JEP 214: Remove GC Combinations Deprecated in JDK 8
bengt.rutisson at oracle.com
Fri Oct 17 07:27:07 UTC 2014
On 2014-10-17 09:13, Kirk Pepperdine wrote:
> Hi Bengt,
> Well, it’s not like you need my support but… from my view out here, the G1 is really coming into it’s own. Just last night I had a discussion with yet another client on an issue they were having that causes memory to spontaneously bloat. Pause times from the parallel collectors are unacceptable, the pause times from CMS are ok but you can’t tune it to work with an operational load and then survive this spurious event. Getting the G1 to work well with a normal operation load while manage the spurious event while we work to kill it is easy…
Thanks for sharing this. It's always nice to get some positive feedback.
We often get to hear when things don't work but rarely hear anything
when they do work... :)
> On Oct 17, 2014, at 7:52 AM, Bengt Rutisson <bengt.rutisson at oracle.com> wrote:
>> Hi Kirk,
>> Thanks you for sharing your support for this change! I also remember the discussion we had regarding the deprecation of iCMS and I appreciate that you took the time to read this JEP too.
>> I'm especially happy that you find the improvements we have done to G1 useful. I agree that there is more work to be done to improve G1 even further, but it is nice to hear that our work is already paying off.
>> On 2014-10-16 22:08, Kirk Pepperdine wrote:
>>> Hi Mark,
>>> I just wanted voice support for this JEP.
>>> If I remember the original conversation correctly my concern was with the removal of iCMS is that was being used in a large number of applications with low latency requirements and I’m not sure that many of these people are aware of anything more than the most loudly advertised latest and greatest features in the JDK. iCMS was really the only option they had. However, in the last couple of months I’ve been been moving clients away from the older more traditional collector combinations to the latest implementations of the G1. The results have been quite good and in some cases stunning. Just yesterday I moved an app and while the bench is still running, the results without any tuning what so ever are far better than any other collector combination after some extensive tuning. That said, the G1 still places a bit of a drag on low latency applications when compared to using iCMS but there are also fewer rough edges (ie failure cases and constraints on heap size). So, the situation for G1 has improved quite dramatically since I voiced my first concerns about removal of iCMS. I think we all need to thank the great work of everyone who’s been working on G1 for this.
More information about the hotspot-gc-dev