gaoyunhaii commented on a change in pull request #13924:
URL: https://github.com/apache/flink/pull/13924#discussion_r601023772
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/SortMergeResultPartition.java
##########
@@ -455,4 +458,28 @@ public int getNumberOfQueuedBuffers(int
targetSubpartition) {
PartitionedFile getResultFile() {
return resultFile;
}
+
+ @VisibleForTesting
+ SortMergeResultPartitionReader getPartitionReader() {
+ return partitionReader;
+ }
+
+ private int[] getRandomSubpartitionOrder(int numSubpartitions) {
Review comment:
Could we achieve the random order with
```
List<Integer> list = IntStream.range(0,
numSubpartitions).boxed().collect(Collectors.toList());
Collections.shuffle(list);
return list.toArray();
```
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/SortMergeResultPartition.java
##########
@@ -256,7 +267,7 @@ private SortBuffer getSortBuffer() {
numSubpartitions,
networkBufferSize,
numBuffersForSort,
- null);
+ subpartitionOrder);
Review comment:
Although might be not related to this PR, it seems to me that it might
be better we always pass the random order and makes this parameter nonnull. We
might also add a test constructor for the tests. But it should be a style issue
and I think I would not argue strongly for it.
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/SortMergeResultPartitionReader.java
##########
@@ -0,0 +1,350 @@
+/*
+ * 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.partition;
+
+import org.apache.flink.annotation.VisibleForTesting;
+import org.apache.flink.core.memory.MemorySegment;
+import org.apache.flink.runtime.io.disk.BatchShuffleReadBufferPool;
+import org.apache.flink.runtime.io.network.buffer.BufferRecycler;
+import org.apache.flink.util.IOUtils;
+
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import javax.annotation.concurrent.GuardedBy;
+
+import java.io.IOException;
+import java.nio.channels.FileChannel;
+import java.nio.file.Path;
+import java.nio.file.StandardOpenOption;
+import java.util.ArrayDeque;
+import java.util.ArrayList;
+import java.util.Collection;
+import java.util.Collections;
+import java.util.HashSet;
+import java.util.List;
+import java.util.PriorityQueue;
+import java.util.Queue;
+import java.util.Set;
+import java.util.concurrent.Executor;
+import java.util.concurrent.TimeoutException;
+
+import static org.apache.flink.util.Preconditions.checkNotNull;
+import static org.apache.flink.util.Preconditions.checkState;
+
+/**
+ * Data reader for {@link SortMergeResultPartition} which can read data for
all downstream tasks
+ * consuming the corresponding {@link SortMergeResultPartition}. It always
tries to read shuffle
+ * data in order of file offset, which maximums the sequential read so can
improve the blocking
+ * shuffle performance.
+ */
+public class SortMergeResultPartitionReader implements Runnable,
BufferRecycler {
Review comment:
We may change the name to something like
`SortMergeResultPartitionReadScheduler` ? Since this class do not have actual
read.
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/SortMergeResultPartition.java
##########
@@ -120,8 +127,10 @@ public SortMergeResultPartition(
bufferCompressor,
bufferPoolFactory);
- this.networkBufferSize = networkBufferSize;
- this.numDataBuffers = new int[numSubpartitions];
+ this.networkBufferSize = readBufferPool.getBufferSize();
+ this.subpartitionOrder = getRandomSubpartitionOrder(numSubpartitions);
Review comment:
+1 for add comments for the usage of the order
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/SortMergeResultPartitionReader.java
##########
@@ -0,0 +1,350 @@
+/*
+ * 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.partition;
+
+import org.apache.flink.annotation.VisibleForTesting;
+import org.apache.flink.core.memory.MemorySegment;
+import org.apache.flink.runtime.io.disk.BatchShuffleReadBufferPool;
+import org.apache.flink.runtime.io.network.buffer.BufferRecycler;
+import org.apache.flink.util.IOUtils;
+
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import javax.annotation.concurrent.GuardedBy;
+
+import java.io.IOException;
+import java.nio.channels.FileChannel;
+import java.nio.file.Path;
+import java.nio.file.StandardOpenOption;
+import java.util.ArrayDeque;
+import java.util.ArrayList;
+import java.util.Collection;
+import java.util.Collections;
+import java.util.HashSet;
+import java.util.List;
+import java.util.PriorityQueue;
+import java.util.Queue;
+import java.util.Set;
+import java.util.concurrent.Executor;
+import java.util.concurrent.TimeoutException;
+
+import static org.apache.flink.util.Preconditions.checkNotNull;
+import static org.apache.flink.util.Preconditions.checkState;
+
+/**
+ * Data reader for {@link SortMergeResultPartition} which can read data for
all downstream tasks
+ * consuming the corresponding {@link SortMergeResultPartition}. It always
tries to read shuffle
+ * data in order of file offset, which maximums the sequential read so can
improve the blocking
+ * shuffle performance.
+ */
+public class SortMergeResultPartitionReader implements Runnable,
BufferRecycler {
+
+ private static final Logger LOG =
LoggerFactory.getLogger(SortMergeResultPartitionReader.class);
+
+ /** Lock used to synchronize multi-thread access to thread-unsafe fields.
*/
+ private final Object lock;
+
+ /** Buffer pool from which to allocate buffers for shuffle data reading. */
+ private final BatchShuffleReadBufferPool bufferPool;
+
+ /** Executor to run the shuffle data reading task. */
+ private final Executor ioExecutor;
+
+ /** Maximum number of buffers can be allocated by this partition reader. */
+ private final int maxRequestedBuffers;
+
+ /** All failed subpartition readers to be released. */
+ @GuardedBy("lock")
+ private final Set<SortMergeSubpartitionReader> failedReaders = new
HashSet<>();
+
+ /** All readers waiting to read data of different subpartitions. */
+ @GuardedBy("lock")
+ private final Set<SortMergeSubpartitionReader> allReaders = new
HashSet<>();
+
+ /** File channel shared by all subpartitions to read data from. */
+ @GuardedBy("lock")
+ private FileChannel dataFileChannel;
+
+ /** File channel shared by all subpartitions to read index from. */
+ @GuardedBy("lock")
+ private FileChannel indexFileChannel;
+
+ /** Number of buffers already allocated and still not recycled by this
partition reader. */
+ @GuardedBy("lock")
+ private int numRequestedBuffers;
+
+ /**
+ * Whether the data reading task is currently running or not. This flag is
used when trying to
+ * submit the data reading task.
+ */
+ @GuardedBy("lock")
+ private boolean isRunning;
+
+ /** Whether this reader has been released or not. */
+ @GuardedBy("lock")
+ private boolean isReleased;
+
+ public SortMergeResultPartitionReader(
+ BatchShuffleReadBufferPool bufferPool, Executor ioExecutor, Object
lock) {
+ this.lock = checkNotNull(lock);
+ this.bufferPool = checkNotNull(bufferPool);
+ this.ioExecutor = checkNotNull(ioExecutor);
+ // one partition reader can consume at most 32M buffers for data read,
it is only an
+ // empirical value which can not be configured currently
+ this.maxRequestedBuffers = Math.max(1, 4 *
bufferPool.getNumBuffersPerRequest());
+
+ bufferPool.initialize();
+ }
+
+ @Override
+ public synchronized void run() {
+ Set<SortMergeSubpartitionReader> finishedReaders = new HashSet<>();
+ Queue<SortMergeSubpartitionReader> availableReaders =
getAvailableReaders();
+
+ Queue<MemorySegment> buffers = new ArrayDeque<>();
+ if (!availableReaders.isEmpty()) {
+ try {
+ buffers.addAll(bufferPool.requestBuffers());
+ } catch (Throwable throwable) {
+ // this can happen only when the buffer pool is destroyed or
buffer request timeout
+ if (!(throwable instanceof TimeoutException) ||
numRequestedBuffers <= 0) {
+ failSubpartitionReaders(availableReaders, throwable);
+ LOG.error("Failed to request buffers for data reading.",
throwable);
Review comment:
Might directly `return` here ? It seems in this case there should be no
read for any available readers.
Also, what would happen with `TimeoutException` ? Don't we need to fail the
subpartition readers in this case ? If there is indeed deadlock and if the
exception is ignored, it seems the downstream tasks would hang forever ?
--
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]
