Skip to main content

G1 benchmarks

5 replies [Last post]
linuxhippy
Offline
Joined: 2004-01-07
Points: 0

Hi,

Does anybody know some benchmark results of the new Garbage First collector for larger workloads/boxes? How well does it compared to CMS on mid-size/large hardware?

Thanks, Clemens

Reply viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.
mbien
Offline
Joined: 2007-04-29
Points: 0

i tried the early access G1 in one of my old pet projects and it was a bit disappointing. The GC (and overall application throughput) was noticeable lower than any of the 3 combinations of the other gcs. I think i ran into a bug since it kept one or two cores (i7 quadcore) under full load even in situations in which almost no garbage was created.

linuxhippy
Offline
Joined: 2004-01-07
Points: 0

peter: Thanks for the link to the pdf. Unfourtunatly it also does not contain benchmark numbers :-/
mbien: Interesting, thanks.

Its strange that for such a hype piece no numbers are available. It seems nobody has really tried that stuff ;)

Thanks, Clemens

fred34
Offline
Joined: 2004-06-19
Points: 0

http://research.sun.com/jtech/pubs/04-g1-paper-ismm.pdf

Here are the benchmarks and graphs!

peter__lawrey
Offline
Joined: 2005-11-01
Points: 0

While this is a relevant piece of work, its a conversion of a pre-1.5 version of Java for a research project, written about 5 years ago. At least one of the researchers worked on the G1 collector, but I would be careful about drawing to much from the results, or even assuming the solution is the same. Note: this compares CMS as it was before Java 1.5 was released.

peter__lawrey
Offline
Joined: 2005-11-01
Points: 0

This presentation by Tony Printezis may be of use to you. I attended a presentation he gave a month ago which used these slides.
http://developers.sun.com/learning/javaoneonline/2008/pdf/TS-5419.pdf

The impression I got was that minor GCs are much the same and should not be impacted. This is the result I have seen with Java 6u14. It is major/full GCs which should be dramatically improved. However the systems I work on do a good job of avoiding major GCs so I'm not sure I could comment whether it works. (I have tested the G1 on systems using a 2 GB heap)

However, I believe this feature is still experimental and not ready for production. e.g. When I use this feature running IntelliJ, the JVM dies within hours so I wouldn't use it in production yet.