[GitHub] [flink] wsry commented on a change in pull request #8290: [FLINK-12070] [network] Implement new bounded blocking subpartitions

[GitBox <gi...@apache.org>]

wsry commented on a change in pull request #8290: [FLINK-12070] [network] Implement new bounded blocking subpartitions
URL: https://github.com/apache/flink/pull/8290#discussion_r281019849
 
 

 ##########
 File path: flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/BoundedBlockingSubpartition.java
 ##########
 @@ -0,0 +1,416 @@
+/*
+ * 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.runtime.io.network.api.EndOfPartitionEvent;
+import org.apache.flink.runtime.io.network.api.serialization.EventSerializer;
+import org.apache.flink.runtime.io.network.buffer.Buffer;
+import org.apache.flink.runtime.io.network.buffer.BufferConsumer;
+import org.apache.flink.runtime.io.network.partition.BoundedBlockingSubpartitionMemory.Writer;
+import org.apache.flink.util.FlinkRuntimeException;
+import org.apache.flink.util.IOUtils;
+
+import org.apache.flink.shaded.netty4.io.netty.util.internal.PlatformDependent;
+
+import javax.annotation.Nullable;
+import javax.annotation.concurrent.GuardedBy;
+
+import java.io.IOException;
+import java.nio.ByteBuffer;
+import java.nio.channels.FileChannel;
+import java.nio.channels.FileChannel.MapMode;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.StandardOpenOption;
+import java.util.ArrayList;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Set;
+import java.util.stream.Collectors;
+
+import static org.apache.flink.util.Preconditions.checkArgument;
+import static org.apache.flink.util.Preconditions.checkState;
+
+/**
+ * An implementation of the ResultSubpartition for a bounded result transferred
+ * in a blocking manner: The result is first produced, then consumed.
+ * The result can be consumed possibly multiple times.
+ *
+ * <p>The implementation creates a temporary memory mapped file and writes all buffers to that
+ * memory and serves the result from that memory. The kernel backs the mapped memory region
+ * with physical memory and file space incrementally as new pages are filled.
+ */
+class BoundedBlockingSubpartition extends ResultSubpartition {
+
+	/** This lock guards the creation of readers and disposal of the memory mapped file. */
+	private final Object lock = new Object();
+
+	/** The current buffer, may be filled further over time. */
+	@Nullable
+	private BufferConsumer currentBuffer;
+
+	/** The memory that we store the data in, via a memory mapped file. */
+	private final MemoryMappedBuffers memory;
+
+	/** All created and not yet released readers. */
+	@GuardedBy("lock")
+	private final Set<BoundedBlockingSubpartitionReader> readers;
+
+	/** Counter for the number of data buffers (not events!) written. */
+	private int numDataBuffersWritten;
+
+	/** The counter for the number of data buffers and events. */
+	private int numBuffersAndEventsWritten;
+
+	/** Flag indicating whether the writing has finished and this is now available for read. */
+	private boolean isFinished;
+
+	/** Flag indicating whether the subpartition has been released. */
+	private boolean isReleased;
+
+	/**
+	 * Common constructor.
+	 */
+	public BoundedBlockingSubpartition(
+			int index,
+			ResultPartition parent,
+			Path filePath) throws IOException {
+		this(index, parent, filePath, Integer.MAX_VALUE);
+	}
+
+	/**
+	 * Constructor for testing. By default regions are rolled over at 2GB (max size of direct buffers
+	 * in Java). This constructor allows tests to pass in a smaller threshold to test rolling over
+	 * without having to actually produce more than 2GB during testing.
+	 */
+	@VisibleForTesting
+	BoundedBlockingSubpartition(
+			int index,
+			ResultPartition parent,
+			Path filePath,
+			int maxMMapRegionSize) throws IOException {
+
+		super(index, parent);
+
+		final FileChannel fc = FileChannel.open(filePath,
+				StandardOpenOption.READ, StandardOpenOption.WRITE, StandardOpenOption.CREATE_NEW);
+
+		this.memory = new MemoryMappedBuffers(filePath, fc, maxMMapRegionSize);
+		this.readers = new HashSet<>();
+	}
+
+	// ------------------------------------------------------------------------
+
+	/**
+	 * Checks if writing is finished.
+	 * Readers cannot be created until writing is finished, and no further writes can happen after that.
+	 */
+	public boolean isFinished() {
+		return isFinished;
+	}
+
+	@Override
+	public boolean isReleased() {
+		return isReleased;
+	}
+
+	@Override
+	public boolean add(BufferConsumer bufferConsumer) throws IOException {
+		if (isFinished()) {
+			bufferConsumer.close();
+			return false;
+		}
+
+		flushCurrentBuffer();
+		currentBuffer = bufferConsumer;
+		return true;
+	}
+
+	@Override
+	public void flush() {
+		// unfortunately, the signature of flush does not allow for any exceptions, so we
+		// need to do this discouraged pattern of runtime exception wrapping
+		try {
+			flushCurrentBuffer();
+		}
+		catch (IOException e) {
+			throw new FlinkRuntimeException(e.getMessage(), e);
+		}
+	}
+
+	private void flushCurrentBuffer() throws IOException {
+		if (currentBuffer != null) {
+			writeAndCloseBufferConsumer(currentBuffer);
+			currentBuffer = null;
+		}
+	}
+
+	private void writeAndCloseBufferConsumer(BufferConsumer bufferConsumer) throws IOException {
+		try {
+			final Buffer buffer = bufferConsumer.build();
+			try {
+				memory.writeBuffer(buffer);
+
+				numBuffersAndEventsWritten++;
+				if (buffer.isBuffer()) {
+					numDataBuffersWritten++;
+				}
+			}
+			finally {
+				buffer.recycleBuffer();
+			}
+		}
+		finally {
+			bufferConsumer.close();
+		}
+	}
+
+	@Override
+	public void finish() throws IOException {
+		synchronized (lock) {
+			checkState(!isReleased, "data partition already released");
+			checkState(!isFinished, "data partition already finished");
+
+			isFinished = true;
+			flushCurrentBuffer();
+			writeAndCloseBufferConsumer(EventSerializer.toBufferConsumer(EndOfPartitionEvent.INSTANCE));
+			memory.finishWrite();
+		}
+	}
+
+	@Override
+	public void release() throws IOException {
+		synchronized (lock) {
+			if (isReleased) {
+				return;
+			}
+
+			isReleased = true;
+			isFinished = true; // for fail fast writes
+
+			checkReaderReferencesAndDispose();
+		}
+	}
+
+	@Override
+	public ResultSubpartitionView createReadView(BufferAvailabilityListener availability) throws IOException {
+		synchronized (lock) {
+			checkState(!isReleased, "data partition already released");
+			checkState(isFinished, "writing of blocking partition not yet finished");
+
+			availability.notifyDataAvailable();
+
+			final List<ByteBuffer> allBuffers = memory.getFullBuffers();
+			final BoundedBlockingSubpartitionReader reader = new BoundedBlockingSubpartitionReader(
+					this, allBuffers, numDataBuffersWritten);
+			readers.add(reader);
+			return reader;
+		}
+	}
+
+	void releaseReaderReference(BoundedBlockingSubpartitionReader reader) throws IOException {
+		synchronized (lock) {
+			if (readers.remove(reader) && isReleased) {
+				checkReaderReferencesAndDispose();
+			}
+		}
+	}
+
+	@GuardedBy("lock")
+	private void checkReaderReferencesAndDispose() throws IOException {
+		assert Thread.holdsLock(lock);
+
+		// To avoid lingering memory mapped files (large resource footprint), we don't
+		// wait for GC to unmap the files, but use a Netty utility to directly unmap the file.
+		// To avoid segmentation faults, we need to wait until all readers have been released.
+
+		if (readers.isEmpty()) {
+			memory.unmapAndDispose();
+		}
+	}
+
+	// ------------------------------ legacy ----------------------------------
+
+	@Override
+	public int releaseMemory() throws IOException {
+		return 0;
+	}
+
+	// ---------------------------- statistics --------------------------------
+
+	@Override
+	public int unsynchronizedGetNumberOfQueuedBuffers() {
+		return 0;
+	}
+
+	@Override
+	protected long getTotalNumberOfBuffers() {
+		return numBuffersAndEventsWritten;
+	}
+
+	@Override
+	protected long getTotalNumberOfBytes() {
+		return memory.getSize();
+	}
+
+	int getBuffersInBacklog() {
+		return numDataBuffersWritten;
+	}
+
+	// ------------------------------------------------------------------------
+
+	/**
+	 * This class is largely a workaround for the fact that a memory mapped region in Java can cannot
+	 * be larger than 2GB (== signed 32 bit int max value).
+	 *
+	 * <p>The class takes {@link Buffer}s and writes them to several memory mapped region, using the
+	 * {@link BoundedBlockingSubpartitionMemory} class.
+	 *
+	 * <p>This class performs absolutely no synchronization and relies on single threaded
+	 * or externally synchronized access.
+	 */
+	private static final class MemoryMappedBuffers {
+
+		/** Memory mappings should be at the granularity of page sizes, for efficiency. */
+		private static final int PAGE_SIZE = PageSizeUtil.getSystemPageSizeOrConservativeMultiple();
+
+		/** The encoder to the current memory mapped region we are writing to. */
+		private BoundedBlockingSubpartitionMemory.Writer currentBuffer;
+
+		/** All memory mapped regions that are already full (completed). */
+		private final ArrayList<ByteBuffer> fullBuffers;
+
+		/** The file channel backing the memory mapped file. */
+		private final FileChannel file;
+
+		/** The path of the memory mapped file. */
+		private final Path filePath;
+
+		/** The offset where the next mapped region should start. */
+		private long nextMappingOffset;
+
+		/** The size of each mapped region. */
+		private final long mappingSize;
+
+		MemoryMappedBuffers(Path filePath, FileChannel file, int maxSizePerByteBuffer) throws IOException {
+			this.filePath = filePath;
+			this.file = file;
+			this.mappingSize = alignSize(maxSizePerByteBuffer);
+			this.fullBuffers = new ArrayList<>(4);
+
+			rollOverToNextBuffer();
+		}
+
+		void writeBuffer(Buffer buffer) throws IOException {
+			if (currentBuffer.writeBuffer(buffer)) {
+				return;
+			}
+
+			rollOverToNextBuffer();
+
+			if (!currentBuffer.writeBuffer(buffer)) {
+				throwTooLargeBuffer(buffer);
+			}
+		}
+
+		List<ByteBuffer> getFullBuffers() {
+			return fullBuffers.stream()
+					.map(ByteBuffer::slice)
+					.collect(Collectors.toList());
+		}
+
+		/**
+		 * Finishes the current region and prevents further writes.
+		 */
+		void finishWrite() throws IOException {
+			fullBuffers.add(currentBuffer.complete());
+			currentBuffer = null; // fail further writes fast
+			file.close(); // won't map further regions from now on
+		}
+
+		/**
+		 * Unmaps the file from memory and deletes the file.
+		 * After calling this method, access to any ByteBuffer obtained from this instance
+		 * will cause a segmentation fault.
+		 */
+		void unmapAndDispose() throws IOException {
+			IOUtils.closeQuietly(file); // in case we dispose before finishing writes
+
+			for (ByteBuffer bb : fullBuffers) {
+				PlatformDependent.freeDirectBuffer(bb);
+			}
+			fullBuffers.clear();
+
+			if (currentBuffer != null) {
+				PlatformDependent.freeDirectBuffer(currentBuffer.complete());
+				currentBuffer = null;
+			}
+
+			// To make this compatible with all versions of Windows, we must wait with
+			// deleting the file until it is unmapped.
+			// See also https://stackoverflow.com/questions/11099295/file-flag-delete-on-close-and-memory-mapped-files/51649618#51649618
+
+			Files.delete(filePath);
+		}
+
+		/**
+		 * Gets the number of bytes of all written data (including the metadata in the buffer headers).
+		 */
+		long getSize() {
+			long size = 0L;
+			for (ByteBuffer bb : fullBuffers) {
+				size += bb.capacity();
+			}
+			if (currentBuffer != null) {
+				size += currentBuffer.getNumBytes();
+			}
+			return size;
+		}
+
+		private void rollOverToNextBuffer() throws IOException {
+			if (currentBuffer != null) {
+				// we need to remember the original buffers, not any slices.
+				// slices have no cleaner, which we need to trigger explicit unmapping
+				fullBuffers.add(currentBuffer.complete());
+			}
+
+			final ByteBuffer mapped = file.map(MapMode.READ_WRITE, nextMappingOffset, mappingSize);
+			currentBuffer = new Writer(mapped);
+			nextMappingOffset += mappingSize;
 
 Review comment:
   The previous mapped bytebuffer may not be full, this may lead to file holes between data blocks, though the read view can be created correctly, is it by design.

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services