JEP 173: Remove Rarely-Used Combinations of Garbage Collectors

[Richard Warburton]

Hey,

Further to previous discussion on this topic - I have asked the London Java
Community's membership about who is using GCs proposed for removal.  Due to
London's strong financial services industries we have a number of people
who have low pause-time requirements.

Several members stated that they were using iCMS in production.  In all
cases that I talked to they had tuned their GC Logs with both CMS and iCMS
and identified situations where iCMS had superior performance.  Sadly I
didn't get information about how large their heap sizes were.  Though
people stated that they hadn't tried retuning with modern G1.

Furthermore Jack Shirazi pointed out that he was actually using SerialOld
in some cases, and with a good tuning reason.  This isn't necessarily a
good reason to keep SerialOld, but I'm trying to push back useful
feedback.  Hope this is helpful.

"Richard, I had a scan across the JVMs I support (several thousand) and
there were several on SerialOld. I have the notes on exactly why one has
that, as I tuned it (quite a while ago). The JVM is used in 2 modes, during
the day it needs to have low pause times and pretty much all these are
young gen as the old gen doesn't fill up. The pause times for the daytime
operating activity were significantly shorter using the ParNew young gen
collector than the scavenge collector, so that was used. This meant that
the old gen collector had to be either CMS or SerialOld.

The evening activity was batch oriented with the old gen getting repeatedly
full and needing to be cleared. The way the batch activity processed was
such that a large number of objects were promoted in spikes. The CMS was
never able to keep up, so it failed over practically every time, meaning
that just setting it to serial old turned out to be better for both
throughput and pause times.

It's possible I could have tuned CMS to achieve the same capability, but it
would have required a lot of effort where choosing Serial was simple and
quick to test. I'm not saying don't drop the combo, I can always work with
whatever is available, but thought I'd give you the use-case."

regards,

  Richard
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20121214/fe4be7f2/attachment.htm>