Rtalk Performance for micro benchmarks (how to improve)

[Charles Oliver Nutter]

Your output is a little hard to parse, but I think I followed it.

First off, it is currently expected that jdk8 is not optimizing as
well as the jdk7 indy logic. I have seen very few benchmarks that are
faster, and my reading of the inlining logs tells me it's just not
inlining everything the jdk7 logic did. I have not yet gotten an hsdis
build to work on OpenJDK8 (help!).

If you still don't see performance comparing favorably to the fastruby
version (which, to be fair, is doing 100% virtual dispatch with no
indy, no user-created guards, etc), then could it simply be that your
indy guard logic and arbitrary precision logic adding all that
overhead? It seems like a lot indeed.

Your times definitely seem a slower than they should be based on the
bytecode. My machine is a 2.2GHz i7, and you're on a 2.8GHz core
2...that's probably about a toss-up on straight-line perf, in my
experience. So yeah, I guess I'd look at guard logic, your actual math
operations, and probably also the size of your numeric objects (my
RFixnum has exactly one field for the long value.

FWIW, my experience with JRuby is that I can come within 1.5x slower
than Java doing the exact same operations through all virtual calls,
which matches almost exactly with fastruby's 30% improvement for this
contrived case.

- Charlie

On Mon, Sep 17, 2012 at 7:46 PM, Mark Roos <mroos at roos.com> wrote:
> After reading Charles' blog on 'fast' ruby I decided to look at how Rtalk
> was
> comparing.  At the same time I loaded the latest JDK8 just to compare.
>
> First jdk 8 runs (excellent) with some things faster but most slower.  But
> to
> my chagrin Rtalk running FIB(35) is much slower than Charles' effort.  700ms
> vs 300ms.
> So I am wondering what I could be doing that could improve.  At the end are
> my
> jvm bytecodes for fib ( which look ok ).  I am doing all 64 bit and my
> integer
> code does do conversions to/from BigIntegers.  I also use a 4K small integer
> cache to help with object creation.
>
> To do some research I have two versions of Hanoi as well.  The Rtalk version
> and
> an implementation in Java where I do the exact work but without
> invokeDynamic.
> Here I run 100% slower vs Charles' 30%.
>
> Suggestions?  Cmd line args, compiler choice? where to look?
>
> Thanks
> mark
>
>
> 64 bit OSX 2.8 GHz core 2 duo
> No cmd line options
> Times are the min of 10 runs.
>
> JDK 7u2 Oracle site
> Starting FIB 35
>   Time (ms) 617
> Starting Fib 40
>   Time (ms) 6982
> Starting Java Hanoi 25 10X
>   Time (ms) 294
> Starting RTALK Hanoi 25
>   Time (ms) 629
>
> JDK 8 ( 9/14)  Google code b56
> Starting FIB 35
>   Time (ms) 776
> Starting Fib 40
>   Time (ms) 8668
> Starting Java Hanoi 25 10X
>   Time (ms) 292
> Starting RTALK Hanoi 25
>   Time (ms) 752
>
> Code generated by Rtalk compiler
> Push Constant is a ConstantCallSite
> Perform is an InvokeDynamic MutableCallSite with a single GWT
>
> CLASS                rtPbc/r957 extends Object
> METHOD               invoke RtObject,RtObject,RtObject, access=9
> FRAME                -1 localVarCnt=2 {ri/core/rtalk/RtObject,
> ri/core/rtalk/RtObject, null, null, null, null} stackCnt=0 {null, null}
> Push Constant        2
>                      aload 0
> Perform              <   line  7
>                      getstatic ri/core/rtalk/RtObject _false
> Lri/core/rtalk/RtObject;
> JUMP                 if_acmpeq LABEL 2
>                      aload 0
> JUMP                 goto LABEL 3
> LABEL                LABEL 2
> FRAME                -1 localVarCnt=2 {ri/core/rtalk/RtObject,
> ri/core/rtalk/RtObject, null, null, null, null} stackCnt=0 {null, null}
>                      aload 1
>                      astore 4
> Push Constant        1
>                      aload 0
> Perform              -   line  42
>                      aload 4
> Perform              supportFib:   line  60
>                      astore 4
>                      aload 1
>                      astore 5
> Push Constant        2
>                      aload 0
> Perform              -   line  87
>                      aload 5
> Perform              supportFib:   line  105
>                      aload 4
> Perform              +   line  123
> LABEL                LABEL 3
> FRAME                -1 localVarCnt=2 {ri/core/rtalk/RtObject,
> ri/core/rtalk/RtObject, null, null, null, null} stackCnt=1
> {ri/core/rtalk/RtObject, null}
>                      areturn
>                      maxStack 2, maxLocals 6
> _______________________________________________
> mlvm-dev mailing list
> mlvm-dev at openjdk.java.net
> http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev
>