harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Aleksey Shipilev" <aleksey.shipi...@gmail.com>
Subject Re: [general][performance] Serialization patches need committer attention
Date Tue, 15 Apr 2008 05:34:03 GMT
Thanks, Tony!

There is microbenchmark in [1], is it too big for regression test?

Thanks,
Aleksey.

[1] https://issues.apache.org/jira/browse/HARMONY-5632

On Tue, Apr 15, 2008 at 7:08 AM, Tony Wu <wuyuehao@gmail.com> wrote:
> Hi, Aleksey
>  I'm looking at 5722. I agree with you on leveraging the Accessor but I
>  think it would be good if you can post some micro-benchmark on this
>  JIRA so that we can utilize it as regression test in future.
>
>
>
>  On 4/15/08, Aleksey Shipilev <aleksey.shipilev@gmail.com> wrote:
>  > Hi all,
>  >
>  > As you might noticed, there is active work on serialization
>  > performance improvements [1]. So far we had committed 4 patches
>  > concerning scalability and boosted the performance on multi-threaded
>  > serialization benchmarks more than 2x times. And this is not the end
>  > of the journey: more issues coming.
>  >
>  > Dear committers, can you please review and commit the patches in ready JIRAs?
>  > I will be happy if these two committed:
>  >
>  >  a. [classlib][luni][performance] ObjectInputStream/ObjectOutputStream
>  > can use optimized API for accessing fields - available at [2]
>  > This issue moves the setField/getField functionality to ObjectAccessor
>  > utility class. This eliminates one serious problem: creation of string
>  > during field name lookup on native side. The second advantage is that
>  > this solution checks security on accessor creation rather than on
>  > every call (like Reflection API does). This improvement gives another
>  > 30-120% boost for serialization workload.
>  >
>  > I expect more with rewriting these accessors on magics, thus
>  > eliminating JNI overhead and native-side work. Those who interested in
>  > this topic, feel free to join in corresponding JIRA [3].
>  >
>  >  b. [classlib][luni][performance] j.u.IdentityHashMap ariphmetic
>  > improvements - available at [4]
>  > After the moving of ObjectStreamClass cache to ThreadLocalCache [5],
>  > performance of serialization depends on ThreadLocal performance and
>  > then on IdentityHashMap performance. Simple benchmark shows that
>  > ThreadLocal performance is devastating: 50x slower than Sun 1.6.0_05
>  > [6]. This is very first ready patch from that chain of issues. It
>  > brings up +300% boost on ThreadLocal benchmark and +10% boost to
>  > serialization benchmark.
>  >
>  > Thanks,
>  > Aleksey.
>  >
>  > [1] https://issues.apache.org/jira/browse/HARMONY-5632
>  > [2] https://issues.apache.org/jira/browse/HARMONY-5722 (ready)
>  > [3] https://issues.apache.org/jira/browse/HARMONY-5725
>  > [4] https://issues.apache.org/jira/browse/HARMONY-5718 (ready)
>  > [5] https://issues.apache.org/jira/browse/HARMONY-5633
>  > [6] https://issues.apache.org/jira/browse/HARMONY-5703
>  >
>
>
>  --
>  Tony Wu
>  China Software Development Lab, IBM
>

Mime
View raw message