add bufferPool for nimble parallel writer (#103)

Summary:

change the buffer class to a bufferPool to handle multihreaded buffers without mutexes

Differential Revision: D64774959

dwio/nimble/common/Buffer.h

@@ -15,6 +15,8 @@
  */
 #pragma once
 
+#include "dwio/nimble/common/Exceptions.h"
+#include "folly/fibers/Semaphore.h"
 #include "velox/buffer/Buffer.h"
 #include "velox/common/memory/Memory.h"
 
@@ -33,12 +35,11 @@
 //   and so on
 
 namespace facebook::nimble {
+using MemoryPool = facebook::velox::memory::MemoryPool;
 
-// Internally manages memory in chunks. Releases memory only upon destruction.
+// Internally manages memory in chunks. releases memory when destroyed
 // Buffer is NOT threadsafe: external locking is required.
 class Buffer {
-  using MemoryPool = facebook::velox::memory::MemoryPool;
-
  public:
   explicit Buffer(
       MemoryPool& memoryPool,
@@ -52,7 +53,6 @@ class Buffer {
   // to, and guarantees for the lifetime of *this that that region will remain
   // valid. Does NOT guarantee that the region is initially 0'd.
   char* reserve(uint64_t bytes) {
-    std::scoped_lock<std::mutex> l(mutex_);
     if (reserveEnd_ + bytes <= chunkEnd_) {
       pos_ = reserveEnd_;
       reserveEnd_ += bytes;
@@ -98,11 +98,73 @@ class Buffer {
   char* reserveEnd_;
   std::vector<velox::BufferPtr> chunks_;
   MemoryPool& memoryPool_;
-  // NOTE: this is temporary fix, to quickly enable parallel access to the
-  // buffer class. In the near future, we are going to templetize this class to
-  // produce a concurrent and a non-concurrent variants, and change the call
-  // sites to use each variant when needed.
+};
+
+// Manages a pool of buffers. Buffers are returned to the pool when released.
+// maxPoolSize should be set to at least 90% of capacity for performance
+class BufferPool {
+ public:
+  explicit BufferPool(
+      MemoryPool& memoryPool,
+      size_t maxPoolSize = std::thread::hardware_concurrency(),
+      size_t initialBufferCount = 20, // should be enough for 1 stripe
+      uint64_t initialChunkSize = kMinChunkSize)
+      : defaultInitialChunkSize_{initialChunkSize},
+        semaphore_{maxPoolSize},
+        memoryPool_{memoryPool} {
+    NIMBLE_CHECK(maxPoolSize > 0, "max pool size must be > 0")
+    NIMBLE_CHECK(
+        initialBufferCount <= maxPoolSize,
+        "initial pool size must be <= max pool size");
+
+    pool_.reserve(maxPoolSize);
+
+    for (size_t i = 0; i < initialBufferCount; ++i) {
+      addBuffer(newBuffer());
+    }
+  }
+
+  MemoryPool& getMemoryPool() {
+    return memoryPool_;
+  }
+
+  // buffer back to the pool.
+  void addBuffer(std::unique_ptr<Buffer> buffer) {
+    std::scoped_lock<std::mutex> lock(mutex_);
+    pool_.push_back(std::move(buffer));
+    semaphore_.signal();
+  }
+
+  // Reserves a buffer from the pool. Adds a new buffer to the pool
+  // while there are buffers available
+  std::unique_ptr<Buffer> reserveBuffer() {
+    std::scoped_lock<std::mutex> lock(mutex_);
+    if (pool_.empty()) {
+      return newBuffer();
+    }
+
+    semaphore_.wait();
+    auto buffer = std::move(pool_.back());
+    pool_.pop_back();
+    return buffer;
+  }
+
+  // Returns estimated number of buffers in the pool
+  size_t size() {
+    return pool_.size();
+  }
+
+ private:
+  static const uint64_t kMinChunkSize = 1LL << 20;
+  const uint64_t defaultInitialChunkSize_ = kMinChunkSize;
   std::mutex mutex_;
+  folly::fibers::Semaphore semaphore_;
+  std::vector<std::unique_ptr<Buffer>> pool_;
+  MemoryPool& memoryPool_;
+
+  std::unique_ptr<Buffer> newBuffer() {
+    return std::make_unique<Buffer>(memoryPool_, defaultInitialChunkSize_);
+  }
 };
 
 } // namespace facebook::nimble

dwio/nimble/common/tests/BufferPoolTests.cpp

@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <gtest/gtest.h>
+#include "dwio/nimble/common/Buffer.h"
+#include "dwio/nimble/common/Exceptions.h"
+#include "folly/executors/CPUThreadPoolExecutor.h"
+#include "velox/common/memory/Memory.h"
+#include "velox/dwio/common/ExecutorBarrier.h"
+
+namespace facebook::nimble::test {
+using MemoryPool = velox::memory::MemoryPool;
+using ExecutorBarrier = velox::dwio::common::ExecutorBarrier;
+
+class BufferPoolTest : public ::testing::Test {
+ protected:
+  static void SetUpTestCase() {}
+
+  void SetUp() override {
+    memPool_ = velox::memory::deprecatedAddDefaultLeafMemoryPool();
+  }
+
+  std::shared_ptr<velox::memory::MemoryPool> memPool_;
+};
+
+TEST_F(BufferPoolTest, CreateBufferPoolWithInitialSize) {
+  auto bufferPool = BufferPool{*memPool_, 10, 10};
+  EXPECT_EQ(bufferPool.size(), 10);
+}
+
+TEST_F(BufferPoolTest, CreateBufferPoolBadMaxPool) {
+  try {
+    auto bufferPool = BufferPool{*memPool_, 0};
+    FAIL();
+  } catch (const NimbleUserError& e) {
+    EXPECT_EQ(e.errorMessage(), "max pool size must be > 0");
+  }
+}
+
+TEST_F(BufferPoolTest, CreateBufferPoolBadInitialSizeMaxSize) {
+  try {
+    auto bufferPool = BufferPool{*memPool_, 1, 2};
+    FAIL();
+  } catch (const NimbleUserError& e) {
+    EXPECT_EQ(e.errorMessage(), "initial pool size must be <= max pool size");
+  }
+}
+
+TEST_F(BufferPoolTest, ReserveBuffer) {
+  auto bufferPool = BufferPool{*memPool_};
+  auto buffer = bufferPool.reserveBuffer();
+}
+
+TEST_F(BufferPoolTest, AddBuffer) {
+  auto bufferPool = BufferPool{*memPool_, 10, 0};
+  auto buffer = bufferPool.reserveBuffer();
+  EXPECT_EQ(bufferPool.size(), 0);
+  bufferPool.addBuffer(std::move(buffer));
+  EXPECT_EQ(bufferPool.size(), 1);
+}
+
+TEST_F(BufferPoolTest, DestroyBufferPool) {
+  auto bufferPoolPtr = std::make_unique<BufferPool>(*memPool_);
+  bufferPoolPtr.reset();
+  EXPECT_NO_THROW();
+}
+
+TEST_F(BufferPoolTest, EmptyBufferPool) {
+  auto bufferPool = BufferPool{*memPool_, 10, 1};
+  EXPECT_EQ(bufferPool.size(), 1);
+
+  auto buffer1 = bufferPool.reserveBuffer();
+  EXPECT_EQ(bufferPool.size(), 0);
+}
+
+TEST_F(BufferPoolTest, OverfillBufferPool) {
+  auto bufferPool = BufferPool{*memPool_, 10, 10};
+  auto throwLambda = [&]() {
+    for (auto i = 0; i < 10; i++) {
+      auto buffer = std::make_unique<Buffer>(*memPool_);
+      bufferPool.addBuffer(std::move(buffer));
+    }
+  };
+
+  EXPECT_NO_THROW(throwLambda());
+  EXPECT_GE(bufferPool.size(), 1);
+}
+
+TEST_F(BufferPoolTest, EmptyFillEmptyBufferPool) {
+  size_t iterations = 10;
+  std::vector<std::unique_ptr<Buffer>> buffers;
+
+  auto bufferPool = BufferPool{*memPool_, iterations, 10};
+  for (auto i = 0; i < iterations; i++) {
+    auto buffer = bufferPool.reserveBuffer();
+    buffers.push_back(std::move(buffer));
+  }
+  EXPECT_EQ(bufferPool.size(), 0);
+  EXPECT_EQ(buffers.size(), iterations);
+  for (auto& buffer : buffers) {
+    bufferPool.addBuffer(std::move(buffer));
+  }
+  EXPECT_EQ(bufferPool.size(), iterations);
+  buffers.clear();
+
+  for (auto i = 0; i < iterations; i++) {
+    auto buffer = bufferPool.reserveBuffer();
+    buffers.push_back(std::move(buffer));
+  }
+  EXPECT_EQ(bufferPool.size(), 0);
+  EXPECT_EQ(buffers.size(), iterations);
+}
+
+TEST_F(BufferPoolTest, ParallelFillPool) {
+  folly::CPUThreadPoolExecutor executor{10};
+  ExecutorBarrier barrier{executor};
+  auto bufferPool = BufferPool{*memPool_, 100};
+  auto fillPool = [&]() {
+    for (auto i = 0; i < 10; i++) {
+      EXPECT_NO_THROW(
+          auto buffer = bufferPool.reserveBuffer();
+          std::this_thread::sleep_for(std::chrono::milliseconds{1000});
+          bufferPool.addBuffer(std::move(buffer)););
+    }
+  };
+
+  for (auto i = 0; i < 10; i++) {
+    barrier.add(fillPool);
+  }
+
+  barrier.waitAll();
+}
+
+} // namespace facebook::nimble::test

dwio/nimble/velox/FieldWriter.cpp

@@ -290,9 +290,10 @@ class SimpleFieldWriter : public FieldWriter {
   void write(const velox::VectorPtr& vector, const OrderedRanges& ranges)
       override {
     auto size = ranges.size();
-    auto& buffer = context_.stringBuffer();
     auto& data = valuesStream_.mutableData();
-
+    auto bufferPtr = context_.bufferPool().reserveBuffer();
+    auto bufferGuard = folly::makeGuard(
+        [&]() { context_.bufferPool().addBuffer(std::move(bufferPtr)); });
     if (auto flat = vector->asFlatVector<SourceType>()) {
       valuesStream_.ensureNullsCapacity(flat->mayHaveNulls(), size);
       bool rangeCopied = false;
@@ -332,6 +333,7 @@ class SimpleFieldWriter : public FieldWriter {
             valuesStream_.mutableNonNulls(),
             Flat<SourceType>{vector},
             [&](SourceType value) {
+              auto& buffer = *bufferPtr;
               data.push_back(
                   C::convert(value, buffer, valuesStream_.extraMemory()));
             });
@@ -349,6 +351,7 @@ class SimpleFieldWriter : public FieldWriter {
           valuesStream_.mutableNonNulls(),
           Decoded<SourceType>{decoded},
           [&](SourceType value) {
+            auto& buffer = *bufferPtr;
             data.push_back(
                 C::convert(value, buffer, valuesStream_.extraMemory()));
           });

dwio/nimble/velox/FieldWriter.h

@@ -95,18 +95,22 @@ struct FieldWriterContext {
       velox::memory::MemoryPool& memoryPool,
       std::shared_ptr<folly::Executor> writeExecutor = nullptr,
       std::unique_ptr<velox::memory::MemoryReclaimer> reclaimer = nullptr,
-      std::function<void(void)> vectorDecoderVisitor = []() {})
+      std::function<void(void)> vectorDecoderVisitor = []() {},
+      size_t maxPoolSize = std::thread::hardware_concurrency(),
+      size_t initialBufferCount = 20)
       : bufferMemoryPool{memoryPool.addLeafChild(
             "field_writer_buffer",
             true,
             std::move(reclaimer))},
         writeExecutor{std::move(writeExecutor)},
         inputBufferGrowthPolicy{
             DefaultInputBufferGrowthPolicy::withDefaultRanges()},
+        bufferPool_{std::make_unique<BufferPool>(
+            *bufferMemoryPool,
+            maxPoolSize,
+            initialBufferCount)},
         decodingContextPool_{std::make_unique<DecodingContextPool>(
-            std::move(vectorDecoderVisitor))} {
-    resetStringBuffer();
-  }
+            std::move(vectorDecoderVisitor))} {}
 
   std::shared_ptr<velox::memory::MemoryPool> bufferMemoryPool;
   std::shared_ptr<folly::Executor> writeExecutor;
@@ -133,13 +137,8 @@ struct FieldWriterContext {
     return decodingContextPool_->addContext(std::move(context));
   }
 
-  Buffer& stringBuffer() {
-    return *buffer_;
-  }
-
-  // Reset writer context for use by next stripe.
-  void resetStringBuffer() {
-    buffer_ = std::make_unique<Buffer>(*bufferMemoryPool);
+  BufferPool& bufferPool() {
+    return *bufferPool_;
   }
 
   const std::vector<std::unique_ptr<StreamData>>& streams() {
@@ -169,7 +168,7 @@ struct FieldWriterContext {
   }
 
  private:
-  std::unique_ptr<Buffer> buffer_;
+  std::unique_ptr<BufferPool> bufferPool_;
   std::unique_ptr<DecodingContextPool> decodingContextPool_;
   std::vector<std::unique_ptr<StreamData>> streams_;
 };

dwio/nimble/velox/VeloxWriter.cpp

@@ -68,7 +68,9 @@ class WriterContext : public FieldWriterContext {
         memoryPool,
         options.writeExecutor,
         options.reclaimerFactory(),
-        options.vectorDecoderVisitor
+        options.vectorDecoderVisitor,
+        options.maxPoolSize,
+        options.initialBufferCount
         },
         options{std::move(options)},
         logger{this->options.metricsLogger} {
@@ -620,9 +622,6 @@ void VeloxWriter::writeChunk(bool lastChunk) {
     LoggingScope scope{*context_->logger};
     velox::CpuWallTimer veloxTimer{context_->stripeFlushTiming};
 
-    if (!encodingBuffer_) {
-      encodingBuffer_ = std::make_unique<Buffer>(*encodingMemoryPool_);
-    }
     streams_.resize(context_->schemaBuilder.nodeCount());
 
     // When writing null streams, we write the nulls as data, and the stream
@@ -666,9 +665,13 @@ void VeloxWriter::writeChunk(bool lastChunk) {
 
     auto encode = [&](StreamData& streamData) {
       const auto offset = streamData.descriptor().offset();
-      auto encoded = encodeStream(*context_, *encodingBuffer_, streamData);
+      auto bufferPtr = context_->bufferPool().reserveBuffer();
+      auto bufferGuard = folly::makeGuard(
+          [&]() { context_->bufferPool().addBuffer(std::move(bufferPtr)); });
+      auto& buffer = *bufferPtr;
+      auto encoded = encodeStream(*context_, buffer, streamData);
       if (!encoded.empty()) {
-        ChunkedStreamWriter chunkWriter{*encodingBuffer_};
+        ChunkedStreamWriter chunkWriter{buffer};
         NIMBLE_DASSERT(offset < streams_.size(), "Stream offset out of range.");
         auto& stream = streams_[offset];
         for (auto& buffer : chunkWriter.encode(encoded)) {
@@ -780,10 +783,6 @@ uint32_t VeloxWriter::writeStripe() {
     uint64_t startSize = writer_.size();
     writer_.writeStripe(context_->rowsInStripe, std::move(streams_));
     stripeSize = writer_.size() - startSize;
-    encodingBuffer_.reset();
-    // TODO: once chunked string fields are supported, move string buffer
-    // reset to writeChunk()
-    context_->resetStringBuffer();
   }
 
   NIMBLE_ASSERT(

dwio/nimble/velox/VeloxWriter.h

@@ -77,7 +77,6 @@ class VeloxWriter {
   TabletWriter writer_;
   std::unique_ptr<FieldWriter> root_;
 
-  std::unique_ptr<Buffer> encodingBuffer_;
   std::vector<Stream> streams_;
   std::exception_ptr lastException_;
   const velox::common::SpillConfig* const spillConfig_;