[ https://issues.apache.org/jira/browse/SPARK-15531?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15300986#comment-15300986 ]
mathieu longtin commented on SPARK-15531: ----------------------------------------- Correct, on a 128G server, just running {{java}} with no argument will try to allocate 32G, regardless of ulimit. It's "expected behavior" according to Oracle. > spark-class tries to use too much memory when running Launcher > -------------------------------------------------------------- > > Key: SPARK-15531 > URL: https://issues.apache.org/jira/browse/SPARK-15531 > Project: Spark > Issue Type: Bug > Components: Deploy > Affects Versions: 1.6.1, 2.0.0 > Environment: Linux running in Univa or Sun Grid Engine > Reporter: mathieu longtin > Priority: Minor > Labels: launcher > > When running Java on a server with a lot of memory but a rather small virtual > memory ulimit, Java will try to allocate a large memory pool and fail: > {code} > # System has 128GB of Ram but ulimit set to 7.5G > $ ulimit -v > 7812500 > $ java -client > Error occurred during initialization of VM > Could not reserve enough space for object heap > Error: Could not create the Java Virtual Machine. > Error: A fatal exception has occurred. Program will exit. > {code} > This is a known issue with Java, but unlikely to get fixed. > As a result, when starting various Spark process (spark-submit, master or > workers), they fail when {{spark-class}} tries to run > {{org.apache.spark.launcher.Main}}. > To fix this, add {{-Xmx128m}} to this line > {code} > "$RUNNER" -Xmx128m -cp "$LAUNCH_CLASSPATH" org.apache.spark.launcher.Main > "$\@" > {code} > (https://github.com/apache/spark/blob/master/bin/spark-class#L71) > We've been using 128m and that works in our setup. Considering all the > launcher does is analyze the arguments and env var and spit out some command, > it should be plenty. All other calls to Java seem to include some value for > -Xmx, so it is not an issue elsewhere. > I don't mind submitting a PR, but I'm sure somebody has opinions on the 128m > (bigger, smaller, configurable, ...), so I'd rather it would be discussed > first. -- 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