Continuous CMS Collections Followed By Concurrent Mode Failure

Elliot Barlas Elliot.Barlas at citrix.com
Tue Jul 15 01:31:50 UTC 2014


Alternatively, could my problem be due to fragmentation?  Are there segments of free address space that are too small to use that could eventually be responsible for exhausting the old generation?

-Elliot

________________________________
From: Kirk Pepperdine [kirk at kodewerk.com]
Sent: Wednesday, July 09, 2014 1:24 AM
To: Srinivas Ramakrishna
Cc: Elliot Barlas; hotspot-gc-dev at openjdk.java.net openjdk.java.net
Subject: Re: Continuous CMS Collections Followed By Concurrent Mode Failure

Hi Ramki,

No need for flame proof suites… The (promotion failed) has been separated from the ParNew. I can easily adjust for this change… it’s just one more in a line of 100s of trivial changes that don’t add value to the GC logs… <sighs> ;-)

— Kirk

On Jul 9, 2014, at 3:45 AM, Srinivas Ramakrishna <ysr1729 at gmail.com<mailto:ysr1729 at gmail.com>> wrote:


On Tue, Jul 8, 2014 at 12:06 AM, Kirk Pepperdine <kirk at kodewerk.com<mailto:kirk at kodewerk.com>> wrote:
...
Ramki, should we file a bug for the ParNew (promotion failed) being corrupted by a CMS cycle?


Can you elaborate what precisely you meant by "corrupted (formatting)" below?

Perhaps you want what seems to be a missing newline or something? You should check with Jon or one of the HotSpot/GC folks on whether and how easy it would be to fix, or otherwise modify your parser. (Kevlar suit donned, for yr response ;-)

-- ramki

That was actually my guess. There is only 1 perm record in the file and although it shows that perm is grossly over-sized (ok, there is only one record ;-)) all of the recovery comes from the CMF which suggests perm is involved. All of the other CMS cycles are clearly due to tenured never being below the initiating occupancy fraction. Even without a young collection the initial marks are constantly reporting an occupancy of >1920xxxK of 1929xxxK. Oddly enough the ParNew’s only once promoted enough to trip the concurrent mode failure and oddly enough each of the CMS cycles (without the intervening ParNew) seem to recover about 1xxK bytes per cycle.

2014-06-10T22:56:18.793-0700: 4999527.565: [GC [1 CMS-initial-mark: 1920286K(1926784K)] 2051254K(2080128K), 0.3388330 secs] [Times: user=0.33 sys=0.00, real=0.33 secs]
…...
2014-06-10T22:56:26.242-0700: 4999535.014: [GC2014-06-10T22:56:26.242-0700: 4999535.014: [ParNew2014-06-10T22:56:26.256-0700: 4999535.028: [CMS-concurrent-abortable-preclean: 1.948/3.114 secs] [Times: user=1.93 sys=0.12, real=3.11 secs]
 (promotion failed)
Desired survivor size 8716288 bytes, new threshold 6 (max 6)
- age   1:    1036320 bytes,    1036320 total
- age   2:     825248 bytes,    1861568 total
- age   3:     119024 bytes,    1980592 total
- age   4:     113784 bytes,    2094376 total
- age   5:     129024 bytes,    2223400 total
- age   6:     154976 bytes,    2378376 total
: 141769K->140729K(153344K), 0.3807730 secs]2014-06-10T22:56:26.623-0700: 4999535.395: [CMS
 (concurrent mode failure): 1920816K->50773K(1926784K), 28.3938140 secs] 2062055K->50773K(2080128K), [CMS Perm : 48657K->41071K(262144K)], 28.7750370 secs] [Times: user=1.65 sys=0.03, real=28.78 secs]
2014-06-10T22:58:41.361-0700: 4999670.133: [GC2014-06-10T22:58:41.361-0700: 4999670.133: [ParNew
Desired survivor size 8716288 bytes, new threshold 6 (max 6)
- age   1:    1239232 bytes,    1239232 total
: 136320K->2239K(153344K), 0.0223680 secs] 187093K->53012K(2080128K), 0.0226340 secs] [Times: user=0.05 sys=0.00, real=0.02 secs]

Note the corrupted formatting of the ParNew (promotion failed).



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/attachments/20140715/cf8e2666/attachment-0001.html>


More information about the hotspot-gc-dev mailing list