TPCDS queries on Gluten+Velox in EMR is considerably slower than OSS Spark

[sagarlakshmipathy]

Backend

VL (Velox)

Bug description

[Expected behavior] Faster query runs compared to OSS Spark [actual behavior] OSS Spark runs in half the time taken by Gluten+Velox Spark.

Spark version

None

Spark configurations

Gluten+Velox+Spark

./spark-3.4.1-bin-hadoop3/bin/spark-shell --master yarn --deploy-mode client --driver-memory 19g --executor-memory 19g --executor-cores 5 --num-executors 32 --jars /home/hadoop/hudi-spark3.4-bundle_2.12-0.14.1.jar,/home/hadoop/hudi-benchmarks-0.1-SNAPSHOT.jar --packages org.apache.hadoop:hadoop-aws:3.2.4 --conf spark.plugins=io.glutenproject.GlutenPlugin --conf spark.memory.offHeap.enabled=true --conf spark.memory.offHeap.size=30g --conf spark.shuffler=org.apache.spark.shuffle.sort.ColumnarShuffleManager --conf spark.serializer=org.apache.spark.serializer.KryoSerializer --conf spark.sql.extensions=org.apache.spark.sql.hudi.HoodieSparkSessionExtension --conf spark.sql.catalog.spark_catalog=org.apache.spark.sql.hudi.catalog.HoodieCatalog --conf spark.hadoop.fs.s3a.aws.credentials.provider=com.amazonaws.auth.DefaultAWSCredentialsProviderChain --conf spark.sql.catalogImplementation=in-memory --conf spark.ui.proxyBase="" --conf 'spark.eventLog.enabled=true' --conf 'spark.eventLog.dir=hdfs:///var/log/spark/apps'

OSS Spark

./spark-3.4.1-bin-hadoop3/bin/spark-shell --master yarn --deploy-mode client --driver-memory 19g --executor-memory 19g --executor-cornum-executors 32 --jars /home/hadoop/hudi-spark3.4-bundle_2.12-0.14.1.jar,/home/hadoop/hudi-benchmarks-0.1-SNAPSHOT.jar --packages org.apache.hadoop:hadoop-aws:3.2.4 --conf spark.serializer=org.apache.spark.serializer.KryoSerializer --conf spark.sql.extensions=org.apache.spark.sql.hudi.HoodieSparkSessionExtension --conf spark.sql.catalog.spark_catalog=org.apache.spark.sql.hudi.catalog.HoodieCatalog --conf spark.hadoop.fs.s3a.aws.credentials.provider=com.amazonaws.auth.DefaultAWSCredentialsProviderChain --conf spark.sql.catalogImplementation=in-memory --conf spark.ui.proxyBase="" --conf 'spark.eventLog.enabled=true' --conf 'spark.eventLog.dir=hdfs:///var/log/spark/apps'

System information

Environment: Amazon EMR - 10 workers, 1 driver all m5.4xlarge OS: Amazon Linux 2

Relevant logs

Wondering what you need me to capture that'll help you

[zhouyuan]

Hi @sagarlakshmipathy Can you please also share the performance number per query? on TPCDS the Q72 is still a trouble for gluten and needs some special config. Here's some discussions: https://github.com/apache/incubator-gluten/issues/1775

Are you testing with HUDI tables by any chance? --conf spark.sql.catalog.spark_catalog=org.apache.spark.sql.hudi.catalog.HoodieCatalog For now the HUDI support is not ready in Gluten. It will actually run with vanilla Spark code, and with a RowtoColumn(memcpy) connect to Gluten native operators. So this will actually bring lots of overhead.

thanks, -yuan

[sagarlakshmipathy]

Query ID	Gluten Velox Spark Hudi (ms)	OSS Spark Hudi
1	22040	16699
2	60531	33095
3	61031	25965
4	360561	172286
5	140865	72149
6	48038	22890
7	106637	44359
8	45072	19636

I didn't bother running the rest of them. I am testing Hudi tables with Gluten. Is there a gh issue/discussion I can +1 to?

[zhouyuan]

It is quite likely due to the fallback of scanning HUDI tables. Here's the issue tracker for unified data lake design, ICEBERG and DELTA LAKE are now both supported(not 100%) now. https://github.com/apache/incubator-gluten/issues/3378

Thanks, -yuan