siknezevic commented on pull request #27246: URL: https://github.com/apache/spark/pull/27246#issuecomment-646872842
> > Spilling SpillReader with 16000 rows: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative > > UnsafeSorterSpillReader_ratio1 568 727 200 0.5 2219.4 1.0X > > UnsafeSorterSpillReader_ratio1024 122902 123262 509 0.0 480085.1 0.0X > > Thanks for the benchmarking! Could you run the benchmark code on both the master with/without this patch? Then, please describe performance differences in the PR description. @maropu The below are results of the following micro-benchmarks: 1) Case when read buffer size is parameterized: OpenJDK 64-Bit Server VM 1.8.0_252-8u252-b09-1~16.04-b09 on Linux 4.4.0-178-generic Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz Spilling SpillReader with 16000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative //------------------------------------------------------------------------------------------------ UnsafeSorterSpillReader_bufferSize1024 1312 / 1448 0.2 5125.1 1.0X UnsafeSorterSpillReader_bufferSize1024*1024 124134 / 125262 0.0 484897.4 0.0X 2) Case when read buffer size is hard-codded to 1024 (1KB) Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz Spilling SpillReader with 16000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative //------------------------------------------------------------------------------------------------ UnsafeSorterSpillReader_bufferSize1024 228 / 351 1.1 892.1 1.0X 3) Case when read buffer size is hard-coded to 1024*1024 (1MB) OpenJDK 64-Bit Server VM 1.8.0_252-8u252-b09-1~16.04-b09 on Linux 4.4.0-178-generic Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz Spilling SpillReader with 16000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative //------------------------------------------------------------------------------------------------ UnsafeSorterSpillReader_bufferSize1024*1024 122350 / 123832 0.0 477930.4 1.0X If you compare the micro-benchmark results of parameterized execution 1) with hard-coded results 2) of 1024 buffer size you can noticed performance hit (around 5X). Created micro-benchmark simulates 100000 joins (creates 100000 iterators). You will notice just 2 sec performance hit for buffer 1024*1024.This is understandable because 1MB buffer size already has big performance impact. With 10TB TPCDS data set I tested spilling with query q14a and buffer size of 1024. Execution with hard-coded read buffer size is faster by 37% (27 min vs 37 min) comparing to the execution when buffer size is parameterized and the same size 1024 is used. Query q14a, for 10TB data set, generates around 180 million joins per partition and when buffer size is parameterized, that translates into 10 min longer execution time. Could you please let me know would it be OK to hard-code the read buffer size to 1024? Taking in the consideration that UnsafeSorterSpillReader already has logic to adjust buffer size to actual data size when needed. Thank you very much for your help ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org
