Bulat Ziganshin wrote:
Hello Simon,
Tuesday, January 10, 2006, 12:26:30 PM, you wrote:
CM> My old version is faster, because the version with makeStableName does CM> very much GC.
CM> MUT time 27.28s ( 28.91s elapsed) CM> GC time 133.98s (140.08s elapsed)
try to add infamous "+RTS -A10m" switch ;)
SM> The real problem seems to be that minor GCs are taking too long. Having SM> said that, you can usually reduce GC overhead with large -A or -H options.
it is the same problem as with scanning large IOArrays on each GC.
Actually I think it's a different problem, with the same workaround.
in this case, i think, table of stable names scanned each time and therefore program works so slow. if ghc will dynamically increase "+RTS -A" area when there are a large IOArrays/stable names table, this would help improve speed significantly.
I'm not keen on this, because its a hack and doesn't address the real cause of the problem. We're working on addressing the array issue, see this ticket I created today: http://cvs.haskell.org/trac/ghc/ticket/650
also, i want to remind my old suggestion to temporarily disable all GCs, or better to temporarily change "-A" area at the programs request. That will help programs like Joels one and my own, where large data structures (say, 5 or 20 mb) are created, used and then completely discarded
You can change the allocation area size from within a program quite easily. Write a little C function to assign to RtsFlags.GcFlags.minAllocAreaSize (#include "RtsFlags.h" first), and call it from Haskell; the next time GC runs it will allocate the larger nursery. Please try this and let me know if it works. Cheers, Simon