THIS IS A TEST INSTANCE. ALL YOUR CHANGES WILL BE LOST!!!!
...
Here's a graph of the GC times with the settings above. The heap size for this graph is -Xms4096M -Xmx6144M. All of the GC pauses are well under a second, and the vast majority of them are under a quarter second. The graph will be updated when more information is available:
gc-7u72-with-8m-region-200ms-45percentoccupancy.png]]
Some notes about the ! G1HeapRegionSize parameter: I was seeing a large number of "humongous allocations" in the GC log, which means that they are larger than 50 percent of a G1 heap region. Those allocations were about 2MB in size, and my 4GB minimum heap results in a default region size of 2MB, which is calculated by dividing the -Xms value by 2048. I suspect that those allocations were for filterCache entries, because the indexes on that instance are each about 16.5 million documents, and the bitset memory structure of a filter for an index that size is a little over 2MB. Setting the region size to 8MB ensures that those allocations are not categorized as humongous. See "Humongous Allocations" section on this blog entry for more information.
...
This is a graph of a session with the settings above on Java 7u72. The heap size for this graph is -Xms4096M -Xmx6144M. I am quite thrilled about this. You may notice that the graph shows a clearly different GC pattern starting at about 8 hours ... this is because a full index rebuild started at just before 8 hours and ran for about 8 hours. The bulk of the longer pauses occurred during the full reindex, but they are not numerous, and there are millions of documents being indexed at the time. Before indexing started, there were no long pauses, and after it stopped, there have only been a handful. I would like to fix all of the long pauses, but I can live with the results I am seeing, especially because the average GC pause time is only six hundredths of a second.
The PerfDisableSharedMem option is there because of something that is called the four month bug.
...
Note to self: The -XX:ParGCCardsPerStrideChunk option was mentioned to me on the #solr IRC channel. With values of 4096 and 32768, the IRC user was able to achieve 15% and 19% reductions in average pause time, respectively, with the maximum pause cut nearly in half. That option was added to the options you can find in the Solr 5.x start script, on a 12GB heap.
The UseLargePages Option
A note about the UseLargePages option: Currently I do not have huge pages allocated in my operating system. This option will not actually do anything unless you allocate memory to huge pages. If you do so, memory usage reporting with the "top" command will probably only show a few hundred MB of resident memory used by your Solr process, even if it is in fact using several gigabytes of heap. If you do enable huge pages in Linux, be aware that you might wish to turn off an operating system feature called transparent huge pages.
...