Hi Karel,

could you try adding `-j8` to `SRC_HC_OPTS` for the build flavor you're using in `mk/build.mk`, and running `gmake -j8` instead of `gmake -j64`. A graph like the one you attached will likely look even worse, but the walltime of your build should hopefully be improved.

The build system seems to currently rely entirely on `make` for parallelism. It doesn't exploit ghc's own parallel `--make` at all, unless you explictly add `-jn` to SRC_HC_OPTS, with n>1 (which also sets the number of capabilities for the runtime system, so also adding `+RTS -Nn` is not needed). 

Case study: One of the first things the build system does is build ghc-cabal and Cabal using the stage 0 compiler, through a single invocation of `ghc --make`. All the later make targets depend on that step to complete first. Because `ghc --make` is not instructed to build in parallel, using `make -j1` or `make -j100000` doesn't make any difference (for that step). I think your graph shows that there are many of more of such bottlenecks.

You would have to find out empirically how to best divide your number of threads (32) between `make` and `ghc --make`. From reading this comment by Simon in #9221 I understand it's better not to call `ghc --make -jn` with `n` higher than the number of physical cores of your machine (8 in your case). Once you get some better parallelism, other flags like `-A` might also have an effect on walltime (see that ticket).

-Thomas

On Sat, Mar 7, 2015 at 11:49 AM, Karel Gardas <karel.gardas@centrum.cz> wrote:

Folks,

first of all, I remember someone already mentioned issue with decreased parallelism of the GHC build recently somewhere but I cann't find it now. Sorry, for that since otherwise I would use this thread if it was on this mailing list.

Anyway, while working on SPARC NCG I'm using T2000 which provides 32 threads/8 core UltraSPARC T1 CPU. The property of this machine is that it's really slow on single-threaded work. To squeeze some perf from it man really needs to push 32 threads of work on it. Now, it really hurts my nerves to see it's lazy building/running just one or two ghc processes. To verify the fact I've created simple script to collect number of ghc processes over time and putting this to graph. The result is in the attached picture. The graph is result of running:

gmake -j64

anyway, the average number of running ghc processes is 4.4 and the median value is 2. IMHO such low number not only hurts build times on something like CMT SPARC machine, but also on let say a cluster of ARM machines using NFS and also on common engineering workstations which provide these days (IMHO!) around 8-16 cores (and double the threads number).

My naive idea(s) for fixing this issue is (I'm assuming no Haskell file imports unused imports here, but perhaps this may be also investigated):

1) provide explicit dependencies which guides make to build in more optimal way

2) hack GHC's make depend to kind of compute explicit dependencies from (1) in an optimal way automatically

3) someone already mentioned using shake for building ghc. I don't know shake but perhaps this is the right direction?

4) hack GHC to compile needed hi file directly in its memory if hi file is not (yet!) available (issue how to get compiling options right here). Also I don't know hi file semantics yet so bear with me on this.


Is there anything else which may be done to fix that issue? Is someone already working on some of those? (I mean those reasonable from the list)?

Thanks!
Karel


_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs