WriteThread::STATE_MEMTABLE_WRITER_LEADER) {
     PERF_TIMER_GUARD(write_memtable_time);
     assert(writer.ShouldWriteToMemtable());
     WriteThread::WriteGroup memtable_write_group;
     write_thread_.EnterAsMemTableWriter(&writer, &memtable_write_group);
     assert(immutable_db_options_.allow_concurrent_memtable_write);
-    auto version_set = versions_->GetColumnFamilySet();
-    memtable_write_group.running.store(0);
-    for (auto it = memtable_write_group.begin();
-         it != memtable_write_group.end(); ++it) {
-      if (!it.writer->ShouldWriteToMemtable()) {
-        continue;
+    if (memtable_write_group.size > 1) {
+      is_leader_thread = true;
+      write_thread_.LaunchParallelMemTableWriters(&memtable_write_group);
+    } else {
+      auto version_set = versions_->GetColumnFamilySet();
+      memtable_write_group.running.store(0);
+      for (auto it = memtable_write_group.begin();
+           it != memtable_write_group.end(); ++it) {
+        if (!it.writer->ShouldWriteToMemtable()) {
+          continue;
+        }
+        WriteBatchInternal::AsyncInsertInto(
+            it.writer, it.writer->sequence, version_set, &flush_scheduler_,
+            &trim_history_scheduler_, ignore_missing_faimly, this,
+            &write_thread_.write_queue_);
       }
-      WriteBatchInternal::AsyncInsertInto(
-          it.writer, it.writer->sequence, version_set, &flush_scheduler_,
-          &trim_history_scheduler_, ignore_missing_faimly, this,
-          &write_thread_.write_queue_);
-    }
-    while (memtable_write_group.running.load(std::memory_order_acquire) > 0) {
-      std::function<void()> work;
-      if (write_thread_.write_queue_.PopFront(work)) {
-        work();
-      } else {
-        std::this_thread::yield();
+      while (memtable_write_group.running.load(std::memory_order_acquire) > 0) {
+        if (!write_thread_.write_queue_.RunFunc()) {
+          std::this_thread::yield();
+        }
+      }
+      MemTableInsertStatusCheck(memtable_write_group.status);
+      versions_->SetLastSequence(memtable_write_group.last_sequence);
+      write_thread_.ExitAsMemTableWriter(&writer, memtable_write_group);
+    }
+  }
+  if (writer.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
+    assert(writer.ShouldWriteToMemtable());
+    auto version_set = versions_->GetColumnFamilySet();
+    WriteBatchInternal::AsyncInsertInto(
+        &writer, writer.sequence, version_set, &flush_scheduler_,
+        &trim_history_scheduler_, ignore_missing_faimly, this,
+        &write_thread_.write_queue_);
+    // Because `LaunchParallelMemTableWriters` has add `write_group->size` to
+    // `running`,
+    // the value of `running` is always larger than one if the leader thread
+    // does not
+    // call `CompleteParallelMemTableWriter`.
+    while (writer.write_group->running.load(std::memory_order_acquire) > 1) {
+      // Write thread could exit and block itself if it is not a leader thread.
+      if (!write_thread_.write_queue_.RunFunc() && !is_leader_thread) {
+        break;
       }
     }
-    MemTableInsertStatusCheck(memtable_write_group.status);
-    versions_->SetLastSequence(memtable_write_group.last_sequence);
-    write_thread_.ExitAsMemTableWriter(&writer, memtable_write_group);
+    // We only allow leader_thread to finish this WriteGroup because there may
+    // be another task which is done by the thread that is not in this
+    // WriteGroup,
+    // and it would not notify the threads in this WriteGroup. So we must make
+    // someone in
+    // this WriteGroup to complete it and leader thread is easy to be decided.
+    if (is_leader_thread) {
+      MemTableInsertStatusCheck(writer.status);
+      versions_->SetLastSequence(writer.write_group->last_sequence);
+      write_thread_.ExitAsMemTableWriter(&writer, *writer.write_group);
+    } else {
+      write_thread_.CompleteParallelMemTableWriter(&writer);
+    }
   }
+
   if (seq_used != nullptr) {
     *seq_used = writer.sequence;
   }
@@ -299,10 +329,10 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
   }
 
   if (immutable_db_options_.enable_multi_thread_write) {
-    autovector<WriteBatch*> updates;
-    updates.push_back(my_batch);
-    return MultiThreadWriteImpl(write_options, updates, callback, log_used,
-                                log_ref, seq_used);
+    std::vector<WriteBatch*> updates(1);
+    updates[0] = my_batch;
+    return MultiBatchWriteImpl(write_options, std::move(updates), callback,
+                               log_used, log_ref, seq_used);
   }
 
   if (immutable_db_options_.enable_pipelined_write) {

db/db_write_test.cc

@@ -458,7 +458,10 @@ TEST_P(DBWriteTest, MultiThreadWrite) {
   if (!options.enable_multi_thread_write) {
     return;
   }
-  constexpr int kNumThreads = 8;
+  constexpr int kNumThreads = 4;
+  constexpr int kNumWrite = 4;
+  constexpr int kNumBatch = 8;
+  constexpr int kBatchSize = 16;
   options.env = mock_env.get();
   options.write_buffer_size = 1024 * 128;
   Reopen(options);
@@ -467,21 +470,20 @@ TEST_P(DBWriteTest, MultiThreadWrite) {
     threads.push_back(port::Thread(
         [&](int index) {
           WriteOptions opt;
-          for (int j = 0; j < 64; j++) {
+          std::vector<WriteBatch> data(kNumBatch);
+          for (int j = 0; j < kNumWrite; j++) {
             std::vector<WriteBatch*> batches;
-            for (int i = 0; i < 4; i++) {
-              WriteBatch* batch = new WriteBatch;
-              for (int k = 0; k < 64; k++) {
+            for (int i = 0; i < kNumBatch; i++) {
+              WriteBatch* batch = &data[i];
+              batch->Clear();
+              for (int k = 0; k < kBatchSize; k++) {
                 batch->Put("key_" + ToString(index) + "_" + ToString(j) + "_" +
                                ToString(i) + "_" + ToString(k),
                            "value" + ToString(k));
               }
               batches.push_back(batch);
             }
-            dbfull()->MultiThreadWrite(opt, batches);
-            for (auto b : batches) {
-              delete b;
-            }
+            dbfull()->MultiBatchWrite(opt, std::move(batches));
           }
         },
         t));
@@ -492,9 +494,9 @@ TEST_P(DBWriteTest, MultiThreadWrite) {
   ReadOptions opt;
   for (int t = 0; t < kNumThreads; t++) {
     std::string value;
-    for (int i = 0; i < 64; i++) {
-      for (int j = 0; j < 4; j++) {
-        for (int k = 0; k < 64; k++) {
+    for (int i = 0; i < kNumWrite; i++) {
+      for (int j = 0; j < kNumBatch; j++) {
+        for (int k = 0; k < kBatchSize; k++) {
           ASSERT_OK(dbfull()->Get(opt,
                                   "key_" + ToString(t) + "_" + ToString(i) +
                                       "_" + ToString(j) + "_" + ToString(k),

db/write_batch.cc

@@ -684,7 +684,7 @@ uint32_t WriteBatchInternal::Count(const WriteBatch* b) {
   return DecodeFixed32(b->rep_.data() + 8);
 }
 
-uint32_t WriteBatchInternal::Count(const autovector<WriteBatch*> b) {
+uint32_t WriteBatchInternal::Count(const std::vector<WriteBatch*> b) {
   uint32_t count = 0;
   for (auto w : b) {
     count += DecodeFixed32(w->rep_.data() + 8);
@@ -2367,30 +2367,38 @@ void WriteBatchInternal::AsyncInsertInto(
     WriteThread::Writer* writer, SequenceNumber sequence,
     ColumnFamilySet* version_set, FlushScheduler* flush_scheduler,
     TrimHistoryScheduler* trim_history_scheduler,
-    bool ignore_missing_column_families, DB* db,
-    SafeQueue<std::function<void()>>* pool) {
+    bool ignore_missing_column_families, DB* db, SafeFuncQueue* pool) {
   auto write_group = writer->write_group;
-  write_group->running.fetch_add(writer->batches.size(),
-                                 std::memory_order_seq_cst);
-  for (auto w : writer->batches) {
-    pool->PushBack([=]() {
+  auto batch_size = writer->batches.size();
+  write_group->running.fetch_add(batch_size);
+  for (size_t i = 0; i < batch_size; i++) {
+    auto f = [=]() {
       ColumnFamilyMemTablesImpl memtables(version_set);
       MemTableInserter inserter(
           sequence, &memtables, flush_scheduler, trim_history_scheduler,
           ignore_missing_column_families, 0 /*recovering_log_number*/, db,
           true /*concurrent_memtable_writes*/, nullptr /*prot_info*/,
           nullptr /*has_valid_writes*/);
       inserter.set_log_number_ref(writer->log_ref);
-      SetSequence(w, sequence);
-      Status s = w->Iterate(&inserter);
+      SetSequence(writer->batches[i], sequence);
+      Status s = writer->batches[i]->Iterate(&inserter);
       if (!s.ok()) {
         std::lock_guard<std::mutex> guard(write_group->leader->StateMutex());
         write_group->status = s;
       }
       inserter.PostProcess();
-      write_group->running.fetch_sub(1, std::memory_order_release);
-    });
-    sequence += WriteBatchInternal::Count(w);
+      write_group->running.fetch_sub(1);
+    };
+    if (i + 1 == batch_size) {
+      // If there is only one WriteBatch written by this thread, It shall do it
+      // by self, because this batch may be large. And every thread does the
+      // latest
+      // one by self will reduce the cost of calling `SafeFuncQueue::Push`.
+      f();
+    } else {
+      pool->Push(std::move(f));
+      sequence += WriteBatchInternal::Count(writer->batches[i]);
+    }
   }
 }
 

db/write_batch_internal.h

@@ -133,7 +133,7 @@ class WriteBatchInternal {
   // Return the number of entries in the batch.
   static uint32_t Count(const WriteBatch* batch);
 
-  static uint32_t Count(const autovector<WriteBatch*> batch);
+  static uint32_t Count(const std::vector<WriteBatch*> batch);
 
   // Set the count for the number of entries in the batch.
   static void SetCount(WriteBatch* batch, uint32_t n);
@@ -157,7 +157,7 @@ class WriteBatchInternal {
     return batch->rep_.size();
   }
 
-  static size_t ByteSize(const autovector<WriteBatch*> batch) {
+  static size_t ByteSize(const std::vector<WriteBatch*> batch) {
     size_t count = 0;
     for (auto w : batch) {
       count += w->rep_.size();
@@ -223,7 +223,7 @@ class WriteBatchInternal {
                               FlushScheduler* flush_scheduler,
                               TrimHistoryScheduler* trim_history_scheduler,
                               bool ignore_missing_column_families, DB* db,
-                              SafeQueue<std::function<void()>>* pool);
+                              SafeFuncQueue* pool);
 
   static Status Append(WriteBatch* dst, const WriteBatch* src,
                        const bool WAL_only = false);