[jira] [Commented] (SPARK-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15001316#comment-15001316 ] Chris Howard commented on SPARK-11154: -- Hi [~tgraves] - this overlaps to a point with [SPARK-3374|https://issues.apache.org/jira/browse/SPARK-3374] and [SPARK-4408|https://issues.apache.org/jira/browse/SPARK-4408] I would agree with [~srowen] that 2.0 provides an opportunity for house keeping / consolidation and there is scope to clean up the config / args for cluster and client modes. I would prefer not to create new configs and would rather stick with the current naming and support k | m | g modifiers unless somebody has a strong view on what to rename the existing configs. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15000441#comment-15000441 ] Sean Owen commented on SPARK-11154: --- I think we'd have to make new properties to maintain compatibility. However I agree it's confusing. I think it's therefore not worth fixing in 1.x. At best, target this for 2.x. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15000438#comment-15000438 ] Thomas Graves commented on SPARK-11154: --- sorry actually thinking about this more I'm torn. New configs might be less confusing then changing it in 2.0 or for new users hitting issue [~sowen] pointed out. But I just hate to see yet more configs andI don't like the change of memory to mem in spark.yarn.am.memory because now they are different from the other spark ones spark.executor.memory, etc. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15000426#comment-15000426 ] Thomas Graves commented on SPARK-11154: --- It seems unnecessary to me to add new configs just to support this. I see this causing confusion to users. Personally I would rather see either adding the support for k/m/g and leaving if none specified defaults to m (although this doesn't match other spark things) or there is discussion going on about spark 2.0 and perhaps there we just change the existing configs to support k/m/g, etc, > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15000419#comment-15000419 ] Dustin Cote commented on SPARK-11154: - [~Kitard] I think the naming convention and strategy makes sense. Someone with more familiarity with the code base should probably comment on the files that need to change though. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=1544#comment-1544 ] Chris Howard commented on SPARK-11154: -- On review it looks like the following conf values need changes - spark.yarn.driver.memoryOverhead - spark.yarn.am.memoryOverhead - spark.yarn.am.memory - spark.yarn.executor.memoryOverhead Changes required to - yarn/ClientArguments.scala - yarn/Client.scala - yarn/YarnAllocator.scala org.apache.spark.deploy.ClientArguments.scala uses org.apache.spark.util.MemoryParam to convert to MB so propose same. If we were to replicate the functionality of deploy.ClientArguments.scala we should also consider allowing setting via command line args. Propose we add new config values to ensure backward compatibility. Anyone issues with naing as follows? - spark.yarn.driver.memoryOverhead -> spark.yarn.driver.memOverhead - spark.yarn.am.memoryOverhead -> spark.yarn.am.memOverhead - spark.yarn.am.memory -> spark.yarn.am.mem - spark.yarn.executor.memoryOverhead -> spark.yarn.executor.memOverhead Docs will also need updating to reflect fact that original option is MB only and new options support k | m | g variants. Happy to run with this if no objections. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14973322#comment-14973322 ] Sean Owen commented on SPARK-11154: --- I think that if this is done at all, it would have to be with a new property. The old one would then be deprecated but continue to function. This would have to be done for all such properties. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14961195#comment-14961195 ] Dustin Cote commented on SPARK-11154: - That's absolutely a good point. It might be more confusing to have a default 'untagged' number be specified in megabytes and allow option appended 'g' for gigabytes. I think it would be safe to allow people to specify the 'm' at the end (this currently doesn't work), but I agree it would be a bad thing to move from a default of megabytes to bytes as it wouldn't be backward compatible. This might just be a usability quirk, but I figured I'd call it out since I see customers trying to use this like Xmx all the time. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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-11154) make specificaition spark.yarn.executor.memoryOverhead consistent with typical JVM options
[ https://issues.apache.org/jira/browse/SPARK-11154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14961184#comment-14961184 ] Sean Owen commented on SPARK-11154: --- Should be for all similar properties, not just this one. The twist is that you have to support the current syntax. 1000 must mean "1000 megabytes". But then someone writing "100" would be surprised to find that it means "100 megabytes". (CM might do just this, note.) Hence I'm actually not sure if this is feasible. > make specificaition spark.yarn.executor.memoryOverhead consistent with > typical JVM options > -- > > Key: SPARK-11154 > URL: https://issues.apache.org/jira/browse/SPARK-11154 > Project: Spark > Issue Type: Improvement > Components: Documentation, Spark Submit >Reporter: Dustin Cote >Priority: Minor > > spark.yarn.executor.memoryOverhead is currently specified in megabytes by > default, but it would be nice to allow users to specify the size as though it > were a typical -Xmx option to a JVM where you can have 'm' and 'g' appended > to the end to explicitly specify megabytes or gigabytes. -- 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