[SPARK-4516] Avoid allocating Netty PooledByteBufAllocators unnecessarily

Turns out we are allocating an allocator pool for every TransportClient (which means that the number increases with the number of nodes in the cluster), when really we should just reuse one for all clients.

This patch, as expected, greatly decreases off-heap memory allocation, and appears to make allocation only proportional to the number of cores.

Author: Aaron Davidson <aaron@databricks.com>

Closes #3465 from aarondav/fewer-pools and squashes the following commits:

36c49da [Aaron Davidson] [SPARK-4516] Avoid allocating unnecessarily Netty PooledByteBufAllocators

(cherry picked from commit 346bc17a2ec8fc9e6eaff90733aa1e8b6b46883e)
Signed-off-by: Patrick Wendell <pwendell@gmail.com>

2 files changed, 8 insertions, 10 deletions

diff --git a/network/common/src/main/java/org/apache/spark/network/client/TransportClientFactory.java b/network/common/src/main/java/org/apache/spark/network/client/TransportClientFactory.java
index 76bce85928..9afd5decd5 100644
--- a/network/common/src/main/java/org/apache/spark/network/client/TransportClientFactory.java
+++ b/network/common/src/main/java/org/apache/spark/network/client/TransportClientFactory.java
@@ -19,7 +19,6 @@ package org.apache.spark.network.client;
 
 import java.io.Closeable;
 import java.io.IOException;
-import java.lang.reflect.Field;
 import java.net.InetSocketAddress;
 import java.net.SocketAddress;
 import java.util.List;
@@ -37,7 +36,6 @@ import io.netty.channel.ChannelInitializer;
 import io.netty.channel.ChannelOption;
 import io.netty.channel.EventLoopGroup;
 import io.netty.channel.socket.SocketChannel;
-import io.netty.util.internal.PlatformDependent;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -67,6 +65,7 @@ public class TransportClientFactory implements Closeable {
 
   private final Class<? extends Channel> socketChannelClass;
   private EventLoopGroup workerGroup;
+  private PooledByteBufAllocator pooledAllocator;
 
   public TransportClientFactory(
       TransportContext context,
@@ -80,6 +79,8 @@ public class TransportClientFactory implements Closeable {
     this.socketChannelClass = NettyUtils.getClientChannelClass(ioMode);
     // TODO: Make thread pool name configurable.
     this.workerGroup = NettyUtils.createEventLoop(ioMode, conf.clientThreads(), "shuffle-client");
+    this.pooledAllocator = NettyUtils.createPooledByteBufAllocator(
+      conf.preferDirectBufs(), false /* allowCache */, conf.clientThreads());
   }
 
   /**
@@ -115,11 +116,8 @@ public class TransportClientFactory implements Closeable {
        // Disable Nagle's Algorithm since we don't want packets to wait
       .option(ChannelOption.TCP_NODELAY, true)
       .option(ChannelOption.SO_KEEPALIVE, true)
-      .option(ChannelOption.CONNECT_TIMEOUT_MILLIS, conf.connectionTimeoutMs());
-
-    // Use pooled buffers to reduce temporary buffer allocation
-    bootstrap.option(ChannelOption.ALLOCATOR, NettyUtils.createPooledByteBufAllocator(
-      conf.preferDirectBufs(), false /* allowCache */, conf.clientThreads()));
+      .option(ChannelOption.CONNECT_TIMEOUT_MILLIS, conf.connectionTimeoutMs())
+      .option(ChannelOption.ALLOCATOR, pooledAllocator);
 
     final AtomicReference<TransportClient> clientRef = new AtomicReference<TransportClient>();
 
diff --git a/network/common/src/main/java/org/apache/spark/network/util/NettyUtils.java b/network/common/src/main/java/org/apache/spark/network/util/NettyUtils.java
index 5c654a6fd6..b3991a6577 100644
--- a/network/common/src/main/java/org/apache/spark/network/util/NettyUtils.java
+++ b/network/common/src/main/java/org/apache/spark/network/util/NettyUtils.java
@@ -109,9 +109,9 @@ public class NettyUtils {
 
   /**
    * Create a pooled ByteBuf allocator but disables the thread-local cache. Thread-local caches
-   * are disabled because the ByteBufs are allocated by the event loop thread, but released by the
-   * executor thread rather than the event loop thread. Those thread-local caches actually delay
-   * the recycling of buffers, leading to larger memory usage.
+   * are disabled for TransportClients because the ByteBufs are allocated by the event loop thread,
+   * but released by the executor thread rather than the event loop thread. Those thread-local
+   * caches actually delay the recycling of buffers, leading to larger memory usage.
    */
   public static PooledByteBufAllocator createPooledByteBufAllocator(
       boolean allowDirectBufs,