On 11 Oct 2016, at 8:42, Heiko W.Rupp wrote:

Hey,

tl;dr: we need to investigate heap usage - especially in the case of compression
kicking in - it looks like there could be a memory leak. Compression timing seems
mostly ok.

Looks like we are making some progress

• metric forwarding to alerts seems to do some memory-wise heavy work, that can be reduced (hwkmetrics-507 and -508 and hawkular-1130 iirc)
• internal hsqldb uses a lot of memory
  • that is "only" a developer concern
• using postgres does not exhibit the memory usage for caching as does hsqldb, but
  • the postgres driver in inventory (v...1201) seems to use finalizers too much
    • we are investigating to use a newer driver that may perform better

I just replaced in a local copy inventory 0.20 with 0.19.3 and this seems to be
smoother sailing - a lot less finalizers that also hold less memory.

I did try to run with -XX:+UseStringDeduplication which requires -XX:+UseG1GC, but
according to MAT it did not have an effect for me. We may want to look into
string interning for the future like we did in JON.