Hi, Flink's HashJoin implementation was designed to gracefully handle inputs that exceed the main memory. It is not explicitly optimized for in-memory processing and does not play fancy tricks like optimizing cache accesses or batching. I assume your benchmark is about in-memory joins only. This was not the main design goal when the join was implemented but robustness. Since most of the development of Flink focuses on streaming applications at the moment, the join implementation has barely been touched in recent years (except for minor extensions and bugfixes).
Regarding your tests, Tuple should give better performance than Row because Row is null-sensitive and serialized a null-mask. There is also a blog post about Flink's join performance [1] which is already a bit dusty but as I said, the algorithm hasn't change much since then. Best, Fabian [1] https://flink.apache.org/news/2015/03/13/peeking-into-Apache-Flinks-Engine-Room.html 2017-05-15 16:26 GMT+02:00 weijie tong <tongweijie...@gmail.com>: > The Flink version is 1.2.0 > > On Mon, May 15, 2017 at 10:24 PM, weijie tong <tongweijie...@gmail.com> > wrote: > >> @Till thanks for your reply. >> >> My code is similar to HashTableITCase.testInMemoryMutableHashTable() >> . It just use the MutableHashTable class , there's no other Flink's >> configuration. The main code body is: >> >> this.recordBuildSideAccessor = RecordSerializer.get(); >>> this.recordProbeSideAccessor = RecordSerializer.get(); >>> final int[] buildKeyPos = new int[]{buildSideJoinIndex}; >>> final int[] probeKeyPos = new int[]{probeSideJoinIndex}; >>> final Class<? extends Value>[] keyType = (Class<? extends Value>[]) new >>> Class[]{BytesValue.class}; >>> this.recordBuildSideComparator = new RecordComparator(buildKeyPos, keyType); >>> this.recordProbeSideComparator = new RecordComparator(probeKeyPos, keyType); >>> this.pactRecordComparator = new >>> HashJoinVectorJointGroupIterator.RecordPairComparator(buildSideJoinIndex); >>> Sequence<Record> buildSideRecordsSeq = >>> makeSequenceRecordOfSameSideSegments(buildSideSegs, localJoinQuery); >>> Sequence<Record> probeSideRecordsSeq = >>> makeSequenceRecordOfSameSideSegments(probeSideSegs, localJoinQuery); >>> List<MemorySegment> memorySegments; >>> int pageSize = hashTableMemoryManager.getTotalNumPages(); >>> try { >>> memorySegments = this.hashTableMemoryManager.allocatePages(MEM_OWNER, >>> pageSize); >>> } >>> catch (MemoryAllocationException e) { >>> LOGGER.error("could not allocate " + pageSize + " pages memory for >>> HashJoin", e); >>> Throwables.propagate(e); >>> return; >>> } >>> try { >>> Stopwatch stopwatch = Stopwatch.createStarted(); >>> UniformRecordGenerator buildInput = new >>> UniformRecordGenerator(buildSideRecordsSeq); >>> UniformRecordGenerator probeInput = new >>> UniformRecordGenerator(probeSideRecordsSeq); >>> join = new MutableHashTable<Record, Record>( >>> recordBuildSideAccessor, >>> recordProbeSideAccessor, >>> recordBuildSideComparator, >>> recordProbeSideComparator, >>> pactRecordComparator, >>> memorySegments, >>> ioManager >>> ); >>> join.open(buildInput,probeInput); >>> >>> LOGGER.info("construct hash table elapsed:" + >>> stopwatch.elapsed(TimeUnit.MILLISECONDS) + "ms"); >>> >>> >> The BytesValue type is self defined one which holds byte[] , but just >> like the original StringValue, also has the same serDe performance. >> >> >> while (join.nextRecord()) { >> Record currentProbeRecord = join.getCurrentProbeRecord(); >> MutableObjectIterator<Record> buildSideIterator = >> join.getBuildSideIterator(); >> while (buildSideIterator.next(reusedBuildSideRow) != null) { >> materializeRecord2OutVector(reusedBuildSideRow, buildSideIndex2Value, >> buildSideIndex2Vector, rowNum); >> materializeRecord2OutVector(currentProbeRecord, probeSideIndex2Value, >> probeSideIndex2Vector, rowNum); >> rowNum++; >> }} >> >> >> >> >> I have tried both the Record ,Row class as the type of records without >> any better improved performance . I also tried batched the input records. >> That means the buildInput or probeInput variables of the first code >> block which iterate one Record a time from another batched Records . >> Batched records's content stay in memory in Drill's ValueVector format. >> Once a record is need to participate in the build or probe phase from a >> iterate.next() call, >> it will be fetched from the batched in memory ValueVector content. But no >> performance gains. >> >> >> The top hotspot profile from Jprofiler is below: >> > >> Hot spot,"Self time (microseconds)","Average Time","Invocations" >> org.apache.flink.types.Record.serialize,1014127,"n/a","n/a" >> org.apache.flink.types.Record.deserialize,60684,"n/a","n/a" >> org.apache.flink.types.Record.copyTo,83007,"n/a","n/a" >> org.apache.flink.runtime.operators.hash.MutableHashTable. >> open,55238,"n/a","n/a" >> org.apache.flink.runtime.operators.hash.MutableHashTable. >> nextRecord,10955,"n/a","n/a" >> org.apache.flink.runtime.memory.MemoryManager.release,33484,"n/a","n/a" >> org.apache.flink.runtime.memory.MemoryManager.allocatePages, >> 104259,"n/a","n/a" >> >> >> My log show that hashjoin.open() method costs too much time. >> > >> construct hash table elapsed:1885ms >> >> >> >> >> On Mon, May 15, 2017 at 6:20 PM, Till Rohrmann <trohrm...@apache.org> >> wrote: >> >>> Hi Weijie, >>> >>> it might be the case that batching the processing of multiple rows can >>> give you an improved performance compared to single row processing. >>> >>> Maybe you could share the exact benchmark base line results and the code >>> you use to test Flink's MutableHashTable with us. Also the Flink >>> configuration and how you run it would be of interest. That way we might be >>> able to see if we can tune Flink a bit more. >>> >>> Cheers, >>> Till >>> >>> On Sun, May 14, 2017 at 5:23 AM, weijie tong <tongweijie...@gmail.com> >>> wrote: >>> >>>> I has a test case to use Flink's MutableHashTable class to do a hash >>>> join on a local machine with 64g memory, 64cores. The test case is one >>>> build table with 14w rows ,one probe table with 320w rows ,the matched >>>> result rows is 12 w. >>>> >>>> It takes 2.2 seconds to complete the join.The performance seems bad. I >>>> ensure there's no overflow, the smaller table is the build side. The >>>> MutableObjectIterator is a sequence of Rows. The Row is composed of several >>>> fields which are byte[]. Through my log,I find the open() method takes >>>> 1.560 seconds. The probe iterates phase takes 680ms. And my Jprofiler's >>>> profile shows the MutableObjectIterator's next() method call is the >>>> hotspot. >>>> >>>> >>>> I want to know how to tune this scenario. I find Drill's HashJoin is >>>> batch model. Its build side's input is a RecordBatch which holds batch of >>>> rows and memory size is approach to L2 cache. Through this strategy it will >>>> gain less method calls (that means call to next() ) and much efficient to >>>> cpu calculation. I also find SQL server's paper noticed the batch model's >>>> performance gains (https://www.microsoft.com/en- >>>> us/research/wp-content/uploads/2013/06/Apollo3-Sigmod-2013-final.pdf) >>>> . I guess the performance's down is due to the single row iterate model. >>>> >>>> >>>> Hope someone to correct my opinion. Also maybe I have a wrong use of >>>> the MutableHashTable. wait for someone to give an advice. >>>> >>> >>> >> >
