[Haskell-cafe] Haskell Speed Myth

[Thomas Davie]

On 24 Aug 2008, at 01:26, Brandon S. Allbery KF8NH wrote:

> On 2008 Aug 23, at 18:34, Krzysztof Skrzętnicki wrote:
>> Recently I wrote computation intensive program that could easily
>> utilize both cores. However, there was overhead just from compiling
>> with -threaded and making some forkIO's. Still, the overhead was not
>> larger than 50% and with 4 cores I would probably still get the
>> results faster - I didn't experience an order of magnitude slowdown.
>> Perhaps it's the issue with OS X.
>
>
> All that's needed for multicore to be a *lot* slower is doing it  
> wrong.  Make sure you're forcing the right things in the right  
> places, or you could quietly be building up thunks on both cores  
> that will cause lots of cross-core signaling or locking.  And, well,  
> make sure the generated code isn't stupid.  Quite possibly the PPC  
> code is an order of magnitude worse than the better-tested Intel code.

Except that the test was running on a Core2Duo, and it runs very fast  
when ghc does the threading on one core.  My personal guess is that to  
do it properly threaded requires *lots* of kernel boundary crosses to  
do the locking etc on OS X (being a nearly-micro-kernel).  The test  
program was almost 100% made up of thread locking code.

Bob