[Haskell-cafe] Re: A suggestion for the next high profile Haskell project

[Alexey Rodriguez Yakushev]

Andy,

The GHC head can currently build against PAPI[1], a library for  
gathering CPU statistics. At the moment you can only gather such  
statistics for AMD Opteron but it shouldn't be difficult to port it  
to other CPUs after a bit of browsing around the PAPI docs.  
Installing PAPI requires installing a linux kernel driver though, so  
it is not for the faint hearted.

We have used this library to find bottlenecks in the current code  
generation and we have implemented ways of correcting them, so expect  
some good news about this in the future.

I should get around to start a wiki page about using PAPI these days,  
but meanwhile feel free to contact me if you need further information  
or help.

Cheers,

Alexey

[1] http://icl.cs.utk.edu/papi/

On Dec 18, 2006, at 22:35, Andy Georges wrote:

> Hi,
>
>>> I have to dispute this Bulat's characterisation here. We can  
>>> solve lots
>>> of nice problems and have high performance *right now*. Particularly
>>> concurrency problems, and ones involving streams of bytestrings.
>>> No need to leave the safety of GHC either, nor resort to low  
>>> level evil
>>> code.
>>
>> let's go further in this long-term discussion. i've read Shootout  
>> problems
>> and concluded that there are only 2 tasks which speed is dependent on
>> code-generation abilities of compiler, all other tasks are  
>> dependent on
>> speed of used libraries. just for example - in one test TCL was  
>> fastest
>> language. why? because this test contained almost nothing but 1000  
>> calls to
>> the regex engine with very large strings and TCL regex engine was  
>> fastest
>
> Maybe it would not be a bad idea to check the number of cache  
> misses, branch mispredictions etc. per instruction executed for the  
> shootout apps, in different languages, and of course, in haskell,  
> on the platforms ghc targets. Do you think it might potentially be  
> interesting to the GHC developing community to have such an  
> overview? It might show potential bottlenecks, I think.
>
> -- Andy
>
>
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe