Fwd: negative pause times
Y. Srinivas Ramakrishna
y.s.ramakrishna at oracle.com
Fri Sep 3 12:45:43 PDT 2010
You are right. i suspect it'll screw up ergonomics (not just for G1,
but for some of our other collectors as well, i suspect).
We need to harden our ergonomics (i.e. at least so it does no evil,
perhaps doing nothing and holding current control settings unchanged)
when such time-freeze anomalies are encountered. I am not sure current
ergonomic control deals properly with such situations, so an RFE
may be in order in case one does not already exist.
Angelika Langer wrote:
> Thanks for the hint.
> The negative pause times appeared in a GC traces that a colleague gave
> us to share his experiences with G1. It was produced on Linux and quite
> possibly using VMware. So, your hint helps explaining the negative
> pause time. Thank you.
> Whether it is worth a bug report we do not know. The question is: does
> the time drift affect G1 and its algorithms? After all, the collector
> gathers statistics in order to figure out how long the next gc pause
> will be. It must take time stamps for this purpose.
> Does G1 work reliably in a situation with time drift problems?
>>> Does anybody know off hand why the G1 log file shows negative pause
>>> 222.121: [GC pause (young) 561M->528M(1005M), 0.0586070 secs]
>>> 222.915: [GC pause (young) 576M->540M(1005M), 0.0447820 secs]
>>> 223.763: [GC pause (young) 602M->562M(1005M), -0.7354870 secs] <=======
>>> 225.572: [GC pause (young) 772M->632M(1005M), 0.0719350 secs]
>> Is this on Linux? Are you running NTP or otherwise causing TOD to
>> be adjusted? (you might want to turn off NTP if the negative times
>> bother you.) Are you running on VMware? (try a non-virtualized
>> run to see if the problem reproduces.) The negative pause times
>> we have seen in the past have usually been related to one of the
>> other of the above.
>> Please file a formal bug using your support credentials should
>> you believe this is an issue that needs delving deeper, especially
>> if you have a usable test case.
> hotspot-gc-use mailing list
> hotspot-gc-use at openjdk.java.net
More information about the hotspot-gc-use