harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Xiao-Feng Li" <xiaofeng...@gmail.com>
Subject Re: [drlvm][gc] IdentityHashCode implementation on magics
Date Fri, 16 May 2008 06:18:58 GMT
On Fri, May 16, 2008 at 1:41 PM, Aleksey Shipilev
<aleksey.shipilev@gmail.com> wrote:
> Aha, so it's literally no way to disable buffered hashcode? That means
> identityHashCode GC helper can't be implemented as magic.

Why not? GC helper doesn't need to include buffered hash.

Thanks,
xiaofeng

> That's too bad, my profiling and measurements shows that
> identityHashCode computation is really hot on serialization workload
> and I'm pretty sure the same picture shows up on any IdentityHashMap
> user. Simple substitution of GC call in
> VMMemoryManager.getIdentityHashCode with ((jint)jobj->object >> 2)
> gives +20% boost to serialization benchmark even with JNI overhead in
> place.
>
> Thanks,
> Aleksey.
>
> On Fri, May 16, 2008 at 3:42 AM, Xiao-Feng Li <xiaofeng.li@gmail.com> wrote:
>> Aleksey, "attached hashcode" works together with "buffered hashcode".
>> "Buffered hashcode" is created in compacting GC (in other words, in
>> major collection). So if normally there is few major collections,
>> there will be very few buffered hashcodes created. In minor
>> collections, it uses the attached hashcode.
>>
>> Thanks,
>> xiaofeng
>>
>> On Fri, May 16, 2008 at 5:06 AM, Aleksey Shipilev
>> <aleksey.shipilev@gmail.com> wrote:
>>> Hi, Xiao Feng, GC gurus,
>>>
>>> I'm back to IdentityHashCode implementation on magics [1]. I had
>>> managed to inline the helper for System.identityHashCode(), but
>>> unfortunately I had stuck with buffered hashcode scheme. Basically,
>>> there is no support for buffered hashcodes in GCHelper and I doubt it
>>> could be implemented there as we need to traverse native data
>>> structure on hashcode lookup. Not saying about performance when there
>>> are many hashcodes in the run, I have a suspicion there is the problem
>>> too.
>>>
>>> So I wonder whether is it possible to turn on "attached hashcode"
>>> instead of "buffered hashcode" somehow to see the performance impact?
>>> I searched through sources and it's still unclear for me.
>>>
>>> Thanks,
>>> Aleksey.
>>>
>>> [1] https://issues.apache.org/jira/browse/HARMONY-5713
>>>
>>
>>
>>
>> --
>> http://xiao-feng.blogspot.com
>>
>



-- 
http://xiao-feng.blogspot.com

Mime
View raw message