On 11/06/2009 05:40, Evan Klitzke wrote:
I've written a multi-threaded Haskell program that I'm trying to debug. Basically what's happening is the program runs for a while, and then at some point one of the threads goes crazy and spins the CPU while allocating memory; this proceeds until the system runs out of available memory. I can't fix this without figuring out what code is being executed in the loop (or at least which thread is misbehaving, which would narrow things down a lot).
I was hopeful that I could compile the program with profiling support and then use +RTS -M100M along with some of the RTS profiling options and get profiling information on CPU and memory usage at the time that my program runs out of memory. The thinking here is that nearly all of the CPU time and heap space will be from the misbehaving thread, so at that point I could do more investigation into exactly what is happening. Unfortunately, this doesn't seem to work; whenever the program terminates due to running out of heap space, the generated .prof file is empty.
We fixed this recently (GHC 6.10.2): http://hackage.haskell.org/trac/ghc/ticket/2592 In 6.12.1 you'll be able to use ThreadScope, our parallel profiling tool. You could try it right now if you're brave enough to compile GHC (it needs GHC 6.11). The ThreadScope code is here: http://code.haskell.org/ThreadScope/ and shortly the Haskell Symposium paper about it will be available (we're just making the final corrections now). Cheers, Simon