Sounds like 64bit is OK. Would be work re-testing that G1GC assert trip with the latest JDK.
On 14 March 2015 at 01:46, Shawn Heisey <apa...@elyograg.org> wrote: > On 3/12/2015 5:11 PM, Markus Jelsma wrote: > > Hello - i would assume off-heap would out perform any heap based data > structure. G1 is only useful if you deal with very large heaps, and it eats > CPU at the same time. As much as G1 is better than CMS in same cases, you > would still have less wasted CPU time and resp. less STW events. > > > > Anyway. if someone has a setup at hand to provide details, please do :) > > I do not have any info about CPU usage with G1 vs. CMS. My Solr servers > are extremely lightly loaded so CPU usage is never a problem. For very > busy servers, that could be an issue, and CMS is probably the way to go. > > I have been able to reduce GC times (both the average time and the "long > collection" time) by a large margin using G1 vs. CMS, with the help of > the hotspot-gc-use mailing list maintained by the OpenJDK project. > > I keep my mad ramblings about GC tuning on the Solr wiki: > > http://wiki.apache.org/solr/ShawnHeisey#GC_Tuning_for_Solr > > The only caveat that I have is that the Lucene project has said never to > use G1GC with Lucene. Their exact words, found on the wiki URL below, > are: "Do not, under any circumstances, run Lucene with the G1 garbage > collector. " > > I can say that I have never had a problem with G1, but if you choose to > use it, you should know that you are going against the advice of the > people who write the code: > > > http://wiki.apache.org/lucene-java/JavaBugs#Oracle_Java_.2F_Sun_Java_.2F_OpenJDK_Bugs > > Thanks, > Shawn > > -- Damien Kamerman
