Re: parallel garbage collection performance

Thanks very much for this information. My observations match your recommendations, insofar as I can test them. Cheers, John On Mon, Jun 25, 2012 at 11:42 PM, Simon Marlow wrote:

On 19/06/12 02:32, John Lato wrote:

...

Thanks for the suggestions.  I'll try them and report back.  Although I've since found that out of 3 not-identical systems, this problem only occurs on one.  So I may try different kernel/system libs and see where that gets me.

-qg is funny.  My interpretation from the results so far is that, when the parallel collector doesn't get stalled, it results in a big win. But when parGC does stall, it's slower than disabling parallel gc entirely.

Parallel GC is usually a win for idiomatic Haskell code, it may or may not be a good idea for things like Repa - I haven't done much analysis of those types of programs yet.  Experiment with the -A flag, e.g. -A1m is often better than the default if your processor has a large cache.

However, the parallel GC will be a problem if one or more of your cores is being used by other process(es) on the machine.  In that case, the GC synchronisation will stall and performance will go down the drain.  You can often see this on a ThreadScope profile as a big delay during GC while the other cores wait for the delayed core.  Make sure your machine is quiet and/or use one fewer cores than the total available.  It's not usually a good idea to use hyperthreaded cores either.

I'm also seeing unpredictable performance on a 32-core AMD machine with NUMA.  I'd avoid NUMA for Haskell for the time being if you can.  Indeed you get unpredictable performance on this machine even for single-threaded code, because it makes a difference on which node the pages of your executable are cached (I heard a rumour that Linux has some kind of a fix for this in the pipeline, but I don't know the details).

...

I had thought the last core parallel slowdown problem was fixed a while ago, but apparently not?

We improved matters by inserting some "yield"s into the spinlock loops.  This helped a lot, but the problem still exists.

Cheers,        Simon

...

Thanks, John

On Tue, Jun 19, 2012 at 8:49 AM, Ben Lippmeier  wrote:

...

On 19/06/2012, at 24:48 , Tyson Whitehead wrote:

...

On June 18, 2012 04:20:51 John Lato wrote:

...

Given this, can anyone suggest any likely causes of this issue, or anything I might want to look for?  Also, should I be concerned about the much larger gc_alloc_block_sync level for the slow run?  Does that indicate the allocator waiting to alloc a new block, or is it something else?  Am I on completely the wrong track?

A total shot in the dark here, but wasn't there something about really bad performance when you used all the CPUs on your machine under Linux?

Presumably very tight coupling that is causing all the threads to stall everytime the OS needs to do something or something?

This can be a problem for data parallel computations (like in Repa). In Repa all threads in the gang are supposed to run for the same time, but if one gets swapped out by the OS then the whole gang is stalled.

I tend to get best results using -N7 for an 8 core machine.

It is also important to enable thread affinity (with the -qa) flag.

For a Repa program on an 8 core machine I use +RTS -N7 -qa -qg

Ben.

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users