RFR: 8191471: Elastic TLABs for G1
per.liden at oracle.com
Wed May 2 11:49:51 UTC 2018
On 05/02/2018 11:28 AM, Stefan Johansson wrote:
> Thanks for the review Per,
> On 2018-04-24 11:19, Per Liden wrote:
>> Hi Stefan,
>> On 04/23/2018 02:57 PM, Stefan Johansson wrote:
>>> Please review these changes to lower the waste in G1 mutator regions.
>>> JBS: https://bugs.openjdk.java.net/browse/JDK-8191471
>>> Webrev: http://cr.openjdk.java.net/~sjohanss/8191471/00/
>> Looks good overall, just two comments:
>> 374 size_t minimal_tlab_size =
>> MAX2(ThreadLocalAllocBuffer::compute_min_size(size), MinTLABSize);
>> Can we move the MAX2-logic into compute_min_size()? It seems to belong
>> in there, rather than here?
> I agree and I changed it. This will however change the logic in
> compute_size a bit, it didn't take MinTLABSize into consideration
> before, but now does. I still think the change is good though.
>> 133 virtual HeapWord* allocate_new_tlab(size_t min_word_size,
>> 134 size_t desired_word_size,
>> 135 size_t* actual_word_size);
>> The argument names now vary between the collectors. I suggest we use
>> the same names in all places and skip the "word" part. Something like:
>> min_size, requested_size, actual_size?
> New webrevs:
> Full: http://cr.openjdk.java.net/~sjohanss/8191471/02
> Inc: http://cr.openjdk.java.net/~sjohanss/8191471/01-02
> Will re-run testing before pushing to make sure the small change don't
> cause any trouble.
>>> G1 might waste excessive amount of memory due to allocations that
>>> would span region boundaries. To lower this waste this patch
>>> addresses this problem in two ways. TLAB allocations are now more
>>> flexible allowing the size to be less then the desired amount as long
>>> as the allocation will fit in the TLAB (defined by min_word_size in
>>> the call).
>>> The G1 allocation code also tries to minimize the waste by keeping
>>> two active allocation regions so that a single large allocation won't
>>> cause waste that could have been used for other allocations.
>>> Functional testing through mach5 with hs-tier 1-3 and jdk-tier 1-3.
>>> Performance testing locally and through aurora without seeing any
More information about the hotspot-gc-dev