I had this problem at my work. I solved by increasing the unix ulimit, because spark is trying to open to many files.
Em 29 de set de 2017 5:05 PM, "Anthony Thomas" <ahtho...@eng.ucsd.edu> escreveu: > Hi Spark Users, > > I recently compiled spark 2.2.0 from source on an EC2 m4.2xlarge instance > (8 cores, 32G RAM) running Ubuntu 14.04. I'm using Oracle Java 1.8. I > compiled using the command: > > export MAVEN_OPTS="-Xmx2g -XX:ReservedCodeCacheSize=512m" > ./build/mvn -DskipTests -Pnetlib-lgpl clean package > > Spark compiles fine, but when running tests (./build/mvn test), at least > one test in "StandaloneDynamicAllocationSuite" of "Spark Project Core" > consistently cause the JVM to crash with errors like: > > Exception in thread "ExecutorRunner for app-20170929185545-0000/0" > java.lang.OutOfMemoryError: unable to create new native thread > > According to "cat /proc/sys/kernel/threads-max" the system can support up > to 120,000 threads which seems like it should be more than enough. The > limits set in ulimit also seem reasonable. Looking at "top," the JVM > doesn't seem to be using anywhere close to 32GB of RAM. > > Has anyone else encountered similar issues or have any suggestions about > where to go in diagnosing the cause of this problem? Alternatively, is > this a problem I can safely ignore? Running some short code segments on > Spark seems to work just fine, but I'm wondering if this will become a > problem at heavy loads. Please let me know if there's any other info that > would be helpful. >
