hadoop-common-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From szets...@apache.org
Subject hadoop git commit: HDFS-7980. Incremental BlockReport will dramatically slow down namenode startup. Contributed by Walter Su
Date Thu, 07 May 2015 18:39:09 GMT
Repository: hadoop
Updated Branches:
  refs/heads/branch-2 bb035ff08 -> 4e1f2eb39


HDFS-7980. Incremental BlockReport will dramatically slow down namenode startup.  Contributed
by Walter Su


Project: http://git-wip-us.apache.org/repos/asf/hadoop/repo
Commit: http://git-wip-us.apache.org/repos/asf/hadoop/commit/4e1f2eb3
Tree: http://git-wip-us.apache.org/repos/asf/hadoop/tree/4e1f2eb3
Diff: http://git-wip-us.apache.org/repos/asf/hadoop/diff/4e1f2eb3

Branch: refs/heads/branch-2
Commit: 4e1f2eb3955a97a70cf127dc97ae49201a90f5e0
Parents: bb035ff
Author: Tsz-Wo Nicholas Sze <szetszwo@hortonworks.com>
Authored: Thu May 7 11:36:35 2015 -0700
Committer: Tsz-Wo Nicholas Sze <szetszwo@hortonworks.com>
Committed: Thu May 7 11:37:33 2015 -0700

----------------------------------------------------------------------
 hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt     |   3 +
 .../server/blockmanagement/BlockManager.java    |   8 +-
 .../blockmanagement/TestBlockManager.java       | 108 +++++++++++++++++++
 3 files changed, 115 insertions(+), 4 deletions(-)
----------------------------------------------------------------------


http://git-wip-us.apache.org/repos/asf/hadoop/blob/4e1f2eb3/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
----------------------------------------------------------------------
diff --git a/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt b/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
index 5b358d5..7fd0472 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
+++ b/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
@@ -386,6 +386,9 @@ Release 2.7.1 - UNRELEASED
     HDFS-8305: HDFS INotify: the destination field of RenameOp should always
     end with the file name (cmccabe)
 
+    HDFS-7980. Incremental BlockReport will dramatically slow down namenode
+    startup.  (Walter Su via szetszwo)
+
 Release 2.7.0 - 2015-04-20
 
   INCOMPATIBLE CHANGES

http://git-wip-us.apache.org/repos/asf/hadoop/blob/4e1f2eb3/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
----------------------------------------------------------------------
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
index 842faac..1b45b0b 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
+++ b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
@@ -1812,7 +1812,7 @@ public class BlockManager {
         return !node.hasStaleStorages();
       }
 
-      if (storageInfo.numBlocks() == 0) {
+      if (storageInfo.getBlockReportCount() == 0) {
         // The first block report can be processed a lot more efficiently than
         // ordinary block reports.  This shortens restart times.
         processFirstBlockReport(storageInfo, newReport);
@@ -2066,7 +2066,7 @@ public class BlockManager {
       final BlockListAsLongs report) throws IOException {
     if (report == null) return;
     assert (namesystem.hasWriteLock());
-    assert (storageInfo.numBlocks() == 0);
+    assert (storageInfo.getBlockReportCount() == 0);
 
     for (BlockReportReplica iblk : report) {
       ReplicaState reportedState = iblk.getState();
@@ -2478,14 +2478,14 @@ public class BlockManager {
     }
 
     // just add it
-    storageInfo.addBlock(storedBlock);
+    AddBlockResult result = storageInfo.addBlock(storedBlock);
 
     // Now check for completion of blocks and safe block count
     int numCurrentReplica = countLiveNodes(storedBlock);
     if (storedBlock.getBlockUCState() == BlockUCState.COMMITTED
         && numCurrentReplica >= minReplication) {
       completeBlock(storedBlock.getBlockCollection(), storedBlock, false);
-    } else if (storedBlock.isComplete()) {
+    } else if (storedBlock.isComplete() && result == AddBlockResult.ADDED) {
       // check whether safe replication is reached for the block
       // only complete blocks are counted towards that.
       // In the case that the block just became complete above, completeBlock()

http://git-wip-us.apache.org/repos/asf/hadoop/blob/4e1f2eb3/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
----------------------------------------------------------------------
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
index 9ce16f2..1e09e19 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
+++ b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
@@ -47,10 +47,14 @@ import org.apache.hadoop.hdfs.protocol.BlockListAsLongs;
 import org.apache.hadoop.hdfs.server.blockmanagement.DatanodeDescriptor.BlockTargetPair;
 import org.apache.hadoop.hdfs.server.common.HdfsServerConstants;
 import org.apache.hadoop.hdfs.server.datanode.DataNodeTestUtils;
+import org.apache.hadoop.hdfs.server.datanode.FinalizedReplica;
+import org.apache.hadoop.hdfs.server.datanode.ReplicaBeingWritten;
 import org.apache.hadoop.hdfs.server.namenode.FSNamesystem;
 import org.apache.hadoop.hdfs.server.namenode.NameNodeAdapter;
 import org.apache.hadoop.hdfs.server.protocol.DatanodeRegistration;
 import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage;
+import org.apache.hadoop.hdfs.server.protocol.ReceivedDeletedBlockInfo;
+import org.apache.hadoop.hdfs.server.protocol.StorageReceivedDeletedBlocks;
 import org.apache.hadoop.ipc.RemoteException;
 import org.apache.hadoop.net.NetworkTopology;
 import org.junit.Assert;
@@ -646,6 +650,110 @@ public class TestBlockManager {
         BlockListAsLongs.EMPTY, null, false);
     assertEquals(1, ds.getBlockReportCount());
   }
+
+  /**
+   * test when NN starts and in same mode, it receives an incremental blockReport
+   * firstly. Then receives first full block report.
+   */
+  @Test
+  public void testSafeModeIBRBeforeFirstFullBR() throws Exception {
+    // pretend to be in safemode
+    doReturn(true).when(fsn).isInStartupSafeMode();
+
+    DatanodeDescriptor node = nodes.get(0);
+    DatanodeStorageInfo ds = node.getStorageInfos()[0];
+    node.isAlive = true;
+    DatanodeRegistration nodeReg =  new DatanodeRegistration(node, null, null, "");
+
+    // register new node
+    bm.getDatanodeManager().registerDatanode(nodeReg);
+    bm.getDatanodeManager().addDatanode(node);
+    assertEquals(node, bm.getDatanodeManager().getDatanode(node));
+    assertEquals(0, ds.getBlockReportCount());
+    // Build a incremental report
+    List<ReceivedDeletedBlockInfo> rdbiList = new ArrayList<>();
+    // Build a full report
+    BlockListAsLongs.Builder builder = BlockListAsLongs.builder();
+
+    // blk_42 is finalized.
+    long receivedBlockId = 42;  // arbitrary
+    BlockInfoContiguous receivedBlock = addBlockToBM(receivedBlockId);
+    rdbiList.add(new ReceivedDeletedBlockInfo(new Block(receivedBlock),
+        ReceivedDeletedBlockInfo.BlockStatus.RECEIVED_BLOCK, null));
+    builder.add(new FinalizedReplica(receivedBlock, null, null));
+
+    // blk_43 is under construction.
+    long receivingBlockId = 43;
+    BlockInfoContiguous receivingBlock = addUcBlockToBM(receivingBlockId);
+    rdbiList.add(new ReceivedDeletedBlockInfo(new Block(receivingBlock),
+        ReceivedDeletedBlockInfo.BlockStatus.RECEIVING_BLOCK, null));
+    builder.add(new ReplicaBeingWritten(receivingBlock, null, null, null));
+
+    // blk_44 has 2 records in IBR. It's finalized. So full BR has 1 record.
+    long receivingReceivedBlockId = 44;
+    BlockInfoContiguous receivingReceivedBlock = addBlockToBM(receivingReceivedBlockId);
+    rdbiList.add(new ReceivedDeletedBlockInfo(new Block(receivingReceivedBlock),
+        ReceivedDeletedBlockInfo.BlockStatus.RECEIVING_BLOCK, null));
+    rdbiList.add(new ReceivedDeletedBlockInfo(new Block(receivingReceivedBlock),
+        ReceivedDeletedBlockInfo.BlockStatus.RECEIVED_BLOCK, null));
+    builder.add(new FinalizedReplica(receivingReceivedBlock, null, null));
+
+    // blk_45 is not in full BR, because it's deleted.
+    long ReceivedDeletedBlockId = 45;
+    rdbiList.add(new ReceivedDeletedBlockInfo(
+        new Block(ReceivedDeletedBlockId),
+        ReceivedDeletedBlockInfo.BlockStatus.RECEIVED_BLOCK, null));
+    rdbiList.add(new ReceivedDeletedBlockInfo(
+        new Block(ReceivedDeletedBlockId),
+        ReceivedDeletedBlockInfo.BlockStatus.DELETED_BLOCK, null));
+
+    // blk_46 exists in DN for a long time, so it's in full BR, but not in IBR.
+    long existedBlockId = 46;
+    BlockInfoContiguous existedBlock = addBlockToBM(existedBlockId);
+    builder.add(new FinalizedReplica(existedBlock, null, null));
+
+    // process IBR and full BR
+    StorageReceivedDeletedBlocks srdb =
+        new StorageReceivedDeletedBlocks(new DatanodeStorage(ds.getStorageID()),
+            rdbiList.toArray(new ReceivedDeletedBlockInfo[rdbiList.size()]));
+    bm.processIncrementalBlockReport(node, srdb);
+    // Make sure it's the first full report
+    assertEquals(0, ds.getBlockReportCount());
+    bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
+        builder.build(), null, false);
+    assertEquals(1, ds.getBlockReportCount());
+
+    // verify the storage info is correct
+    assertTrue(bm.getStoredBlock(new Block(receivedBlockId)).findStorageInfo
+        (ds) >= 0);
+    assertTrue(((BlockInfoContiguousUnderConstruction) bm.
+        getStoredBlock(new Block(receivingBlockId))).getNumExpectedLocations() > 0);
+    assertTrue(bm.getStoredBlock(new Block(receivingReceivedBlockId))
+        .findStorageInfo(ds) >= 0);
+    assertNull(bm.getStoredBlock(new Block(ReceivedDeletedBlockId)));
+    assertTrue(bm.getStoredBlock(new Block(existedBlock)).findStorageInfo
+        (ds) >= 0);
+  }
+
+  private BlockInfoContiguous addBlockToBM(long blkId) {
+    Block block = new Block(blkId);
+    BlockInfoContiguous blockInfo =
+        new BlockInfoContiguous(block, (short) 3);
+    BlockCollection bc = Mockito.mock(BlockCollection.class);
+    Mockito.doReturn((short) 3).when(bc).getBlockReplication();
+    bm.blocksMap.addBlockCollection(blockInfo, bc);
+    return blockInfo;
+  }
+
+  private BlockInfoContiguous addUcBlockToBM(long blkId) {
+    Block block = new Block(blkId);
+    BlockInfoContiguousUnderConstruction blockInfo =
+        new BlockInfoContiguousUnderConstruction(block, (short) 3);
+    BlockCollection bc = Mockito.mock(BlockCollection.class);
+    Mockito.doReturn((short) 3).when(bc).getBlockReplication();
+    bm.blocksMap.addBlockCollection(blockInfo, bc);
+    return blockInfo;
+  }
   
   /**
    * Tests that a namenode doesn't choose a datanode with full disks to 


Mime
View raw message