ParNew 4x slower under 1.7 vs 1.6?

[Andrew Mulholland]

On 8/10/13 2:40 AM, "Bernd Eckenfels" <bernd.eckenfels at googlemail.com>
wrote:

>Am 10.08.2013, 00:48 Uhr, schrieb Andrew Mulholland
><anmulholland at expedia.com>:
>> Under 1.6, we see:
>> - average ParNew pause times of ~100ms
>>    - with standard deviation of 0.12
>> - max of ~800ms
>> - cumulative total of ~11 seconds ParNew GC during the test run
>>
>> Under 1.7 we see:
>> -  average ParNew pause times of ~440ms
>>     - with standard deviation of 0.26
>> - max of ~1600s
>> - cumulative total of ~40 seconds PewNew GC during the test run
>
>Sounds pretty mich like it is using less threads for the collection. Did
>you tried to run it with -XX:+PrintFlagsFinal and have a look at
>ParallelGCThreads? (CommandLineFlags shows only the ones which get
>calculated based on some compat).

Thanks for the reply..  In a diff between print flags for the different
versions, I didn't initially pick up anything here - as I saw that:

1.6:    uintx ParallelGCThreads                        := 8
{product}
1.7:    uintx ParallelGCThreads                         = 8
{product}


Is there any the significance in the := 8 on 1.6  vs the = 8 on 1.7?

I've tried overriding the ParallelGCThreads, to 6, 8 etc, but seen no
difference in performance on 1.7

The server I'm testing on has 8 cores.


Thanks

Andrew