Significant indy slowdown on JRuby + redblack bench

[Ben Evans]

On Wed, Feb 23, 2011 at 8:11 PM, Charles Oliver Nutter
<headius at headius.com>wrote:

> On Wed, Feb 23, 2011 at 1:25 PM, John Rose <john.r.rose at oracle.com> wrote:
> > These results are very useful, and you are very welcome to post them.  We
> all need to see how things are shaping up in user experiences so that we can
> best allocate our time.
> >
> > Here's the "but":  Christian and I are working on bugs and
> refactoring/renaming this month and next.  This means that we have to treat
> performance issues (both the good and the bad) as a side activity for now.
>  But ("but" squared) you can tempt us all you want to work on performance,
> and sometimes you will succeed.
> >
> > At some point, we'll have to take stock of performance issues and start
> shooting them down.  At that point, I hope you guys will have a juicy list
> of performance pain points for us.
>
> That's all perfectly fair. Perhaps there's a public bug tracker we
> could use to start aggregating perf issues and conversations?
>
> Now a request for pure speculation: How much room do you think we have
> to improve perf from here? As far as what I'm hoping to see, I'd like
> to have indy be competitive with my dynopt logic, since they are
> structurally identical. If that's possible, it could have enormous
> ramifications for JRuby performance in general. I'm very excited by
> that possibility.
>

I'd also add that there's potentially a difference between the performance
at 7u0 and (say) 7u4 or 7u6.

To my mind, what matters for the 7u0 release is the completeness and
correctness of the feature and supporting APIs. We know from experience
tracking the 6uX release train that performance of new features can and will
improve post major release.

indy is a feature that is not going to be used by very many developers -
only really by framework or language designers. In fact, most of the initial
users are on this mailing list, and would hopefully understand if the
performance at 7u0 was still in the "room for improvement" camp - provided
that the API etc are stable, and actually ship with 7u0.

MethodHandles are a slightly different case, as I do see them being used
more widely by end-user developers, rather than framework-builders. I'm
hoping to get some cycles to do some benchmarking on some simple use cases
soon - will post anything interesting I find.

Thanks,

Ben
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/mlvm-dev/attachments/20110224/6f615082/attachment.html