From cae55de2cd1f9ea068f3410c8bdea14cf55738cb Mon Sep 17 00:00:00 2001 From: Sanford Ryza Date: Fri, 16 Aug 2013 08:11:04 +0000 Subject: [PATCH] MAPREDUCE-5462. In map-side sort, swap entire meta entries instead of indexes for better cache performance. (Sandy Ryza) git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1514608 13f79535-47bb-0310-9956-ffa450edef68 --- hadoop-mapreduce-project/CHANGES.txt | 3 ++ .../org/apache/hadoop/mapred/MapTask.java | 51 ++++++++----------- 2 files changed, 23 insertions(+), 31 deletions(-) diff --git a/hadoop-mapreduce-project/CHANGES.txt b/hadoop-mapreduce-project/CHANGES.txt index 613a38a52f..a412371966 100644 --- a/hadoop-mapreduce-project/CHANGES.txt +++ b/hadoop-mapreduce-project/CHANGES.txt @@ -189,6 +189,9 @@ Release 2.1.1-beta - UNRELEASED MAPREDUCE-5446. TestJobHistoryEvents and TestJobHistoryParsing have race conditions (jlowe via kihwal) + MAPREDUCE-5462. In map-side sort, swap entire meta entries instead of + indexes for better cache performance. (Sandy Ryza) + BUG FIXES MAPREDUCE-5385. Fixed a bug with JobContext getCacheFiles API. (Omkar Vinit diff --git a/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapred/MapTask.java b/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapred/MapTask.java index e9d3ed7886..99f9eac81f 100644 --- a/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapred/MapTask.java +++ b/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapred/MapTask.java @@ -884,10 +884,10 @@ public static class MapOutputBuffer byte[] kvbuffer; // main output buffer private final byte[] b0 = new byte[0]; - private static final int INDEX = 0; // index offset in acct - private static final int VALSTART = 1; // val offset in acct - private static final int KEYSTART = 2; // key offset in acct - private static final int PARTITION = 3; // partition offset in acct + private static final int VALSTART = 0; // val offset in acct + private static final int KEYSTART = 1; // key offset in acct + private static final int PARTITION = 2; // partition offset in acct + private static final int VALLEN = 3; // length of value private static final int NMETA = 4; // num meta ints private static final int METASIZE = NMETA * 4; // size in bytes @@ -1151,10 +1151,10 @@ public synchronized void collect(K key, V value, final int partition distanceTo(keystart, valend, bufvoid)); // write accounting info - kvmeta.put(kvindex + INDEX, kvindex); kvmeta.put(kvindex + PARTITION, partition); kvmeta.put(kvindex + KEYSTART, keystart); kvmeta.put(kvindex + VALSTART, valstart); + kvmeta.put(kvindex + VALLEN, distanceTo(valstart, valend)); // advance kvindex kvindex = (kvindex - NMETA + kvmeta.capacity()) % kvmeta.capacity(); } catch (MapBufferTooSmallException e) { @@ -1224,17 +1224,11 @@ int distanceTo(final int i, final int j, final int mod) { } /** - * For the given meta position, return the dereferenced position in the - * integer array. Each meta block contains several integers describing - * record data in its serialized form, but the INDEX is not necessarily - * related to the proximate metadata. The index value at the referenced int - * position is the start offset of the associated metadata block. So the - * metadata INDEX at metapos may point to the metadata described by the - * metadata block at metapos + k, which contains information about that - * serialized record. + * For the given meta position, return the offset into the int-sized + * kvmeta buffer. */ int offsetFor(int metapos) { - return kvmeta.get(metapos * NMETA + INDEX); + return metapos * NMETA; } /** @@ -1260,16 +1254,17 @@ public int compare(final int mi, final int mj) { kvmeta.get(kvj + VALSTART) - kvmeta.get(kvj + KEYSTART)); } + final byte META_BUFFER_TMP[] = new byte[METASIZE]; /** - * Swap logical indices st i, j MOD offset capacity. + * Swap metadata for items i, j * @see IndexedSortable#swap */ public void swap(final int mi, final int mj) { - final int kvi = (mi % maxRec) * NMETA + INDEX; - final int kvj = (mj % maxRec) * NMETA + INDEX; - int tmp = kvmeta.get(kvi); - kvmeta.put(kvi, kvmeta.get(kvj)); - kvmeta.put(kvj, tmp); + int iOff = (mi % maxRec) * METASIZE; + int jOff = (mj % maxRec) * METASIZE; + System.arraycopy(kvbuffer, iOff, META_BUFFER_TMP, 0, METASIZE); + System.arraycopy(kvbuffer, jOff, kvbuffer, iOff, METASIZE); + System.arraycopy(META_BUFFER_TMP, 0, kvbuffer, jOff, METASIZE); } /** @@ -1601,9 +1596,9 @@ private void sortAndSpill() throws IOException, ClassNotFoundException, while (spindex < mend && kvmeta.get(offsetFor(spindex % maxRec) + PARTITION) == i) { final int kvoff = offsetFor(spindex % maxRec); - key.reset(kvbuffer, kvmeta.get(kvoff + KEYSTART), - (kvmeta.get(kvoff + VALSTART) - - kvmeta.get(kvoff + KEYSTART))); + int keystart = kvmeta.get(kvoff + KEYSTART); + int valstart = kvmeta.get(kvoff + VALSTART); + key.reset(kvbuffer, keystart, valstart - keystart); getVBytesForOffset(kvoff, value); writer.append(key, value); ++spindex; @@ -1729,14 +1724,8 @@ private void spillSingleRecord(final K key, final V value, private void getVBytesForOffset(int kvoff, InMemValBytes vbytes) { // get the keystart for the next serialized value to be the end // of this value. If this is the last value in the buffer, use bufend - final int nextindex = kvoff == kvend - ? bufend - : kvmeta.get( - (kvoff - NMETA + kvmeta.capacity() + KEYSTART) % kvmeta.capacity()); - // calculate the length of the value - int vallen = (nextindex >= kvmeta.get(kvoff + VALSTART)) - ? nextindex - kvmeta.get(kvoff + VALSTART) - : (bufvoid - kvmeta.get(kvoff + VALSTART)) + nextindex; + final int vallen = kvmeta.get(kvoff + VALLEN); + assert vallen >= 0; vbytes.reset(kvbuffer, kvmeta.get(kvoff + VALSTART), vallen); }