RFR(XXS) 8190729: Adjustment to anonymous metadata space chunk allocation algorithm (original) (raw)

Thomas Stüfe thomas.stuefe at gmail.com
Mon Nov 27 18:36:51 UTC 2017

Hi Zhengyu,

looks good.

(I wonder whether if you just could have given the SpaceManager a back pointer to its enclosing Metaspace instead of handing down _space_type and mdType, but then, those are cleanups we can do later.)

In SpaceManager, could you please make _space_type - and _mdtype too - const?

Why do you do this only for non-class metaspace?

Could please you embellish the comment in calc_chunk_size() a bit?

Would it be possible to have a regression test for this one, maybe as part of the NMT tests? So that if we change the allocator logic in the future, we can be sure not to reintroduce that bug. To me, it would be okay to do this in a follow up item, if you are concerned about the jdk10 code freeze.

Thank you,

Kind Regards, Thomas

On Mon, Nov 27, 2017 at 6:09 PM, Zhengyu Gu <zgu at redhat.com> wrote:

Hi,

To follow up recent discussion of anonymous metadata space memory usage, I would like submit this simple fix for review. By continuing allocating anonymous metadata space from SpecializeChunk pool, up to 4 chunks, it reduces the waste from 60+% to about 30%.

Bug: https://bugs.openjdk.java.net/browse/JDK-8190729 Webrev: http://cr.openjdk.java.net/~zgu/8190729/webrev.00/ Test: hotspottier1runtime on Linux x64 (fastdebug and release) Thanks, -Zhengyu

More information about the hotspot-dev mailing list