[Haskell-cafe] Parallel weirdness

[Brandon S. Allbery KF8NH]

On Apr 19, 2008, at 11:50 , Andrew Coppin wrote:
> Bulat Ziganshin wrote:
>> there are plenty of reasons: first, -threaded make i/o overlapped
>> with calculations.
>
> Not with -N1.

Depending on how it's implemented (I not being a ghc guru), possibly  
even with -N1 as long as it's using the thread-capable runtime.   
(Note that "make -j2" is known to be optimal on single-processor  
machines, specifically because I/O tends to overlap with CPU.)

>> second, parallel version may exhibit better cpu
>> cache behavior - such as processing all data in cache before sending
>> it back to memory
>
> Again, with -N1, it is *still* only using 1 CPU core.

And again, this may well be an effect of using the thread-*capable*  
runtime.  You can't generally multiplex memory accesses in SMP, so  
you may well want to delay and batch operations to/from main memory  
as much as possible to reduce lock contention for memory access.

> Well, based on the results I've seen so far, it seems that  
> parallelism is a complete waste of time because it doesn't gain you  
> anything. And that doesn't make a lot of sense...

Easy parallelism is still an unsolved problem; naive parallelism  
generally is no better than sequential and often worse, because naive  
parallelism generally fails to account for lock / resource  
contention.  (Note that resource locking will be done by the threaded  
runtime even with only one thread, so you will see some slowdowns  
especially in I/O-related code.)  Haskell can only help you so much  
with this; you need to design your algorithms to parallel properly.

In addition, laziness can result in naive parallelism being a no-op  
because the only thing parallelized is some operation that trivially  
returns a lazy thunk that is later forced in the main thread.   
Careful strictness analysis is necessary in non-strict languages to  
make sure you are actually parallelizing what you want to.

-- 
brandon s. allbery [solaris,freebsd,perl,pugs,haskell] allbery at kf8nh.com
system administrator [openafs,heimdal,too many hats] allbery at ece.cmu.edu
electrical and computer engineering, carnegie mellon university    KF8NH