You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@hive.apache.org by ha...@apache.org on 2020/07/23 04:05:55 UTC
[hive] 01/02: HIVE-23843: Improve key evictions in
VectorGroupByOperator (Rajesh Balamohan via Ashutosh Chauhan,
Zoltan Haindrich)
This is an automated email from the ASF dual-hosted git repository.
hashutosh pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/hive.git
commit 380be9afd1a364fe0ff83e61e17ba4ced12f29a0
Author: Rajesh Balamohan <rb...@apache.org>
AuthorDate: Tue Jul 14 10:00:14 2020 +0530
HIVE-23843: Improve key evictions in VectorGroupByOperator (Rajesh Balamohan via Ashutosh Chauhan, Zoltan Haindrich)
Signed-off-by: Ashutosh Chauhan <ha...@apache.org>
---
.../ql/exec/vector/VectorAggregationBufferRow.java | 12 +-
.../hive/ql/exec/vector/VectorGroupByOperator.java | 53 +++++++--
.../ql/exec/vector/TestVectorGroupByOperator.java | 125 ++++++++++++++++++---
3 files changed, 168 insertions(+), 22 deletions(-)
diff --git a/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorAggregationBufferRow.java b/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorAggregationBufferRow.java
index 494db35..a7ef154 100644
--- a/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorAggregationBufferRow.java
+++ b/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorAggregationBufferRow.java
@@ -28,7 +28,8 @@ public class VectorAggregationBufferRow {
private VectorAggregateExpression.AggregationBuffer[] aggregationBuffers;
private int version;
private int index;
-
+ private int accessed = 0;
+
public VectorAggregationBufferRow(
VectorAggregateExpression.AggregationBuffer[] aggregationBuffers) {
this.aggregationBuffers = aggregationBuffers;
@@ -80,5 +81,12 @@ public class VectorAggregationBufferRow {
aggregationBuffers[i].reset();
}
}
-
+
+ public int getAccessCount() {
+ return accessed;
+ }
+
+ public void incrementAccessCount() {
+ accessed++;
+ }
}
diff --git a/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByOperator.java b/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByOperator.java
index 9f81e8e..85535f5 100644
--- a/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByOperator.java
+++ b/ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByOperator.java
@@ -151,6 +151,10 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
private float memoryThreshold;
private boolean isLlap = false;
+
+ // tracks overall access count in map agg buffer any given time.
+ private long totalAccessCount;
+
/**
* Interface for processing mode: global, hash, unsorted streaming, or group batch
*/
@@ -251,7 +255,7 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
* This mode is very simple, there are no keys to consider, and only flushes one row at closing
* The one row must flush even if no input was seen (NULLs)
*/
- private class ProcessingModeGlobalAggregate extends ProcessingModeBase {
+ final class ProcessingModeGlobalAggregate extends ProcessingModeBase {
/**
* In global processing mode there is only one set of aggregation buffers
@@ -288,12 +292,13 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
/**
* Hash Aggregate mode processing
*/
- private class ProcessingModeHashAggregate extends ProcessingModeBase {
+ final class ProcessingModeHashAggregate extends ProcessingModeBase {
/**
* The global key-aggregation hash map.
*/
- private Map<KeyWrapper, VectorAggregationBufferRow> mapKeysAggregationBuffers;
+ @VisibleForTesting
+ Map<KeyWrapper, VectorAggregationBufferRow> mapKeysAggregationBuffers;
/**
* Total per hashtable entry fixed memory (does not depend on key/agg values).
@@ -334,7 +339,8 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
/**
* A soft reference used to detect memory pressure
*/
- private SoftReference<Object> gcCanary = new SoftReference<Object>(new Object());
+ @VisibleForTesting
+ SoftReference<Object> gcCanary = new SoftReference<Object>(new Object());
/**
* Counts the number of time the gcCanary died and was resurrected
@@ -387,10 +393,19 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
sumBatchSize = 0;
mapKeysAggregationBuffers = new HashMap<KeyWrapper, VectorAggregationBufferRow>();
+ if (groupingSets != null && groupingSets.length > 0) {
+ this.maxHtEntries = this.maxHtEntries / groupingSets.length;
+ LOG.info("New maxHtEntries: {}, groupingSets len: {}", maxHtEntries, groupingSets.length);
+ }
computeMemoryLimits();
LOG.debug("using hash aggregation processing mode");
}
+ @VisibleForTesting
+ int getMaxHtEntries() {
+ return maxHtEntries;
+ }
+
@Override
public void doProcessBatch(VectorizedRowBatch batch, boolean isFirstGroupingSet,
boolean[] currentGroupingSetsOverrideIsNulls) throws HiveException {
@@ -502,6 +517,10 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
mapKeysAggregationBuffers.put(kw.copyKey(), aggregationBuffer);
numEntriesHashTable++;
numEntriesSinceCheck++;
+ } else {
+ // for access tracking
+ aggregationBuffer.incrementAccessCount();
+ totalAccessCount++;
}
aggregationBatchInfo.mapAggregationBufferSet(aggregationBuffer, i);
}
@@ -540,6 +559,16 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
}
}
+ int computeAvgAccess() {
+ if (numEntriesHashTable == 0) {
+ return 0;
+ }
+ int avgAccess = (int) (totalAccessCount / numEntriesHashTable);
+ LOG.debug("totalAccessCount:{}, numEntries:{}, avgAccess:{}",
+ totalAccessCount, numEntriesHashTable, avgAccess);
+ return avgAccess;
+ }
+
/**
* Flushes the entries in the hash table by emiting output (forward).
* When parameter 'all' is true all the entries are flushed.
@@ -561,6 +590,7 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
maxHashTblMemory/1024/1024,
gcCanary.get() == null ? "dead" : "alive"));
}
+ int avgAccess = computeAvgAccess();
/* Iterate the global (keywrapper,aggregationbuffers) map and emit
a row for each key */
@@ -568,10 +598,17 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
mapKeysAggregationBuffers.entrySet().iterator();
while(iter.hasNext()) {
Map.Entry<KeyWrapper, VectorAggregationBufferRow> pair = iter.next();
+ if (!all && avgAccess >= 1) {
+ // Retain entries when access pattern is > than average access
+ if (pair.getValue().getAccessCount() > avgAccess) {
+ continue;
+ }
+ }
writeSingleRow((VectorHashKeyWrapperBase) pair.getKey(), pair.getValue());
if (!all) {
+ totalAccessCount -= pair.getValue().getAccessCount();
iter.remove();
--numEntriesHashTable;
if (++entriesFlushed >= entriesToFlush) {
@@ -582,6 +619,7 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
if (all) {
mapKeysAggregationBuffers.clear();
+ totalAccessCount = 0;
numEntriesHashTable = 0;
}
@@ -674,7 +712,7 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
* Streaming processing mode on ALREADY GROUPED data. Each input VectorizedRowBatch may
* have a mix of different keys. Intermediate values are flushed each time key changes.
*/
- private class ProcessingModeStreaming extends ProcessingModeBase {
+ final class ProcessingModeStreaming extends ProcessingModeBase {
/**
* The aggregation buffers used in streaming mode
@@ -816,7 +854,7 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
* writeGroupRow does this and finally increments outputBatch.size.
*
*/
- private class ProcessingModeReduceMergePartial extends ProcessingModeBase {
+ final class ProcessingModeReduceMergePartial extends ProcessingModeBase {
private boolean first;
private boolean isLastGroupBatch;
@@ -896,7 +934,8 @@ public class VectorGroupByOperator extends Operator<GroupByDesc>
/**
* Current processing mode. Processing mode can change (eg. hash -> streaming).
*/
- private transient IProcessingMode processingMode;
+ @VisibleForTesting
+ transient IProcessingMode processingMode;
private static final long serialVersionUID = 1L;
diff --git a/ql/src/test/org/apache/hadoop/hive/ql/exec/vector/TestVectorGroupByOperator.java b/ql/src/test/org/apache/hadoop/hive/ql/exec/vector/TestVectorGroupByOperator.java
index 3835987..c22a833 100644
--- a/ql/src/test/org/apache/hadoop/hive/ql/exec/vector/TestVectorGroupByOperator.java
+++ b/ql/src/test/org/apache/hadoop/hive/ql/exec/vector/TestVectorGroupByOperator.java
@@ -19,7 +19,6 @@
package org.apache.hadoop.hive.ql.exec.vector;
import static org.junit.Assert.assertEquals;
-import static org.junit.Assert.assertFalse;
import static org.junit.Assert.assertNotNull;
import static org.junit.Assert.assertTrue;
@@ -39,9 +38,9 @@ import java.util.Set;
import org.apache.calcite.util.Pair;
import org.apache.hadoop.hive.common.type.HiveDecimal;
import org.apache.hadoop.hive.conf.HiveConf;
-import org.apache.hadoop.hive.llap.LlapDaemonInfo;
import org.apache.hadoop.hive.llap.io.api.LlapProxy;
import org.apache.hadoop.hive.ql.CompilationOpContext;
+import org.apache.hadoop.hive.ql.exec.KeyWrapper;
import org.apache.hadoop.hive.ql.exec.Operator;
import org.apache.hadoop.hive.ql.exec.OperatorFactory;
import org.apache.hadoop.hive.ql.exec.vector.expressions.aggregates.VectorUDAFCountStar;
@@ -76,8 +75,6 @@ import org.apache.hadoop.hive.serde2.io.DoubleWritable;
import org.apache.hadoop.hive.serde2.io.HiveDecimalWritable;
import org.apache.hadoop.hive.serde2.io.ShortWritable;
import org.apache.hadoop.hive.serde2.io.TimestampWritableV2;
-import org.apache.hadoop.hive.serde2.objectinspector.PrimitiveObjectInspector.PrimitiveCategory;
-import org.apache.hadoop.hive.serde2.typeinfo.PrimitiveTypeInfo;
import org.apache.hadoop.hive.serde2.typeinfo.TypeInfo;
import org.apache.hadoop.hive.serde2.typeinfo.TypeInfoFactory;
import org.apache.hadoop.io.BooleanWritable;
@@ -499,6 +496,31 @@ public class TestVectorGroupByOperator {
}
});
+ // vrb of 1 row each
+ FakeVectorRowBatchFromObjectIterables data = getDataForRollup();
+
+ long countRowsProduced = 0;
+ for (VectorizedRowBatch unit: data) {
+ // after 24 rows, we'd have seen all the keys
+ // find 14 keys in the hashmap
+ // but 24*0.5 = 12
+ // won't turn off hash mode because of the 3 grouping sets
+ // if it turns off the hash mode, we'd get 14 + 3*(100-24) rows
+ countRowsProduced += unit.size;
+ vgo.process(unit, 0);
+
+ if (countRowsProduced >= 100) {
+ break;
+ }
+
+ }
+ vgo.close(false);
+ // all groupings
+ // 10 keys generates 14 rows with the rollup
+ assertEquals(1+3+10, outputRowCount);
+ }
+
+ FakeVectorRowBatchFromObjectIterables getDataForRollup() throws HiveException {
// k1 has nDV of 2
Iterable<Object> k1 = new Iterable<Object>() {
@Override
@@ -578,33 +600,110 @@ public class TestVectorGroupByOperator {
};
// vrb of 1 row each
- FakeVectorRowBatchFromObjectIterables data = new FakeVectorRowBatchFromObjectIterables(
+ return new FakeVectorRowBatchFromObjectIterables(
2,
new String[] {"long", "long", "long", "long"},
k1,
k2,
v,
v); // output col
+ }
+
+ @Test
+ public void testRollupAggregationWithFlush() throws HiveException {
+
+ List<String> mapColumnNames = new ArrayList<String>();
+ mapColumnNames.add("k1");
+ mapColumnNames.add("k2");
+ mapColumnNames.add("v");
+ VectorizationContext ctx = new VectorizationContext("name", mapColumnNames);
+
+ // select count(v) from name group by rollup (k1,k2);
+
+ Pair<GroupByDesc,VectorGroupByDesc> pair = buildKeyGroupByDesc (ctx, "count",
+ "v", TypeInfoFactory.longTypeInfo,
+ new String[] { "k1", "k2" },
+ new TypeInfo[] {TypeInfoFactory.longTypeInfo, TypeInfoFactory.longTypeInfo});
+ GroupByDesc desc = pair.left;
+ VectorGroupByDesc vectorDesc = pair.right;
+
+ desc.setGroupingSetsPresent(true);
+ ArrayList<Long> groupingSets = new ArrayList<>();
+ // groupingSets
+ groupingSets.add(0L);
+ groupingSets.add(1L);
+ groupingSets.add(2L);
+ desc.setListGroupingSets(groupingSets);
+ // add grouping sets dummy key
+ ExprNodeDesc groupingSetDummyKey = new ExprNodeConstantDesc(TypeInfoFactory.longTypeInfo, 0L);
+ // this only works because we used an arraylist in buildKeyGroupByDesc
+ // don't do this in actual compiler
+ desc.getKeys().add(groupingSetDummyKey);
+ // groupingSet Position
+ desc.setGroupingSetPosition(2);
+
+ CompilationOpContext cCtx = new CompilationOpContext();
+
+ desc.setMinReductionHashAggr(0.5f);
+ // Set really low check interval setting
+ hconf.set("hive.groupby.mapaggr.checkinterval", "10");
+ hconf.set("hive.vectorized.groupby.checkinterval", "10");
+
+ Operator<? extends OperatorDesc> groupByOp = OperatorFactory.get(cCtx, desc);
+
+ VectorGroupByOperator vgo =
+ (VectorGroupByOperator) Vectorizer.vectorizeGroupByOperator(groupByOp, ctx, vectorDesc);
+
+ FakeCaptureVectorToRowOutputOperator out = FakeCaptureVectorToRowOutputOperator.addCaptureOutputChild(cCtx, vgo);
+ vgo.initialize(hconf, null);
+
+ //Get the processing mode
+ VectorGroupByOperator.ProcessingModeHashAggregate processingMode =
+ (VectorGroupByOperator.ProcessingModeHashAggregate) vgo.processingMode;
+ assertEquals(333333,
+ ((VectorGroupByOperator.ProcessingModeHashAggregate)vgo.processingMode).getMaxHtEntries());
+
+ this.outputRowCount = 0;
+ out.setOutputInspector(new FakeCaptureVectorToRowOutputOperator.OutputInspector() {
+ @Override
+ public void inspectRow(Object row, int tag) throws HiveException {
+ ++outputRowCount;
+ }
+ });
+
+ FakeVectorRowBatchFromObjectIterables data = getDataForRollup();
long countRowsProduced = 0;
+ long numElementsToBeRetained = 0;
+ int avgAccess = 0;
for (VectorizedRowBatch unit: data) {
- // after 24 rows, we'd have seen all the keys
- // find 14 keys in the hashmap
- // but 24*0.5 = 12
- // won't turn off hash mode because of the 3 grouping sets
- // if it turns off the hash mode, we'd get 14 + 3*(100-24) rows
countRowsProduced += unit.size;
vgo.process(unit, 0);
if (countRowsProduced >= 100) {
+ // note down avg access
+ avgAccess = processingMode.computeAvgAccess();
+ numElementsToBeRetained = getElementsHigherThan(processingMode.mapKeysAggregationBuffers, avgAccess);
+ // trigger flush explicitly on next iteration
+ processingMode.gcCanary.clear();
break;
}
+ }
+
+ // This processing would trigger flush
+ for (VectorizedRowBatch unit: data) {
+ vgo.process(unit, 0);
+ long freqElementsAfterFlush = getElementsHigherThan(processingMode.mapKeysAggregationBuffers, avgAccess);
+ assertTrue("After flush: " + freqElementsAfterFlush + ", before flush: " + numElementsToBeRetained,
+ (freqElementsAfterFlush >= numElementsToBeRetained));
+ break;
}
vgo.close(false);
- // all groupings
- // 10 keys generates 14 rows with the rollup
- assertEquals(1+3+10, outputRowCount);
+ }
+
+ long getElementsHigherThan(Map<KeyWrapper, VectorAggregationBufferRow> aggMap, int avgAccess) {
+ return aggMap.values().stream().filter(v -> (v.getAccessCount() > avgAccess)).count();
}
@Test