CMS collection keep working during holiday
Jon.Masamitsu at Sun.COM
Mon Oct 6 15:17:16 UTC 2008
On 10/05/08 10:38, Ken-- at newsgroupstats.hk wrote:
> Anyone help? Someone has exactly same case to me and he posted the case at
> Jon Masamitsu's blog:
> Trying the new CMS in 1.6 and it seems to work nicely for a while but then
> it starts kicking out these CMS aborts. I can find nothing anywhere talking
> about what is happening. This happens over and over until a full GC happens.
>>From what I can tell the CMS collector never completes and is therefore not
> doing anything except taking CPU time. CMS: abort preclean due to time
> 407.606: [CMS-concurrent-abortable-preclean: 0.149/5.058 secs]
This "abort" message stops a second precleaning phase. This second
precleaning phase is only meant to run for a lesser amount of time
and the "abort" does not indicate any problem. That phase is meant
to be "aborted". It's rather an unfortunate choice of words
on our part.
> The CMS will be triggered by -XX:CMSInitiatingOccupancyFraction=70 but it
> seems that it has no effect but continue keeping CMS until a Full GC (caused
> by 'concurrent mode failure'). In my case, it always happen in non-busy
> hours but the CMS behave normally in busy / peak hours.
> I think it is not related to 'cms abort preclean' but it's really taking CPU
> time but do nothing (around 4 hours) and it has chance to have a big GC when
> the old-gen is full.
The CMS collections cycles are being started because the old (tenured)
generation does have > 70% occupancy (at least in this fragment from
the log that your sent earlier).
76632.391: [GC[YG occupancy: 176683 K (873856 K)]76632.391: [Rescan
(parallel) , 0.1555281 secs]76632.546: [weak refs processing, 0.0003456
secs] [1 CMS-remark: 2230653K(3145728K)] 2407337K(4019584K), 0.1559517 secs]
Looking at [1 CMS-remark: 2230653K(3145728K)] the 2230653K is the
amount of the old generation that is being used and 3145728K is
the total old generation size and 2230653K/3145728K is just over
I didn't find a concurrent mode failure in your mail from 10/1. Can
you sent a log showing the output with the concurrent mode failure?
> Best Regards,
More information about the hotspot-gc-dev