I'm having a problem with Solr under Tomcat unexpectedly taking a long time to respond to queries. As part of some stress testing, I wrote a bot that just does random word searches on my Solr install, and my responses typically come back in 10-50 ms. The queries are just 1-3 random words from /usr/share/dict/words, and I cap off the results at 2500 hits.
The queries run just fine and I typically get responses up to 50ms for large result sets. Here's an example of my log: TIME HITS MS SEARCH WORDS 12:33:20 1 5 hoovey Aruru kwachas 12:33:20 8 5 blinis twyver 12:33:20 2500 34 prework burlily sunshine 12:33:20 1928 30 rendu Solly 12:33:20 unnethe 12:33:20 gadwell afterpeak 12:33:20 792 14 steen 12:33:20 4 7 blanchi repaving 12:33:20 32 6 torbanite Storz ungag 12:33:20 7 5 chemostat 12:33:20 15 6 Guauaenok Adao lakist 12:33:20 6 6 bechance viny 12:33:20 20 6 chagigah 12:33:22 532 2404 bonne 12:33:22 143 9 nonman Norrie 12:33:22 24 6 repealers 12:33:22 Pfosi laniard locutory 12:33:22 51 6 sexipolar wordsmith enshield 12:33:22 loggiest Aryanise koels 12:33:22 fogyish unforcing 12:33:22 4 5 Millvale chokies 12:33:22 5 6 Melfa ripal Olva 12:33:22 15 6 apio Heraea latimeria 12:33:22 4 5 nonnitric parleying See that one line where it 2404ms to return? I get those about once a minute, but not at a regular interval. I ran this for two hours and got 122 spikes in 120 minutes. I ran it overnight and came in to work to find that there were 1283 spikes in 1260 minutes. So that one-a-minute is a pattern. As I write this, I'm in IRC with Chris Hostetter and he says: --snip-- Probably need to tweak your garbage collector settings to something that doesn't involve "stop the world" ... the specifics of the changes largely depend on what JVM you are using, what options you already have set, etc. markrmiller wrote a good blog about this a little while back: http://searchhub.org/dev/2011/03/27/garbage-collection-bootcamp-1-0/ There's also some notes here in the LucidWorks Solr Ref Guide: http://lucidworks.lucidimagination.com/display/solr/JVM+Settings --snip-- GC certainly sounds like a reasonable suspect. Any other suggestions? Any hints on Solr-specific GC tuning? I'm currently scouring Google. Thanks, xoa -- Andy Lester => a...@petdance.com => www.petdance.com => AIM:petdance
