zhijiangW commented on a change in pull request #7368: [FLINK-10742][network] Let Netty use Flink's buffers directly in credit-based mode URL: https://github.com/apache/flink/pull/7368#discussion_r375740161
########## File path: flink-runtime/src/main/java/org/apache/flink/runtime/io/network/netty/ZeroCopyNettyMessageDecoder.java ########## @@ -0,0 +1,280 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you 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. + */ + +package org.apache.flink.runtime.io.network.netty; + +import org.apache.flink.runtime.io.network.buffer.Buffer; +import org.apache.flink.shaded.netty4.io.netty.buffer.ByteBuf; +import org.apache.flink.shaded.netty4.io.netty.channel.ChannelHandlerContext; +import org.apache.flink.shaded.netty4.io.netty.channel.ChannelInboundHandlerAdapter; + +import static org.apache.flink.runtime.io.network.netty.NettyMessage.FRAME_HEADER_LENGTH; +import static org.apache.flink.runtime.io.network.netty.NettyMessage.MAGIC_NUMBER; +import static org.apache.flink.util.Preconditions.checkState; + +/** + * Decodes messages from the fragmentary netty buffers. This decoder assumes the + * messages have the following format: + * +-----------------------------------+--------------------------------+ + * | FRAME_HEADER || MESSAGE_HEADER | DATA BUFFER (Optional) | + * +-----------------------------------+--------------------------------+ + * and it decodes each part in order. + * + * This decoder tries best to eliminate copying. For the frame header and message header, + * it only cumulates data when they span multiple input buffers. For the buffer part, it + * copies directly to the input channels to avoid future copying. + * + * The format of the frame header is + * +------------------+------------------+--------+ + * | FRAME LENGTH (4) | MAGIC NUMBER (4) | ID (1) | + * +------------------+------------------+--------+ + */ +public class ZeroCopyNettyMessageDecoder extends ChannelInboundHandlerAdapter { + + private static final int INITIAL_MESSAGE_HEADER_BUFFER_LENGTH = 128; + + /** The parser to parse the message header. */ + private final NettyMessageParser messageParser; + + /** The buffer used to cumulate the frame header part. */ + private ByteBuf frameHeaderBuffer; + + /** The buffer used to receive the message header part. */ + private ByteBuf messageHeaderBuffer; + + /** Which part of the current message is being decoded. */ + private DecodeStep decodeStep; + + /** How many bytes have been decoded in current step. */ + private int decodedBytesOfCurrentStep; + + /** The intermediate state when decoding the current message. */ + private final MessageDecodeIntermediateState intermediateState; + + ZeroCopyNettyMessageDecoder(NettyMessageParser messageParser) { + this.messageParser = messageParser; + this.intermediateState = new MessageDecodeIntermediateState(); + } + + @Override + public void channelActive(ChannelHandlerContext ctx) throws Exception { + super.channelActive(ctx); + + frameHeaderBuffer = ctx.alloc().directBuffer(NettyMessage.FRAME_HEADER_LENGTH); Review comment: I thought of two things here: - Considering the sender side while writing the header of `BufferResponse`, it would always allocate the buffer from the netty allocator for every message. On receiver side now we reuse the same header buffer every time. I am not sure whether we can get benefits from this change. If so, the sender side can also take this way future. If not, we do not need to explicitly maintain these two headers in decoder and it would make things more simple. - If we want to maintain the static header buffers for every decoder, another option is to create direct memory directly, I mean bypassing the netty allocator. Because we ever wondered the memory overhead boost in netty internal management. Another rough thought is if we can make all the memory allocation via the unified allocator, then it is easy for tracing all the memory usages. I mean also allocating these header buffers via existing `NetworkBufferAllocator` if possible. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
