RFR: 7169803: Usage of pretenured value is not correct
David Lindholm
david.lindholm at oracle.com
Mon Jun 15 12:07:25 UTC 2015
Tao and Jon,
I have changed the patch now so that pretenured is not part of the
calculation at all:
http://cr.openjdk.java.net/~david/JDK-7169803/webrev.01/
Thanks,
David
On 2015-06-14 20:58, Tao Mao wrote:
> The code seems to only consider survisor_size and promoted_size and
> ignore pretenured size for most part, except here
>
> avg_promoted()->sample(promoted + _avg_pretenured->padded_average());
>
> The naming of avg_promoted() is probably not correct either if we want
> to make it consistent with what it takes in.
>
> To take all involved sizes into considerations, we need to fix all
> over the places when we make decisions based on the sizes.
>
>
>
> Jon, you could be right. The chances are GCSizePolicy is just doing
> all right and pretenured size won't matter too much whether it's in
> the calculation or not. That needs to be instrumented, of course.
>
> Thanks.
> Tao
>
>
>
>
>
> On Thu, Jun 11, 2015 at 10:22 AM, Jon Masamitsu
> <jon.masamitsu at oracle.com <mailto:jon.masamitsu at oracle.com>> wrote:
>
>
>
> On 06/11/2015 12:47 AM, Tao Mao wrote:
>> Hi Jon,
>>
>> The definition is: "Promoted" is objects that are promoted (or
>> tenured) during yGC while "PRE-tenured" is objects that are
>> allocated directly in old gen. I believe this is so, right?
>
> Tao,
>
> Yes, you're right about the definitions. I think then the question is
> what do you want to passed to
>
> avg_promoted()->sample()
>
> I would say that you want to pass promoted as calculated by
>
> 489 size_t promoted = old_gen->used_in_bytes() -
> old_gen_used_before;
>
> I think the bug is that _avg_pretenured->padded_average() should
> not be
> passed to sample() in the original code
>
> 1307 avg_promoted()->sample(promoted +
> _avg_pretenured->padded_average());
>
> Since _avg_pretenured() should have been calculated using bytes
> but was using
> words, the bug was had less of an affect. Also the amount
> pretenured was likely
> small in the majority of cases. And the end affect would be that
> the number
> used for promoted would be an over-estimate and just make the
> ergonomics more
> conservative.
>
> David,
>
> If you could provide me with jprt binaries for before and after
> your fix,
> I'll do some experiments to see if there are any change in the number
> of GC (young and full) to see if we're changing the ergonomics. It
> could be that the bug is a "lucky" mistake that makes things better.
> Unlikely but I'd like to look.
>
> Thanks.
>
>
> Jon
>
>
>>
>> Thanks.
>> Tao Mao
>>
>>
>>
>> On Wed, Jun 10, 2015 at 6:48 PM, Jon Masamitsu
>> <jon.masamitsu at oracle.com <mailto:jon.masamitsu at oracle.com>> wrote:
>>
>>
>>
>> On 6/10/2015 5:51 PM, Tao Mao wrote:
>>> I think David's right. Promoted objects and pretenured
>>> objects are different guys. This would resolve the second
>>> issue in the ticket JDK-7169803.
>>
>> Tao,
>>
>> How would you define each?
>>
>> Jon
>>
>>
>>>
>>> Thanks.
>>> Tao Mao
>>>
>>> On Wed, Jun 10, 2015 at 5:08 PM, Jon Masamitsu
>>> <jon.masamitsu at oracle.com <mailto:jon.masamitsu at oracle.com>>
>>> wrote:
>>>
>>>
>>>
>>> On 6/10/2015 1:30 AM, David Lindholm wrote:
>>>
>>> Jon,
>>>
>>> Thanks for taking a look at this. No, I don't think
>>> this will lead to double counting. This calculation
>>> is already there, it is just buggy. Before this
>>> patch the code added the total amount of promoted
>>> memory this collection with the average size of the
>>> pretenured objects, and used this as a sample. Now
>>> the code adds the total amount of promoted memory
>>> with the total size of the pretenured objects since
>>> last collection, and uses this as a sample instead.
>>>
>>>
>>> Aren't "promoted" and
>>> "total_pretenured_since_last_promotion()" approximately
>>> the same
>>> thing? In share/vm/gc/parallel/psScavenge.cpp
>>>
>>> 489 size_t promoted = old_gen->used_in_bytes() -
>>> old_gen_used_before;
>>> 490 size_policy->update_averages(_survivor_overflow,
>>> survived, promoted);
>>>
>>> so "promoted" is the change in the old gen between the
>>> before and after the
>>> young gen collection.
>>>
>>> "total_pretenured_size_last_promotion()" is
>>>
>>> 258 void tenured_allocation(size_t size) {
>>> 259 _avg_pretenured->sample(size);
>>> 260 add_pretenured_since_last_promotion(size)
>>>
>>>
>>> which seems to me to be calculating the same thing (sum
>>> of allocations into the old gen).
>>>
>>> Not so?
>>>
>>> Jon
>>>
>>>
>>>
>>>
>>> Thanks,
>>> David
>>>
>>> On 2015-06-09 21:37, Jon Masamitsu wrote:
>>>
>>> David,
>>>
>>> http://cr.openjdk.java.net/~david/JDK-7169803/webrev.00/src/share/vm/gc/parallel/psAdaptiveSizePolicy.cpp.frames.html
>>> <http://cr.openjdk.java.net/%7Edavid/JDK-7169803/webrev.00/src/share/vm/gc/parallel/psAdaptiveSizePolicy.cpp.frames.html>
>>>
>>>
>>> 1308 avg_promoted()->sample(promoted +
>>> total_pretenured_since_last_promotion());
>>>
>>>
>>> Is including both "promoted" and
>>> "total_pretenured_since_last_promotion()"
>>> double counting?
>>>
>>> Jon
>>>
>>> On 06/09/2015 02:06 AM, David Lindholm wrote:
>>>
>>> Hi,
>>>
>>> Please review this patch that corrects the
>>> usage of the pretenured value. There were 2
>>> issues: words and bytes were mixed up and
>>> the addition was done with the wrong value.
>>> See bug for details.
>>>
>>> Webrev:
>>> http://cr.openjdk.java.net/~david/JDK-7169803/webrev.00/
>>> <http://cr.openjdk.java.net/%7Edavid/JDK-7169803/webrev.00/>
>>> Bug:
>>> https://bugs.openjdk.java.net/browse/JDK-7169803
>>>
>>>
>>> Testing: Passed JPRT
>>>
>>>
>>> Thanks,
>>> David
>>>
>>>
>>>
>>>
>>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20150615/39f1b082/attachment.htm>
More information about the hotspot-gc-dev
mailing list