[Cuis] Benching Cog (was Re: A few proposed is: additions)

[Casey ransberger]

Sadly, don't have my machine with me. On Cog, run it two or three times before benching. I can't recall if it's the second or third run that does the trick (I think it's the second,) but Cog doesn't actually start putting JIT'd methods into the poly whosit whatchamacallit cache (PIC) on the first run. It's sort of a stopgap until the work is done to make it smarter about inlining machine code, IIRC.

The fact that your slowest run was took the same amount of time on the interpreter suggests to me that on your slowest run, the method implementation wasn't JIT'd or cached. Of course I'm a VM noob so don't take my word for it; CC Eliot so he has an opportunity to set me straight (which would be a very nice thing to do!) if I'm wrong here.

Casey



On Jun 18, 2012, at 8:33 AM, "Phil (list)" <pbpublist at gmail.com> wrote:

> Casey,
> 
> On Jun 18, 2012, at 3:33 AM, Casey Ransberger wrote:
> 
>> Top post: a lot of iterations to see the difference? That's different. Still wondering if I'm still missing the point of this thread, but are you running on Cog or the interpreter?
> 
> I'm running Cog on an i7-2600.  On that configuration, the fastest test completes in 0.07 seconds and the slowest in 0.7 seconds.  I just ran the same tests under the classic VM and the fastest test completes in 0.7 seconds and the slowest in 3 seconds (the relative performance of old vs. new changes to from 2x faster to 2x slower).  What results are you seeing on your machine? (if necessary, feel free to lower the iterations in LetsGetDownToIssness #timing from 5,000,000)
> 
> Thanks,
> Phil
> _______________________________________________
> Cuis mailing list
> Cuis at jvuletich.org
> http://jvuletich.org/mailman/listinfo/cuis_jvuletich.org