[jira] [Commented] (SPARK-11005) Spark 1.5 Shuffle performance - (sort-based shuffle)
[ https://issues.apache.org/jira/browse/SPARK-11005?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14949059#comment-14949059 ] Sean Owen commented on SPARK-11005: --- [~vnayak053] coud I ask you to put this on the mailing list? It's not quite a JIRA as you're asking a question. There's a similar thread at this very moment about why jobs may not scale linearly: https://www.mail-archive.com/user@spark.apache.org/msg38382.html > Spark 1.5 Shuffle performance - (sort-based shuffle) > > > Key: SPARK-11005 > URL: https://issues.apache.org/jira/browse/SPARK-11005 > Project: Spark > Issue Type: Question > Components: Shuffle, SQL >Affects Versions: 1.5.0 > Environment: 6 node cluster with 1 master and 5 worker nodes. > Memory > 100 GB each > Cores = 72 each > Input data ~94 GB >Reporter: Sandeep Pal > > In case of terasort by Spark SQL with 20 total cores(4 cores/ executor), the > performance of the map tasks is 14 minutes (around 26s-30s each) where as if > I increase the number of cores to 60(12 cores /executor), the performance of > map degrades to 30 minutes ( ~2.3 minutes per task). I believe the map tasks > are independent of each other in the shuffle. > Each map task has 128 MB input (HDFS block size) in both of the above cases. > So, what makes the performance degradation with increasing number of cores. > -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-11005) Spark 1.5 Shuffle performance - (sort-based shuffle)
[ https://issues.apache.org/jira/browse/SPARK-11005?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14949117#comment-14949117 ] Sandeep Pal commented on SPARK-11005: - I think, this thread is about the different problem. Anyway, I will put it in the mailing list. Thanks > Spark 1.5 Shuffle performance - (sort-based shuffle) > > > Key: SPARK-11005 > URL: https://issues.apache.org/jira/browse/SPARK-11005 > Project: Spark > Issue Type: Question > Components: Shuffle, SQL >Affects Versions: 1.5.0 > Environment: 6 node cluster with 1 master and 5 worker nodes. > Memory > 100 GB each > Cores = 72 each > Input data ~94 GB >Reporter: Sandeep Pal > > In case of terasort by Spark SQL with 20 total cores(4 cores/ executor), the > performance of the map tasks is 14 minutes (around 26s-30s each) where as if > I increase the number of cores to 60(12 cores /executor), the performance of > map degrades to 30 minutes ( ~2.3 minutes per task). I believe the map tasks > are independent of each other in the shuffle. > Each map task has 128 MB input (HDFS block size) in both of the above cases. > So, what makes the performance degradation with increasing number of cores. > -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-11005) Spark 1.5 Shuffle performance - (sort-based shuffle)
[ https://issues.apache.org/jira/browse/SPARK-11005?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14949305#comment-14949305 ] Sandeep Pal commented on SPARK-11005: - Ok, Thanks > Spark 1.5 Shuffle performance - (sort-based shuffle) > > > Key: SPARK-11005 > URL: https://issues.apache.org/jira/browse/SPARK-11005 > Project: Spark > Issue Type: Question > Components: Shuffle, SQL >Affects Versions: 1.5.0 > Environment: 6 node cluster with 1 master and 5 worker nodes. > Memory > 100 GB each > Cores = 72 each > Input data ~94 GB >Reporter: Sandeep Pal > > In case of terasort by Spark SQL with 20 total cores(4 cores/ executor), the > performance of the map tasks is 14 minutes (around 26s-30s each) where as if > I increase the number of cores to 60(12 cores /executor), the performance of > map degrades to 30 minutes ( ~2.3 minutes per task). I believe the map tasks > are independent of each other in the shuffle. > Each map task has 128 MB input (HDFS block size) in both of the above cases. > So, what makes the performance degradation with increasing number of cores. > -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org