Hi, your observation sounds like a bug to me and we have to further investigate it. I assume that you’re running a batch job, right? Could you maybe share your complete configuration and the job to reproduce the problem with us?
I think that your investigation that direct buffers are not properly freed and garbage collected can be right. I will open a JIRA issue to further investigate and solve the problem. Thanks for reporting :-) At the moment, one way to solve this problem is, as you’ve already stated, to set taskmanager.memory.preallocate: true in your configuration. For batch jobs, this should actually improve the runtime performance at the cost of a slightly longer start-up time for your TaskManagers. Cheers, Till On Sun, Jun 19, 2016 at 6:16 PM, CPC <acha...@gmail.com> wrote: > Hi, > > I think i found some information regarding this behavior. In jvm it is > almost imposible to free allocated memory via ByteBuffer.allocateDirect. > There is no explicit way to say jvm "free this direct bytebuffer". In some > forums they said you can free memory with below method: > >> def releaseBuffers(buffers:List[ByteBuffer]):List[ByteBuffer] = { >> >> if(!buffers.isEmpty){ >> >> val cleanerMethod = buffers.head.getClass.getMethod("cleaner") >> >> cleanerMethod.setAccessible(true) >> >> buffers.foreach{buffer=> >> >> val cleaner = cleanerMethod.invoke(buffer) >> >> val cleanMethod = cleaner.getClass().getMethod("clean") >> >> cleanMethod.setAccessible(true) >> >> cleanMethod.invoke(cleaner) >> >> } >> >> } >> >> List.empty[ByteBuffer] >> >> } >> >> > but since cleaner method is an internal method ,above is not recommended > and not working in every jvm and java 9 does not support it also. I also > made some tests with above method and behavior is not predictable. If > memory allocated by some other thread and that thread exit then it release > memory. Actually GC controls directMemory buffers. If there is no gc > activity and memory is allocated and then dereferenced by different threads > memory usage goes beyond intended and machine goes to swap then os kills > taskmanager. In my tests i saw that behaviour: > > Suppose that thread A allocated 8gb memory exit and there is no reference > to allocated memory > than thread B allocated 8gb memory exit and there is no reference to > allocated memory > > when i look at direct memory usage from jvisualvm it looks like > below(-Xmx512m -XX:MaxDirectMemorySize=12G) > > [image: Inline images 1] > > but RSS of the process is 16 GB. If i call System.gc at that point RSS > drops to 8GB but not to expected point. > > This is why Apache cassandra guys select sun.misc.Unsafe( > http://cassandra-user-incubator-apache-org.3065146.n2.nabble.com/Off-heap-caching-through-ByteBuffer-allocateDirect-when-JNA-not-available-td6977711.html > ). > > I think currently only way to limit memory usage in flink if you want to > use same taskmanager across jobs is via "taskmanager.memory.preallocate: > true". Since it allocate memory at the beginning and not freed its memory > usage stays constant. > > PS: Sorry for my english i am not a native speaker. I hope i can explain > what i intended to :) > > > > On 18 June 2016 at 16:36, CPC <acha...@gmail.com> wrote: > >> Hello, >> >> I repeated the same test with conf values. >> >>> taskmanager.heap.mb: 6500 >>> >>> taskmanager.memory.off-heap: true >>> >>> taskmanager.memory.fraction: 0.9 >>> >>> >> i set TM_MAX_OFFHEAP_SIZE="6G" in taskmanager sh. Taskmanager started >> with >> >>> capacman 14543 323 56.0 17014744 13731328 pts/1 Sl 16:23 35:25 >>> /home/capacman/programlama/java/jdk1.7.0_75/bin/java >>> -XX:+UseConcMarkSweepGC -XX:+CMSClassUnloadingEnabled -Xms650M -Xmx650M >>> -XX:MaxDirectMemorySize=6G -XX:MaxPermSize=256m >>> -Dlog.file=/home/capacman/Data/programlama/flink-1.0.3/log/flink-capacman-taskmanager-0-capacman-Aspire-V3-771.log >>> -Dlog4j.configuration=file:/home/capacman/Data/programlama/flink-1.0.3/conf/log4j.properties >>> -Dlogback.configurationFile=file:/home/capacman/Data/programlama/flink-1.0.3/conf/logback.xml >>> -classpath >>> /home/capacman/Data/programlama/flink-1.0.3/lib/flink-dist_2.11-1.0.3.jar:/home/capacman/Data/programlama/flink-1.0.3/lib/flink-python_2.11-1.0.3.jar:/home/capacman/Data/programlama/flink-1.0.3/lib/log4j-1.2.17.jar:/home/capacman/Data/programlama/flink-1.0.3/lib/slf4j-log4j12-1.7.7.jar::: >>> org.apache.flink.runtime.taskmanager.TaskManager --configDir >>> /home/capacman/Data/programlama/flink-1.0.3/conf >>> >> >> but memory usage reach up to 13Gb. Could somebodey explain me why memory >> usage is so high? I expect it to be at most 8GB with some jvm internal >> overhead. >> >> [image: Inline images 1] >> >> [image: Inline images 2] >> >> On 17 June 2016 at 20:26, CPC <acha...@gmail.com> wrote: >> >>> Hi, >>> >>> I am making some test about offheap memory usage and encounter an odd >>> behavior. My taskmanager heap limit is 12288 Mb and when i set >>> "taskmanager.memory.off-hep:true" for every job it allocates 11673 Mb off >>> heap area at most which is heapsize*0.95(value of >>> taskmanager.memory.fraction). But when i submit second job it allocated >>> another 11GB and does not free memory since MaxDirectMemorySize set to >>> -XX:MaxDirectMemorySize=${TM_MAX_OFFHEAP_SIZE}" which is >>> TM_MAX_OFFHEAP_SIZE="8388607T" and my laptop goes to swap then kernel oom >>> killed taskmanager. If i hit perform gc from visualvm between jobs then it >>> release direct memory but memory usage of taskmanager in ps command is >>> still around 20GB(RSS) and 27GB(virtual size) in that case i could submit >>> my test job a few times without oom killed task manager but after 10 submit >>> it killed taskmanager again. I dont understand why jvm memory usage is >>> still high even if all direct memory released. Do you have any idea? Then >>> i set MaxDirectMemorySize to 12 GB in this case it freed direct memory >>> without any explicit gc triggering from visualvm but jvm process memory >>> usage was still high around 20GB(RSS) and 27GB(virtual size). After again >>> maybe 10 submit it killed taskmanager. I think this is a bug and make it >>> imposible to reuse taskmanagers without restarting them in standalone mode. >>> >>> [image: Inline images 1] >>> >>> [image: Inline images 2] >>> >> >> >