RFR: 8020277: Young GC could be extremely slow due to assertion in ObjectStartArray::object_starts_in_range

[Stefan Johansson]

On 2014-01-21 14:44, Mikael Gerdin wrote:
> Stefan,
>
> On Tuesday 21 January 2014 14.17.20 Stefan Johansson wrote:
>> Hi,
>>
>> Can I have a couple of reviews for this fix for:
>> https://bugs.openjdk.java.net/browse/JDK-8020277
>>
>> Webrev:
>> http://cr.openjdk.java.net/~sjohanss/8020277/webrev.00/
> Looks good.
> /Mikael
Thanks for reviewing Mikael!
Stefan
>
>> Summary:
>> There have been a couple of bugs filed about varying/slow gc times for
>> ParallelScavenge in debug builds. The reason for the slow down is an
>> assertion in object_starts_in_range() that tries to verify that the
>> range (mapping to one card) really doesn't have any objects starting in
>> it. The assertion uses object_start() to find the first object starting
>> before the end of the range. The reason for this being really slow is
>> that object_start() won't stop looking for an object until it finds it,
>> even if it's out of the range we're currently verifying. So on a heap
>> with many large objects that span more than one card, a lot of extra
>> work is done.
>>
>> My proposed fix is to just remove this assertion. An alternative would
>> be to limit the verification to just check the range we are interested
>> in. I started looking at doing this but realized that such a limitation
>> would end up being very similar to what is currently done in
>> object_starts_in_range() and with this in mind I think we can just
>> remove the assertion. I checked JBS to see if there were many bugs
>> reported with this assertion as cause, but I only found this bug and
>> 8009462 which also lists this assertion as the reason for slow GC times.
>>
>> Testing:
>> * Manually verified that jit/graph/cgt10 doesn't hang anymore.
>>
>> Thanks,
>> Stefan