ES is using hybrid MMap/NIO FS as default. Downside of this setup is that on our nodes ES (1.5.2) only leaves ~300 MB free for the operating system (Ubuntu 14.04.2) ... is there a way to limit the Off-Heap memory footprint used by Lucene or is the only way to control it re-indexing to pure NIO FS (hard, because large index but possible because SSDs used for storage). I have also not found what percentage of memory MMap leaves unused ... always around 300-350 MB which sometimes is causing troubles for the OS ...
-XX:MaxDirectMemorySize JVM directive with a limit to trigger garbage collection on off-heap memory.
Note, please check
elasticsearch.in.sh to remove/comment the line
-XX:+DisableExplicitGC avoids any garbage collection calls to release off-heap memory.
Thanks, will try it.
ES_DIRECT_SIZE I have found too, the
-XX:+DisableExplicitGC I would have overlooked.