Memory leak with parMap

[igloo]

With the attached All.hs, if you run the program without any arguments (so that it uses parMap rnf in gen_blocks_list) then its memory usage keeps going up, even though at the top level it's just running the same computation 5 times in a row. Memory usage goes up to over 100M.

If you give an argument (any argument, so that it uses map instead) then memory usage is constant and only a few megs.

Happens in 6.8.2 and the HEAD.

[igloo]

Forgot to say: Compiling with

ghc -Wall -fforce-recomp All.hs --make -fasm -O2 -threaded -fglasgow-exts -funbox-strict-fields -fbang-patterns -debug

I haven't experimented with other flags.

[simonmar]

I think sparks are currently treated as roots by the GC, which is wrong. I won't get around to fixing this in 6.8.3, but I'll do it in the new GC code in 6.10.

[simonmar]

Now fixed, and I've added the program to our test suite.

[simonmar]

unfixed. Sparks really should be treated as GC roots, because many uses of par rely on this behaviour (including most of the strategies library).

I don't have any ideas for how to fix this yet.

[simonmar]

The  new parallel package on Hackage fixes the space leak in this program (change rnf to rdeepseq to compile it). I'm leaving the ticket open and moving it onto the 6.14.1 branch, because we still need to change the GC policy so that sparks are not treated as GC roots.

[simonmar]

Fixed:

Tue May 25 08:04:35 PDT 2010  Simon Marlow <marlowsd@gmail.com>
  * Make sparks into weak pointers (#2185)
  The new strategies library (parallel-2.0+, preferably 2.2+) is now
  required for parallel programming, otherwise parallelism will be lost.