Hi Paulo! Hmm, interesting. The high discrepancy between virtual and physical memory usually means that the process either maps large files into memory, or that it pre-allocates a lot of memory without immediately using it. Neither of these things are done by Flink.
Could this be an effect of either the Docker environment (mapping certain kernel spaces / libraries / whatever) or a result of one of the libraries (gRPC or so)? Stephan On Mon, Dec 19, 2016 at 12:32 PM, Paulo Cezar <paulo.ce...@gogeo.io> wrote: > - Are you using RocksDB? > > No. > > > - What is your flink configuration, especially around memory settings? > > I'm using default config with 2GB for jobmanager and 5GB for taskmanagers. > I'm starting flink via "./bin/yarn-session.sh -d -n 5 -jm 2048 -tm 5120 -s > 4 -nm 'Flink'" > > - What do you use for TaskManager heap size? Any manual value, or do you > let Flink/Yarn set it automatically based on container size? > > No manual values here. YARN config is pretty much default with maximum > allocation of 12GB of physical memory and ratio between virtual memory to > physical memory 2.1 (via yarn.nodemanager.vmem-pmem-ratio). > > > - Do you use any libraries or connectors in your program? > > I'm using flink-connector-kafka-0.10_2.11, a MongoDB client, a gRPC > client and some http libraries like unirest and Apache HttpClient. > > - Also, can you tell us what OS you are running on? > > My YARN cluster runs on Docker containers (docker version 1.12) with > images based on Ubuntu 14.04. Host OS is Ubuntu 14.04.4 LTS (GNU/Linux > 3.19.0-65-generic x86_64). > >