I'll need to analyse the program to see what's going on. There was a small change to the scheduler between 7.2.1 and 7.2.2 that could conceivably have made a difference in this scenario, but it was aimed at fixing a bug rather than improvement performance. Another possibility is a difference in OS scheduling behaviour between yours and Daniel Fischer's setup. In microbenchmarks like this, it's easy for a difference in OS scheduling behaviour to make a large difference in performance if it happens consistently. Cheers, Simon On 23/01/2012 12:49, John Lato wrote:

Hi Simon,

I'm not certain that your explanation matches what I observed.

All of my tests were done on a 4-core machine, executing with "+RTS -N", which should be the same as "+RTS -N4" I believe.

With 1 Haskell thread (the main thread) and 4 process threads (via pthreads), I saw a significant performance degradation compared to 5 Haskell threads (main + 4 via forkIO) and 4 process threads. As I understand your explanation, if C callbacks are scheduled according to available capabilities, there should be no difference between these situations.

I observed this with GHC-7.2.1, however Daniel Fischer reported that, with ghc-7.2.2, he observed different behavior (which matches your explanation AFAICT). Is it possible that the scheduling of callbacks into Haskell changed between those versions?

Thanks, John L.

...

From: Simon Marlow Subject: Re: Runtime performance degradation for multi-threaded C FFI callback To: Sanket Agrawal Cc: glasgow-haskell-users Message-ID:<4F1D2F4D.9050709@gmail.com> Content-Type: text/plain; charset=ISO-8859-1; format=flowed

On 21/01/2012 15:35, Sanket Agrawal wrote:

...

Hi Edward,

I was just going to get back to you about it. I did find out that the issue was indeed one GHC thread dealing with 5 C threads for callback (1:5 mapping) - so, the C threads were blocking on callback waiting for the only GHC thread to be available. I updated the code to do 1:1 mapping - 5 GHC threads for 5 C threads. That proved to be almost linearly scalable.

This is almost right, except that your callbacks are not waiting for a GHC *thread*, but what we call a "capability", which is roughly speaking "permission to execute Haskell code". The +RTS -N option chooses the number of capabilities.

I expect that with -N1, your program is spending a lot of time just switching between the different OS threads.

It's possible that we could make the runtime more flexible here. I recently made it possible to modify the number of capabilities at runtime, so it's conceivable that the runtime could automatically add capabilities if it is being called from multiple OS threads.

...

John Latos suggested the above approach two days back, but I didn't get to test the idea until now.

It doesn't seem to matter whether number of GHC threads are increased, if the mapping between GHC threads and C threads is not 1:1. I got 1:1 mapping by doing forkIO for each C thread. Is it really possible to do 7:5 mapping (that is 7 GHC threads to choose from, for 5 C threads during callback)? I can't think of a way to do it. Not that I need it. I am just curious if that is possible.

Just think of +RTS -N7 as being 7 *locks*, not 7 threads. Then it makes perfect sense to have 7 locks available for 5 threads.

Cheers, Simon