HDFS-11187. Optimize disk access for last partial chunk checksum of Finalized replica. Contributed by Gabor Bota, Wei-Chiu Chuang.

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/datanode/BlockSender.java

@@ -170,8 +170,13 @@ class BlockSender implements java.io.Closeable {
    * See {{@link BlockSender#isLongRead()}
    */
   private static final long LONG_READ_THRESHOLD_BYTES = 256 * 1024;
-  
 
+  // The number of bytes per checksum here determines the alignment
+  // of reads: we always start reading at a checksum chunk boundary,
+  // even if the checksum type is NULL. So, choosing too big of a value
+  // would risk sending too much unnecessary data. 512 (1 disk sector)
+  // is likely to result in minimal extra IO.
+  private static final long CHUNK_SIZE = 512;
   /**
    * Constructor
    * 
@@ -241,12 +246,6 @@ class BlockSender implements java.io.Closeable {
       synchronized(datanode.data) { 
         replica = getReplica(block, datanode);
         replicaVisibleLength = replica.getVisibleLength();
-        if (replica instanceof FinalizedReplica) {
-          // Load last checksum in case the replica is being written
-          // concurrently
-          final FinalizedReplica frep = (FinalizedReplica) replica;
-          chunkChecksum = frep.getLastChecksumAndDataLen();
-        }
       }
       // if there is a write in progress
       if (replica instanceof ReplicaBeingWritten) {
@@ -254,6 +253,10 @@ class BlockSender implements java.io.Closeable {
         waitForMinLength(rbw, startOffset + length);
         chunkChecksum = rbw.getLastChecksumAndDataLen();
       }
+      if (replica instanceof FinalizedReplica) {
+        chunkChecksum = getPartialChunkChecksumForFinalized(
+            (FinalizedReplica)replica);
+      }
 
       if (replica.getGenerationStamp() < block.getGenerationStamp()) {
         throw new IOException("Replica gen stamp < block genstamp, block="
@@ -329,12 +332,8 @@ class BlockSender implements java.io.Closeable {
         }
       }
       if (csum == null) {
-        // The number of bytes per checksum here determines the alignment
-        // of reads: we always start reading at a checksum chunk boundary,
-        // even if the checksum type is NULL. So, choosing too big of a value
-        // would risk sending too much unnecessary data. 512 (1 disk sector)
-        // is likely to result in minimal extra IO.
-        csum = DataChecksum.newDataChecksum(DataChecksum.Type.NULL, 512);
+        csum = DataChecksum.newDataChecksum(DataChecksum.Type.NULL,
+            (int)CHUNK_SIZE);
       }
 
       /*
@@ -410,6 +409,37 @@ class BlockSender implements java.io.Closeable {
     }
   }
 
+  private ChunkChecksum getPartialChunkChecksumForFinalized(
+      FinalizedReplica finalized) throws IOException {
+    // There are a number of places in the code base where a finalized replica
+    // object is created. If last partial checksum is loaded whenever a
+    // finalized replica is created, it would increase latency in DataNode
+    // initialization. Therefore, the last partial chunk checksum is loaded
+    // lazily.
+
+    // Load last checksum in case the replica is being written concurrently
+    final long replicaVisibleLength = replica.getVisibleLength();
+    if (replicaVisibleLength % CHUNK_SIZE != 0 &&
+        finalized.getLastPartialChunkChecksum() == null) {
+      // the finalized replica does not have precomputed last partial
+      // chunk checksum. Recompute now.
+      try {
+        finalized.loadLastPartialChunkChecksum();
+        return new ChunkChecksum(finalized.getVisibleLength(),
+            finalized.getLastPartialChunkChecksum());
+      } catch (FileNotFoundException e) {
+        // meta file is lost. Continue anyway to preserve existing behavior.
+        DataNode.LOG.warn(
+            "meta file " + finalized.getMetaFile() + " is missing!");
+        return null;
+      }
+    } else {
+      // If the checksum is null, BlockSender will use on-disk checksum.
+      return new ChunkChecksum(finalized.getVisibleLength(),
+          finalized.getLastPartialChunkChecksum());
+    }
+  }
+
   /**
    * close opened files.
    */

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/datanode/FinalizedReplica.java

@@ -18,7 +18,6 @@
 package org.apache.hadoop.hdfs.server.datanode;
 
 import java.io.File;
-import java.io.FileNotFoundException;
 import java.io.IOException;
 
 import org.apache.hadoop.hdfs.protocol.Block;
@@ -30,6 +29,7 @@ import org.apache.hadoop.hdfs.server.datanode.fsdataset.FsVolumeSpi;
  */
 public class FinalizedReplica extends ReplicaInfo {
   private boolean unlinked;      // copy-on-write done for block
+  private byte[] lastPartialChunkChecksum;
 
   /**
    * Constructor
@@ -41,9 +41,24 @@ public class FinalizedReplica extends ReplicaInfo {
    */
   public FinalizedReplica(long blockId, long len, long genStamp,
       FsVolumeSpi vol, File dir) {
+    this(blockId, len, genStamp, vol, dir, null);
+  }
+
+  /**
+   * Constructor.
+   * @param blockId block id
+   * @param len replica length
+   * @param genStamp replica generation stamp
+   * @param vol volume where replica is located
+   * @param dir directory path where block and meta files are located
+   * @param checksum the last partial chunk checksum
+   */
+  public FinalizedReplica(long blockId, long len, long genStamp,
+      FsVolumeSpi vol, File dir, byte[] checksum) {
     super(blockId, len, genStamp, vol, dir);
+    this.setLastPartialChunkChecksum(checksum);
   }
-  
+
   /**
    * Constructor
    * @param block a block
@@ -51,7 +66,20 @@ public class FinalizedReplica extends ReplicaInfo {
    * @param dir directory path where block and meta files are located
    */
   public FinalizedReplica(Block block, FsVolumeSpi vol, File dir) {
+    this(block, vol, dir, null);
+  }
+
+  /**
+   * Constructor.
+   * @param block a block
+   * @param vol volume where replica is located
+   * @param dir directory path where block and meta files are located
+   * @param checksum the last partial chunk checksum
+   */
+  public FinalizedReplica(Block block, FsVolumeSpi vol, File dir,
+      byte[] checksum) {
     super(block, vol, dir);
+    this.setLastPartialChunkChecksum(checksum);
   }
 
   /**
@@ -61,6 +89,7 @@ public class FinalizedReplica extends ReplicaInfo {
   public FinalizedReplica(FinalizedReplica from) {
     super(from);
     this.unlinked = from.isUnlinked();
+    this.setLastPartialChunkChecksum(from.getLastPartialChunkChecksum());
   }
 
   @Override  // ReplicaInfo
@@ -104,30 +133,18 @@ public class FinalizedReplica extends ReplicaInfo {
         + "\n  unlinked          =" + unlinked;
   }
 
-  /**
-   * gets the last chunk checksum and the length of the block corresponding
-   * to that checksum.
-   * Note, need to be called with the FsDataset lock acquired. May improve to
-   * lock only the FsVolume in the future.
-   * @throws IOException
-   */
-  public ChunkChecksum getLastChecksumAndDataLen() throws IOException {
-    ChunkChecksum chunkChecksum = null;
-    try {
-      byte[] lastChecksum = getVolume().loadLastPartialChunkChecksum(
-          getBlockFile(), getMetaFile());
-      if (lastChecksum != null) {
-        chunkChecksum =
-            new ChunkChecksum(getVisibleLength(), lastChecksum);
-      }
-    } catch (FileNotFoundException e) {
-      // meta file is lost. Try to continue anyway.
-      DataNode.LOG.warn("meta file " + getMetaFile() +
-          " is missing!");
-    } catch (IOException ioe) {
-      DataNode.LOG.warn("Unable to read checksum from meta file " +
-          getMetaFile(), ioe);
-    }
-    return chunkChecksum;
+  public byte[] getLastPartialChunkChecksum() {
+    return lastPartialChunkChecksum;
+  }
+
+  public void setLastPartialChunkChecksum(byte[] checksum) {
+    lastPartialChunkChecksum = checksum;
+  }
+
+  public void loadLastPartialChunkChecksum()
+      throws IOException {
+    byte[] lastChecksum = getVolume().loadLastPartialChunkChecksum(
+        getBlockFile(), getMetaFile());
+    setLastPartialChunkChecksum(lastChecksum);
   }
 }

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/datanode/fsdataset/impl/FsDatasetImpl.java

@@ -1127,10 +1127,8 @@ class FsDatasetImpl implements FsDatasetSpi<FsVolumeImpl> {
         v, newBlkFile.getParentFile(), Thread.currentThread(), bytesReserved);
 
     // load last checksum and datalen
-    byte[] lastChunkChecksum = v.loadLastPartialChunkChecksum(
-        replicaInfo.getBlockFile(), replicaInfo.getMetaFile());
     newReplicaInfo.setLastChecksumAndDataLen(
-        replicaInfo.getNumBytes(), lastChunkChecksum);
+        replicaInfo.getNumBytes(), replicaInfo.getLastPartialChunkChecksum());
 
     File newmeta = newReplicaInfo.getMetaFile();
 
@@ -1621,6 +1619,7 @@ class FsDatasetImpl implements FsDatasetSpi<FsVolumeImpl> {
          ReplicaState.FINALIZED) {
       newReplicaInfo = (FinalizedReplica)
              ((ReplicaUnderRecovery)replicaInfo).getOriginalReplica();
+      newReplicaInfo.loadLastPartialChunkChecksum();
     } else {
       FsVolumeImpl v = (FsVolumeImpl)replicaInfo.getVolume();
       File f = replicaInfo.getBlockFile();
@@ -1632,6 +1631,19 @@ class FsDatasetImpl implements FsDatasetSpi<FsVolumeImpl> {
       File dest = v.addFinalizedBlock(
           bpid, replicaInfo, f, replicaInfo.getBytesReserved());
       newReplicaInfo = new FinalizedReplica(replicaInfo, v, dest.getParentFile());
+
+      byte[] checksum = null;
+      // copy the last partial checksum if the replica is originally
+      // in finalized or rbw state.
+      if (replicaInfo.getState() == ReplicaState.FINALIZED) {
+        FinalizedReplica finalized = (FinalizedReplica)replicaInfo;
+        checksum = finalized.getLastPartialChunkChecksum();
+      } else if (replicaInfo.getState() == ReplicaState.RBW) {
+        ReplicaBeingWritten rbw = (ReplicaBeingWritten)replicaInfo;
+        checksum = rbw.getLastChecksumAndDataLen().getChecksum();
+      }
+      newReplicaInfo.setLastPartialChunkChecksum(checksum);
+
       if (v.isTransientStorage()) {
         ramDiskReplicaTracker.addReplica(bpid, replicaInfo.getBlockId(), v);
         datanode.getMetrics().addRamDiskBytesWrite(replicaInfo.getNumBytes());

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/MiniDFSCluster.java

@@ -2637,6 +2637,29 @@ public class MiniDFSCluster {
         blk.getBlockPoolId()), blk.getBlockId()), blk.getBlockName());
   }
 
+  /**
+   * Return all block files in given directory (recursive search).
+   */
+  public static List<File> getAllBlockFiles(File storageDir) {
+    List<File> results = new ArrayList<File>();
+    File[] files = storageDir.listFiles();
+    if (files == null) {
+      return null;
+    }
+    for (File f : files) {
+      if (f.getName().startsWith(Block.BLOCK_FILE_PREFIX) &&
+          !f.getName().endsWith(Block.METADATA_EXTENSION)) {
+        results.add(f);
+      } else if (f.isDirectory()) {
+        List<File> subdirResults = getAllBlockFiles(f);
+        if (subdirResults != null) {
+          results.addAll(subdirResults);
+        }
+      }
+    }
+    return results;
+  }
+
   /**
    * Get the latest metadata file correpsonding to a block
    * @param storageDir storage directory

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/namenode/TestListCorruptFileBlocks.java

@@ -89,7 +89,7 @@ public class TestListCorruptFileBlocks {
       File storageDir = cluster.getInstanceStorageDir(0, 1);
       File data_dir = MiniDFSCluster.getFinalizedDir(storageDir, bpid);
       assertTrue("data directory does not exist", data_dir.exists());
-      List<File> metaFiles = MiniDFSCluster.getAllBlockMetadataFiles(data_dir);
+      List<File> metaFiles = MiniDFSCluster.getAllBlockFiles(data_dir);
       assertTrue("Data directory does not contain any blocks or there was an "
           + "IO error", metaFiles != null && !metaFiles.isEmpty());
       File metaFile = metaFiles.get(0);
@@ -169,7 +169,7 @@ public class TestListCorruptFileBlocks {
       File data_dir = MiniDFSCluster.getFinalizedDir(storageDir, 
           cluster.getNamesystem().getBlockPoolId());
       assertTrue("data directory does not exist", data_dir.exists());
-      List<File> metaFiles = MiniDFSCluster.getAllBlockMetadataFiles(data_dir);
+      List<File> metaFiles = MiniDFSCluster.getAllBlockFiles(data_dir);
       assertTrue("Data directory does not contain any blocks or there was an "
           + "IO error", metaFiles != null && !metaFiles.isEmpty());
       File metaFile = metaFiles.get(0);