Flink

NetworkEnvironment 是一个TaskManager对应一个,而不是一个task对应一个

其中最关键的是networkBufferPool,
operator产生的中间结果,ResultPartition,或是input数据,InputGate
都是需要memory来暂存的,这就需要networkBufferPool来管理这部分内存

/**
 * Network I/O components of each {@link TaskManager} instance. The network environment contains
 * the data structures that keep track of all intermediate results and all data exchanges.
 *
 * When initialized, the NetworkEnvironment will allocate the network buffer pool.
 * All other components (netty, intermediate result managers, ...) are only created once the
 * environment is "associated" with a TaskManager and JobManager. This happens as soon as the
 * TaskManager actor gets created and registers itself at the JobManager.
 */
public class NetworkEnvironment {
    
    private final NetworkEnvironmentConfiguration configuration;
    
    private final NetworkBufferPool networkBufferPool;
    
    private ConnectionManager connectionManager;
    
    private ResultPartitionManager partitionManager;
    
    private ResultPartitionConsumableNotifier partitionConsumableNotifier;
    
    /**
     * ExecutionEnvironment which is used to execute remote calls with the
     * {@link JobManagerResultPartitionConsumableNotifier}
     */
    private final ExecutionContext executionContext;
    
    /**
     * Initializes all network I/O components.
     */
    public NetworkEnvironment(
        ExecutionContext executionContext,
        FiniteDuration jobManagerTimeout,
        NetworkEnvironmentConfiguration config) throws IOException {
        
        // create the network buffers - this is the operation most likely to fail upon
        // mis-configuration, so we do this first
        try {
            networkBufferPool = new NetworkBufferPool(config.numNetworkBuffers(),
                config.networkBufferSize(), config.memoryType());
        }
        catch (Throwable t) {
            throw new IOException("Cannot allocate network buffer pool: " + t.getMessage(), t);
        }
    }
}

 

NetworkBufferPool

先看看networkBufferPool,

首先,它管理了一堆的BufferPool,而不是buffer,因为一个task manager只有一个networkBufferPool,所以对于每个task,需要分配一个buffer pool
再者,它的内存管理和memory manager一样的模式,从heap或off-heap申请相应数量的segments放入availableMemorySegments中

image

可以看到底下黄色部分,就是分配给networkBufferPool的heap

/**
 * The NetworkBufferPool is a fixed size pool of {@link MemorySegment} instances
 * for the network stack.
 *
 * The NetworkBufferPool creates {@link LocalBufferPool}s from which the individual tasks draw
 * the buffers for the network data transfer. When new local buffer pools are created, the
 * NetworkBufferPool dynamically redistributes the buffers between the pools.
 */
public class NetworkBufferPool implements BufferPoolFactory {

    private final int totalNumberOfMemorySegments; //该Pool所管理的所有MemorySegment的数量

    private final int memorySegmentSize; //memorySegment的大小,size

    private final Queue<MemorySegment> availableMemorySegments; //可用的MemorySegment队列

    private final Set<LocalBufferPool> managedBufferPools = new HashSet<LocalBufferPool>(); //管理一组LocalBufferPool,每个task需要分配一个

    public final Set<LocalBufferPool> allBufferPools = new HashSet<LocalBufferPool>();

    private int numTotalRequiredBuffers;

    /**
     * Allocates all {@link MemorySegment} instances managed by this pool.
     */
    public NetworkBufferPool(int numberOfSegmentsToAllocate, int segmentSize, MemoryType memoryType) {
        
        this.totalNumberOfMemorySegments = numberOfSegmentsToAllocate;
        this.memorySegmentSize = segmentSize;

        final long sizeInLong = (long) segmentSize;

        try {
            this.availableMemorySegments = new ArrayBlockingQueue<MemorySegment>(numberOfSegmentsToAllocate); //availableMemorySegments按totalNumberOfMemorySegments分配
        }
        catch (OutOfMemoryError err) {
        
        }

        try {
            if (memoryType == MemoryType.HEAP) { //可以选择是从heap或off-heap分配
                for (int i = 0; i < numberOfSegmentsToAllocate; i++) {
                    byte[] memory = new byte[segmentSize];
                    availableMemorySegments.add(MemorySegmentFactory.wrapPooledHeapMemory(memory, null));
                }
            }
            else if (memoryType == MemoryType.OFF_HEAP) {
                for (int i = 0; i < numberOfSegmentsToAllocate; i++) {
                    ByteBuffer memory = ByteBuffer.allocateDirect(segmentSize);
                    availableMemorySegments.add(MemorySegmentFactory.wrapPooledOffHeapMemory(memory, null));
                }
            }
            else {
                throw new IllegalArgumentException("Unknown memory type " + memoryType);
            }
        }
    }

    public MemorySegment requestMemorySegment() {
        return availableMemorySegments.poll(); //request就是从availableMemorySegments里面取一个
    }

    // This is not safe with regard to destroy calls, but it does not hurt, because destroy happens
    // only once at clean up time (task manager shutdown).
    public void recycle(MemorySegment segment) {
        availableMemorySegments.add(segment); //而回收就是放回availableMemorySegments
    }

    @Override
    public BufferPool createBufferPool(int numRequiredBuffers, boolean isFixedSize) throws IOException {
        // It is necessary to use a separate lock from the one used for buffer
        // requests to ensure deadlock freedom for failure cases.
        synchronized (factoryLock) {

            // Ensure that the number of required buffers can be satisfied.
            // With dynamic memory management this should become obsolete.
            if (numTotalRequiredBuffers + numRequiredBuffers > totalNumberOfMemorySegments) { //确定已经required的加上这次require的没有超过总量
                throw new IOException(String.format("Insufficient number of network buffers: " +
                                "required %d, but only %d available. The total number of network " +
                                "buffers is currently set to %d. You can increase this " +
                                "number by setting the configuration key '%s'.",
                        numRequiredBuffers,
                        totalNumberOfMemorySegments - numTotalRequiredBuffers,
                        totalNumberOfMemorySegments,
                        ConfigConstants.TASK_MANAGER_NETWORK_NUM_BUFFERS_KEY));
            }

            this.numTotalRequiredBuffers += numRequiredBuffers; //增加numTotalRequiredBuffers

            // We are good to go, create a new buffer pool and redistribute
            // non-fixed size buffers.
            LocalBufferPool localBufferPool = new LocalBufferPool(this, numRequiredBuffers); //创建LocalBufferPool,这时并不会把segement给他,request是lazy的

            // The fixed size pools get their share of buffers and don't change
            // it during their lifetime.
            if (!isFixedSize) { //如果不是Fixed,可以动态把多的segment分配出去
                managedBufferPools.add(localBufferPool);
            }

            allBufferPools.add(localBufferPool); //管理localBufferPool

            redistributeBuffers();

            return localBufferPool;
        }
    }
    
    // Must be called from synchronized block
    //目的就是把多余的segement也分配出去,利用起来
    private void redistributeBuffers() throws IOException {
        int numManagedBufferPools = managedBufferPools.size();

        if (numManagedBufferPools == 0) {
            return; // necessary to avoid div by zero when no managed pools
        }

        // All buffers, which are not among the required ones
        int numAvailableMemorySegment = totalNumberOfMemorySegments - numTotalRequiredBuffers; //多的Segments

        // Available excess (not required) buffers per pool
        int numExcessBuffersPerPool = numAvailableMemorySegment / numManagedBufferPools; //多的平均到每个bufferpool

        // Distribute leftover buffers in round robin fashion
        int numLeftoverBuffers = numAvailableMemorySegment % numManagedBufferPools; //余数

        int bufferPoolIndex = 0;

        for (LocalBufferPool bufferPool : managedBufferPools) {
            int leftoverBuffers = bufferPoolIndex++ < numLeftoverBuffers ? 1 : 0; //余数可能是1或0

            bufferPool.setNumBuffers(bufferPool.getNumberOfRequiredMemorySegments() + numExcessBuffersPerPool + leftoverBuffers); //在getNumberOfRequiredMemorySegments的基础上加上多余的
        }
    }

可看到,当一个task需要申请buffer pool时,要先createBufferPool
即,在从availableMemorySegments中取出相应数量的segement,封装成LocalBufferPool,返回

这里有个managedBufferPools,表示bufferpool的size是可以动态变化的,
redistributeBuffers会平均将现有可用的segments分配到所有当前的managedBufferPools上去

 

LocalBufferPool

class LocalBufferPool implements BufferPool {

    private final NetworkBufferPool networkBufferPool; //总的bufferPool

    // The minimum number of required segments for this pool
    private final int numberOfRequiredMemorySegments; //要求申请的MemorySegments的个数,最小个数
    
    // The current size of this pool
    private int currentPoolSize; //实际的MemorySegments的个数,如果不是fixed,可能会多

    // The currently available memory segments. These are segments, which have been requested from
    // the network buffer pool and are currently not handed out as Buffer instances.
    private final Queue<MemorySegment> availableMemorySegments = new ArrayDeque<MemorySegment>(); //缓存MemorySegment的队列

    // Buffer availability listeners, which need to be notified when a Buffer becomes available.
    // Listeners can only be registered at a time/state where no Buffer instance was available.
    private final Queue<EventListener<Buffer>> registeredListeners = new ArrayDeque<EventListener<Buffer>>();

    // Number of all memory segments, which have been requested from the network buffer pool and are
    // somehow referenced through this pool (e.g. wrapped in Buffer instances or as available segments).
    private int numberOfRequestedMemorySegments; //已经分配的MemorySegments的个数

    private boolean isDestroyed;

    private BufferPoolOwner owner; //owner复杂去释放networkBufferPool的buffer

    LocalBufferPool(NetworkBufferPool networkBufferPool, int numberOfRequiredMemorySegments) {
        this.networkBufferPool = networkBufferPool;
        this.numberOfRequiredMemorySegments = numberOfRequiredMemorySegments; //初始化的时候,numberOfRequiredMemorySegments,currentPoolSize相等
        this.currentPoolSize = numberOfRequiredMemorySegments;
    }

    @Override
    public int getMemorySegmentSize() {
        return networkBufferPool.getMemorySegmentSize(); //MemorySegment本身的size
    }

    @Override
    public int getNumBuffers() {
        synchronized (availableMemorySegments) {
            return currentPoolSize; //当前local pool的size
        }
    }

    private Buffer requestBuffer(boolean isBlocking) throws InterruptedException, IOException {
        synchronized (availableMemorySegments) {
            returnExcessMemorySegments(); //把多申请的MemorySegment还回去,如果动态的情况下,是可能的

            boolean askToRecycle = owner != null;

            while (availableMemorySegments.isEmpty()) { //如果availableMemorySegments中没有现成的

                if (numberOfRequestedMemorySegments < currentPoolSize) { //只有在numberOfRequestedMemorySegments小于currentPoolSize,才能继续申请
                    final MemorySegment segment = networkBufferPool.requestMemorySegment(); //从networkBufferPool中申请一块

                    if (segment != null) {
                        numberOfRequestedMemorySegments++;
                        availableMemorySegments.add(segment);

                        continue; //如果申请到继续
                    }
                }

                if (askToRecycle) { //如果申请不到,说明networkBufferPool也没有buffer了
                    owner.releaseMemory(1); //试图让owner去让networkBufferPool释放一块
                }

                if (isBlocking) {
                    availableMemorySegments.wait(2000);
                }
                else {
                    return null;
                }
            }

            return new Buffer(availableMemorySegments.poll(), this);
        }
    }

    @Override
    public void recycle(MemorySegment segment) {
        synchronized (availableMemorySegments) {
            if (isDestroyed || numberOfRequestedMemorySegments > currentPoolSize) {
                returnMemorySegment(segment); //直接还回networkBufferPool
            }
            else {
                EventListener<Buffer> listener = registeredListeners.poll();

                if (listener == null) { //如果没有listen,直接把segment放回availableMemorySegments
                    availableMemorySegments.add(segment);
                    availableMemorySegments.notify(); //触发通知availableMemorySegments有新的segment
                }
                else {
                    try {
                        listener.onEvent(new Buffer(segment, this)); //如果有listener,触发onEvent让listener去处理这个segment
                    }
                    catch (Throwable ignored) {
                        availableMemorySegments.add(segment);
                        availableMemorySegments.notify();
                    }
                }
            }
        }
    }

    @Override
    public void setNumBuffers(int numBuffers) throws IOException {
        synchronized (availableMemorySegments) {
            checkArgument(numBuffers >= numberOfRequiredMemorySegments, "Buffer pool needs at least " + numberOfRequiredMemorySegments + " buffers, but tried to set to " + numBuffers + ".");

            currentPoolSize = numBuffers;

            returnExcessMemorySegments();

            // If there is a registered owner and we have still requested more buffers than our
            // size, trigger a recycle via the owner.
            if (owner != null && numberOfRequestedMemorySegments > currentPoolSize) {
                owner.releaseMemory(numberOfRequestedMemorySegments - numBuffers);
            }
        }
    }
}

 

associateWithTaskManagerAndJobManager

NetworkEnvironment首先需要做的是associate,然后才能用

NetworkEnvironment 中有很多组件,是需要在绑定TaskManagerAndJobManager时,才需要去初始化的

/**
 * This associates the network environment with a TaskManager and JobManager.
 * This will actually start the network components.
 *
 * @param jobManagerGateway Gateway to the JobManager.
 * @param taskManagerGateway Gateway to the TaskManager.
 *
 * @throws IOException Thrown if the network subsystem (Netty) cannot be properly started.
 */
public void associateWithTaskManagerAndJobManager(
        ActorGateway jobManagerGateway,
        ActorGateway taskManagerGateway) throws IOException
{
    synchronized (lock) {

        if (this.partitionConsumableNotifier == null &&
            this.partitionManager == null &&
            this.taskEventDispatcher == null &&
            this.connectionManager == null)
        {
            // good, not currently associated. start the individual components

            LOG.debug("Starting result partition manager and network connection manager");
            this.partitionManager = new ResultPartitionManager();
            this.taskEventDispatcher = new TaskEventDispatcher();
            this.partitionConsumableNotifier = new JobManagerResultPartitionConsumableNotifier(
                executionContext,
                jobManagerGateway,
                taskManagerGateway,
                jobManagerTimeout);

            this.partitionStateChecker = new JobManagerPartitionStateChecker(
                    jobManagerGateway, taskManagerGateway);

            // -----  Network connections  -----
            final Option<NettyConfig> nettyConfig = configuration.nettyConfig();
            connectionManager = nettyConfig.isDefined() ? new NettyConnectionManager(nettyConfig.get())
                                                        : new LocalConnectionManager();

            try {
                LOG.debug("Starting network connection manager");
                connectionManager.start(partitionManager, taskEventDispatcher, networkBufferPool);
            }
            catch (Throwable t) {
                throw new IOException("Failed to instantiate network connection manager: " + t.getMessage(), t);
            }
        }
        else {
            throw new IllegalStateException(
                    "Network Environment is already associated with a JobManager/TaskManager");
        }
    }
}

主要是初始化一系列组件,TaskEventDispatcher,ConnectionManager, ResultPartitionManager

JobManagerResultPartitionConsumableNotifier, JobManagerPartitionStateChecker

 

对于ConnectionManager,这里如果定义了netty,会创建NettyConnectionManager
这里面,主要是初始化Netty client和Netty server
否则是创建LocalConnectionManager

 

而对于ResultPartitionManager, 主要就是用于track所有的result partitions,

核心结构为, Table<ExecutionAttemptID, IntermediateResultPartitionID, ResultPartition> registeredPartitions = HashBasedTable.create();

这个会记录所有的ResultPartition

/**
 * The result partition manager keeps track of all currently produced/consumed partitions of a
 * task manager.
 */
public class ResultPartitionManager implements ResultPartitionProvider {

    private static final Logger LOG = LoggerFactory.getLogger(ResultPartitionManager.class);
    
    public final Table<ExecutionAttemptID, IntermediateResultPartitionID, ResultPartition>
            registeredPartitions = HashBasedTable.create();
    
    private boolean isShutdown;
    
    public void registerResultPartition(ResultPartition partition) throws IOException {
        synchronized (registeredPartitions) {
            checkState(!isShutdown, "Result partition manager already shut down.");
            
            ResultPartitionID partitionId = partition.getPartitionId();
            
            ResultPartition previous = registeredPartitions.put(partitionId.getProducerId(), partitionId.getPartitionId(), partition);
        }
    }
}

 

JobManagerResultPartitionConsumableNotifier,比较关键,通知JobMananger,ResultPartition已经ready,可以开始consume

private static class JobManagerResultPartitionConsumableNotifier implements ResultPartitionConsumableNotifier {

    /**
     * {@link ExecutionContext} which is used for the failure handler of {@link ScheduleOrUpdateConsumers}
     * messages.
     */
    private final ExecutionContext executionContext;

    private final ActorGateway jobManager;

    private final ActorGateway taskManager;

    private final FiniteDuration jobManagerMessageTimeout;

    @Override
    public void notifyPartitionConsumable(JobID jobId, final ResultPartitionID partitionId) {

        final ScheduleOrUpdateConsumers msg = new ScheduleOrUpdateConsumers(jobId, partitionId); //通知jobmanager,去deployconsumer

        Future<Object> futureResponse = jobManager.ask(msg, jobManagerMessageTimeout); //等JobManager的回复

        futureResponse.onFailure(new OnFailure() { //失败,即无法deploy consumer
            @Override
            public void onFailure(Throwable failure) {
                LOG.error("Could not schedule or update consumers at the JobManager.", failure);

                // Fail task at the TaskManager
                FailTask failMsg = new FailTask(
                        partitionId.getProducerId(),
                        new RuntimeException("Could not notify JobManager to schedule or update consumers",
                                failure));

                taskManager.tell(failMsg);
            }
        }, executionContext);
    }
}

 

RegisterTask

在NetworkEnvironment中比较重要的操作,是注册task,需要为task的resultpartition和inputgate分配bufferpool

public void registerTask(Task task) throws IOException {
    final ResultPartition[] producedPartitions = task.getProducedPartitions();
    final ResultPartitionWriter[] writers = task.getAllWriters();

    ResultPartitionConsumableNotifier jobManagerNotifier;

    synchronized (lock) {
        for (int i = 0; i < producedPartitions.length; i++) {
            final ResultPartition partition = producedPartitions[i];
            final ResultPartitionWriter writer = writers[i];

            // Buffer pool for the partition
            BufferPool bufferPool = null;

            try {
                bufferPool = networkBufferPool.createBufferPool(partition.getNumberOfSubpartitions(), false); //创建LocalPool,注意Reqired的segment数目是Subpartitions的数目,即一个subP一个segment
                partition.registerBufferPool(bufferPool); //把localPool注册到ResultPartition

                partitionManager.registerResultPartition(partition); //注册到partitionManager
            }

            // Register writer with task event dispatcher
            taskEventDispatcher.registerWriterForIncomingTaskEvents(writer.getPartitionId(), writer);
        }

        // Setup the buffer pool for each buffer reader
        final SingleInputGate[] inputGates = task.getAllInputGates();

        for (SingleInputGate gate : inputGates) {
            BufferPool bufferPool = null;

            try {
                bufferPool = networkBufferPool.createBufferPool(gate.getNumberOfInputChannels(), false);
                gate.setBufferPool(bufferPool);
            }

        // Copy the reference to prevent races with concurrent shut downs
        jobManagerNotifier = partitionConsumableNotifier;
    }

    for (ResultPartition partition : producedPartitions) {
        // Eagerly notify consumers if required.
        if (partition.getEagerlyDeployConsumers()) { //如果是eager的方式,通知jobmanager,可以deploy consumer了
            jobManagerNotifier.notifyPartitionConsumable(
                    partition.getJobId(), partition.getPartitionId());
        }
    }
}
原文地址:https://www.cnblogs.com/fxjwind/p/6150231.html