vanzin commented on issue #25299: [SPARK-27651][Core] Avoid the network when shuffle blocks are fetched from the same host URL: https://github.com/apache/spark/pull/25299#issuecomment-558285542 > Now we are having to wait to fetch the local blocks first before the host local blocks where as before that could have been in parallel with the local blocks. "Host local" blocks are basically local blocks after you fill in the local dirs cache. And "fetch" is a misnomer in this case; you'll get a pointer to the data (which mostly involves metadata operations like finding the file, but not actually opening it), and it will only be actually opened when the task starts running. So not really any waiting involved. Sounds like a gain to me - while before you'd have RPC + open on the remote side + transfer data over socket + cache it in memory in executor (or, worst, write it to disk again + open another file), now you only do an "open file".
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
