goody wrote > What I haven’t found is any hard evidence from performance tests showing > that G1 beats out a properly tuned CMS config. I am not basing my > statements on the sole findings of the test results I linked to. I’ve > found multiple sources which all state the same thing, specifically some > tests done against JIRA with default configurations of the various jvm > collectors showing CMS still on top. The performance on G1 just isn’t > there yet.
The key phrase there is "properly tuned" CMS. CMS is an absolute nightmare to tune. When you tune CMS, you have to tune a lot of knobs that are all based on the object allocation rate and how long the objects live for. If you have an application, such as a cache server, that has a warming period (where it is primarily filling its heap with cached objects, but also serving up content) and then a longer state where it is mainly serving up those requests (but with some of the cache being invalidated over time) - you have two massively different object allocation rates. There is no proper way to tune CMS for that - you have to either choose to have optimal settings for one of those scenarios or have a set of compromised settings that are kind of "okish" for both of them. G1 can handle both large and small heaps just fine, but the areas you are going to find G1 used exclusively are apps that make use of lots of heap. Getting CMS to perform well with a 50 or 60GB heap is an exercise in futility - but G1 can handle that with ease. There was an article published by linked in a while ago where they claim they had better performance from CMS than G1 - but that article also had a bunch of incorrect statements in it (I was going to offer suggestions to them on how to re-run their tests or fix their G1 cases, but their blog didn't allow comments so their article will likely live on forever with incorrect statements and be referenced by others who assume that the people at linkedIn know what they are doing) If you want to talk about GC tuning - I would suggest the hotspot-gc-users mailing list. It is frequented by the hotspot developers (who are experts in both CMS and G1) and also by GC tuning enthusiasts. Another good source would be Charlie Hunt and Monica Beckwith's presentations on G1 tuning. Charlie Hunt you may recognize as probably one of the leading experts on GC tuning. He has written probably the best book on the market about how to tune the CMS garbage collector - "Java Performance" - and in his presentation at Java One, _he_ openly admitted how much he hates tuning CMS (it really is a _massive_ pain in the rear to get right) and suggests people try using G1. The default tuning for G1 is going to behave far better and with less surprises than the default tuning for CMS - so for them to ship with G1 by default is a very good decision. http://www.infoq.com/presentations/java-g1 http://www.oracle.com/technetwork/articles/java/g1gc-1984535.html A very good video presentation on G1: https://www.parleys.com/share_channel.html#play/525528dbe4b0a43ac12124d7 -- View this message in context: http://forums.jfrog.org/Artifactory-3-0-4-performance-tuning-and-G1GC-vs-CMS-tp7579846p7579879.html Sent from the Artifactory - Users mailing list archive at Nabble.com. ------------------------------------------------------------------------------ Learn Graph Databases - Download FREE O'Reilly Book "Graph Databases" is the definitive new guide to graph databases and their applications. Written by three acclaimed leaders in the field, this first edition is now available. Download your free book today! http://p.sf.net/sfu/NeoTech _______________________________________________ Artifactory-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/artifactory-users
