RFR (S): 8155229: Tune thread usage for live data clearing

Derek White derek.white at oracle.com
Thu Apr 28 19:14:46 UTC 2016


On 4/28/16 10:28 AM, Mikael Gerdin wrote:
> Hi Thomas,
>
> On 2016-04-28 13:16, Thomas Schatzl wrote:
>> Hi all,
>>
>>    during testing of 8077144 I found that live bitmap clear can be
>> significantly sped up by not using that many threads.
>>
>> A good conservative estimate (tends to use too many threads) that still
>> cuts mark bitmap clear time by >50% is to give each thread at least 1
>> MB of data to clear.
>>
>> CR:
>> https://bugs.openjdk.java.net/browse/JDK-8155229
>> Webrev:
>> http://cr.openjdk.java.net/~tschatzl/8155229/webrev/

Hi Thomas,

Looks good!
>
> The fix looks good.
>
> One thing though,
> now that I think about it, would it make sense to add an additional 
> logging tag to
>
> +  log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT 
> " work units.", cl.name(), num_workers, num_chunks);
>
> Since these logging statements are not really about GC ergonomics but 
> rather GC worker ergonomics.
> This can be fixed further on too, I just want to hear if anyone else 
> gets the same feeling.
>
> /Mikael

Hi Mikael,

I have a slight preference for leaving as-is. From my (very) little 
experience with UL, I think keeping the tag set small is more helpful 
than a having a huge number of specific tags.

  - Derek

>> Testing:
>> jprt, webrev, various performance tests
>>
>> Depends on 8155232.
>>
>> Thanks,
>>    Thomas
>>



More information about the hotspot-gc-dev mailing list