yucai commented on issue #19788: [SPARK-9853][Core] Optimize shuffle fetch of contiguous partition IDs URL: https://github.com/apache/spark/pull/19788#issuecomment-456871550 @tgravescs thanks for looking at this! The shuffle block is stored like below: format is s"shuffle_$shuffleId_$mapId_$reduceId", refering to BlockId.scala. <img src="https://user-images.githubusercontent.com/2989575/51620952-7ae1e900-1f6e-11e9-873b-7e51b04968c5.png"; height="200" width="1000"> Before this PR, to read map output file 0's data (e.g.: reducer 5 to 10), whose block Ids are from shuffle_0_0_5 to shuffle_0_0_10, Spark needs 6 disk IOs + 6 network IOs. After this PR, to read the same map output file 0's data, we only needs 1 disk IO and 1 network IO, this way can reduce IO dramatically. We did this kind of merge in both client and server side. In my previous benchmark testing, merge IO will have very obvious improvment in shuffle read when adaptive execution enabled.
---------------------------------------------------------------- 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 --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
![https://user-images.githubusercontent.com/2989575/51620952-7ae1e900-1f6e-11e9-873b-7e51b04968c5.png"](https://user-images.githubusercontent.com/2989575/51620952-7ae1e900-1f6e-11e9-873b-7e51b04968c5.png"){kind=link}