ilooner commented on a change in pull request #1606: DRILL-6845: Semi-Hash-Join
to skip incoming build duplicates, automatically stop skipping if too few
URL: https://github.com/apache/drill/pull/1606#discussion_r248161996
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/physical/impl/join/HashJoinSpillControlImpl.java
##########
@@ -18,39 +18,211 @@
package org.apache.drill.exec.physical.impl.join;
+ import org.apache.drill.exec.ExecConstants;
import org.apache.drill.exec.memory.BufferAllocator;
+ import org.apache.drill.exec.ops.FragmentContext;
+ import org.apache.drill.exec.record.RecordBatch;
import org.apache.drill.exec.record.RecordBatchMemoryManager;
- import org.apache.drill.exec.record.RecordBatchSizer;
- import org.apache.drill.exec.record.VectorContainer;
+ import org.slf4j.Logger;
+ import org.slf4j.LoggerFactory;
+
+ import javax.annotation.Nullable;
+ import java.util.Set;
+
+ import static org.apache.drill.exec.record.JoinBatchMemoryManager.LEFT_INDEX;
+ import static
org.apache.drill.exec.record.JoinBatchMemoryManager.RIGHT_INDEX;
/**
* This class is currently used only for Semi-Hash-Join that avoids duplicates
by the use of a hash table
- * The method {@link
HashJoinMemoryCalculator.HashJoinSpillControl#shouldSpill(VectorContainer)}
returns true if the memory available now to the allocator if not enough
+ * The method {@link
HashJoinMemoryCalculator.HashJoinSpillControl#shouldSpill()} returns true if
the memory available now to the allocator if not enough
* to hold (a multiple of, for safety) a new allocated batch
*/
-public class HashJoinSpillControlImpl implements
HashJoinMemoryCalculator.HashJoinSpillControl {
+public class HashJoinSpillControlImpl implements
HashJoinMemoryCalculator.BuildSidePartitioning {
+ private static final Logger logger =
LoggerFactory.getLogger(HashJoinSpillControlImpl.class);
+
private BufferAllocator allocator;
private int recordsPerBatch;
private int minBatchesInAvailableMemory;
private RecordBatchMemoryManager batchMemoryManager;
+ private int initialPartitions;
+ private int numPartitions;
+ private int recordsPerPartitionBatchProbe;
+ private FragmentContext context;
- HashJoinSpillControlImpl(BufferAllocator allocator, int recordsPerBatch, int
minBatchesInAvailableMemory, RecordBatchMemoryManager batchMemoryManager) {
+ HashJoinSpillControlImpl(BufferAllocator allocator, int recordsPerBatch, int
minBatchesInAvailableMemory, RecordBatchMemoryManager batchMemoryManager,
FragmentContext context) {
this.allocator = allocator;
this.recordsPerBatch = recordsPerBatch;
this.minBatchesInAvailableMemory = minBatchesInAvailableMemory;
this.batchMemoryManager = batchMemoryManager;
+ this.context = context;
}
@Override
- public boolean shouldSpill(VectorContainer currentVectorContainer) {
- assert currentVectorContainer.hasRecordCount();
- assert currentVectorContainer.getRecordCount() == recordsPerBatch;
- // Expected new batch size like the current, plus the Hash Value vector (4
bytes per HV)
- long batchSize = new
RecordBatchSizer(currentVectorContainer).getActualSize() + 4 * recordsPerBatch;
+ public boolean shouldSpill() {
+ // Expected new batch size like the current, plus the Hash Values vector
(4 bytes per HV)
+ long batchSize = (
batchMemoryManager.getRecordBatchSizer(RIGHT_INDEX).getRowAllocWidth() + 4 ) *
recordsPerBatch;
long reserveForOutgoing = batchMemoryManager.getOutputBatchSize();
long memoryAvailableNow = allocator.getLimit() -
allocator.getAllocatedMemory() - reserveForOutgoing;
boolean needsSpill = minBatchesInAvailableMemory * batchSize >
memoryAvailableNow;
+ if ( needsSpill ) {
+ logger.debug("should spill now - batch size {}, mem avail {}, reserved
for outgoing {}", batchSize, memoryAvailableNow, reserveForOutgoing);
+ }
return needsSpill; // go spill if too little memory is available
}
+ @Override
+ public void initialize(boolean firstCycle,
+ boolean reserveHash,
+ RecordBatch buildSideBatch,
+ RecordBatch probeSideBatch,
+ Set<String> joinColumns,
+ boolean probeEmpty,
+ long memoryAvailable,
+ int initialPartitions,
+ int recordsPerPartitionBatchBuild,
+ int recordsPerPartitionBatchProbe,
+ int maxBatchNumRecordsBuild,
+ int maxBatchNumRecordsProbe,
+ int outputBatchSize,
+ double loadFactor) {
+ this.initialPartitions = initialPartitions;
+ this.recordsPerPartitionBatchProbe = recordsPerPartitionBatchProbe;
+
+ calculateMemoryUsage();
+ }
+
+ @Override
+ public void setPartitionStatSet(HashJoinMemoryCalculator.PartitionStatSet
partitionStatSet) {
+ // Do nothing
+ }
+
+ @Override
+ public int getNumPartitions() {
+ return numPartitions;
+ }
+
+ /**
+ * Calculate the number of partitions possible for the given available memory
+ * start at initialPartitions and adjust down (in powers of 2) as needed
+ */
+ private void calculateMemoryUsage() {
+ long reserveForOutgoing = batchMemoryManager.getOutputBatchSize();
+ long memoryAvailableNow = allocator.getLimit() -
allocator.getAllocatedMemory() - reserveForOutgoing;
+
+ // Expected new batch size like the current, plus the Hash Values vector
(4 bytes per HV)
+ long batchSize = (
batchMemoryManager.getRecordBatchSizer(RIGHT_INDEX).getRowAllocWidth() + 4 ) *
recordsPerBatch;
+ long hashTableSize = batchSize /* the keys in the HT */ +
+ 4 * context.getOptions().getLong(ExecConstants.MIN_HASH_TABLE_SIZE_KEY)
/* the initial hash table buckets */ +
+ (2 + 2) * recordsPerBatch; /* the hash-values and the links */
+
+ for ( numPartitions = initialPartitions; numPartitions > 2; numPartitions
/= 2 ) { // need at least 2
+ // each partition needs at least one internal build batch, and a minimum
hash-table
+ // adding "min batches" to create some safety slack
+ if ( memoryAvailableNow >
+ ( numPartitions + minBatchesInAvailableMemory ) * (batchSize +
hashTableSize) ) {
+ break; // got enough memory
+ }
+ }
+ logger.debug("Spill control chosen to use {} partitions", numPartitions);
+ }
+
+ @Override
+ public long getBuildReservedMemory() {
+ return 0;
+ }
+
+ @Override
+ public long getMaxReservedMemory() {
+ return 0;
+ }
+
+ @Override
+ public String makeDebugString() {
+ return "Spill Control " +
HashJoinMemoryCalculatorImpl.HashJoinSpillControl.class.getCanonicalName();
+ }
+
+ @Nullable
+ @Override
+ public HashJoinMemoryCalculator.PostBuildCalculations next() {
+ return new
SpillControlPostBuildCalculationsImpl(recordsPerPartitionBatchProbe,allocator,recordsPerBatch,minBatchesInAvailableMemory,batchMemoryManager);
+ }
+
+ @Override
+ public HashJoinState getState() {
+ return HashJoinState.BUILD_SIDE_PARTITIONING;
+ }
+
+
+ /**
+ * The purpose of this class is to provide the method {@link #shouldSpill}
that ensures that enough memory is available to
+ * hold all the probe incoming batches for those partitions that spilled
(else need to spill more of them, for more memory).
+ */
+ public static class SpillControlPostBuildCalculationsImpl implements
HashJoinMemoryCalculator.PostBuildCalculations {
+ private static final Logger logger =
LoggerFactory.getLogger(SpillControlPostBuildCalculationsImpl.class);
+
+ private final int recordsPerPartitionBatchProbe;
+ private BufferAllocator allocator;
+ private int recordsPerBatch;
+ private int minBatchesInAvailableMemory;
+ private RecordBatchMemoryManager batchMemoryManager;
+ private int numPartitionsSpilled;
+ private boolean probeEmpty;
+
+ public SpillControlPostBuildCalculationsImpl(final int
recordsPerPartitionBatchProbe,
+ BufferAllocator allocator,
int recordsPerBatch, int minBatchesInAvailableMemory, RecordBatchMemoryManager
batchMemoryManager) {
+ this.allocator = allocator;
+ this.recordsPerBatch = recordsPerBatch;
+ this.minBatchesInAvailableMemory = minBatchesInAvailableMemory;
+ this.batchMemoryManager = batchMemoryManager;
+ this.recordsPerPartitionBatchProbe = recordsPerPartitionBatchProbe;
+ }
+
+ @Override
+ public void initialize(boolean hasProbeData, int numPartitionsSpilled) {
+ this.probeEmpty = hasProbeData;
+ this.numPartitionsSpilled = numPartitionsSpilled;
+ }
+
+ @Override
+ public int getProbeRecordsPerBatch() {
+ return recordsPerPartitionBatchProbe;
+ }
+
+ @Override
+ public boolean shouldSpill() {
+ if ( numPartitionsSpilled == 0 ) { return false; } // no extra memory is
needed if all the build side is in memory
Review comment:
- "Indeed the incoming batch read from the spill is not specifically
accounted for. However there's a "slack" (+ min batches in memory) that can
cover that small gap." We should not be introducing another tuning parameter
with minBatchesInMemory, we already have the safetyFactor and
fragmentationFactor tuning parameters. How about we do something like the
following psuedo code:
```
if (firstCycle) {
partitionsMemoryUsage = allocator.getAllocatedMemory() ;
probeReserveSpace = 0;
} else {
partitionsMemoryUsage = allocator.getAllocatedMemory() -
probeBatch.getSize();
probeReserveSpace =
probeSizePredictor.predictBatchSize(maxNumProbeRecords, false)
}
```
This would allow us to account for the probe data safely while using the
existing safetyFactor and fragementationFactor parameters and would allow us to
remove minBatchesInMemory.
- PartitionStatSet collects all the stats from the HashPartitions directly.
So if a HashPartition is spilled the PartitionStatSet will know. I've unit
tested this and the other memory calculators use this logic without issue. The
idea was to decouple the stats needed for memory calculations from the rest of
the HashJoin operator, so that the memory calculators can be unit tested.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
