[Haskell-cafe] Re: Odd parallel haskell observations

[Peter Wortmann]

GHC seems to have a few bottlenecks once you start to really stress-test
its I/O performance. Using a newer HEAD ghc actually gives less awful
performance:

scpmw at cslin209 ~/test $ ghc --make -O2 -threaded -rtsopts get.hs
[1 of 1] Compiling Main             ( get.hs, get.o )
Linking get ...
scpmw at cslin209 ~/test $ ./get 1 10000
1 6.25696s
scpmw at cslin209 ~/test $ ./get 1 10000
1 5.409605s
scpmw at cslin209 ~/test $ ./get 2 10000
2 3.827393s
scpmw at cslin209 ~/test $ ./get 2 10000 +RTS -N2
2 4.274985s
scpmw at cslin209 ~/test $ ./get 3 10000
3 3.692725s
scpmw at cslin209 ~/test $ ./get 3 10000 +RTS -N2
3 4.186283s
scpmw at cslin209 ~/test $ ./get 3 10000 +RTS -N3
3 4.303649s

That it still does not speed up might be the result of Haskell's
internal implementation of non-blocking I/O. If I understand the
situation correctly, all events are actually passing through one OS
thread (the I/O manager) right now. That would explain nicely why you
can't get more than single-core performance out of your program.

In case you are interested in the details, search for the new "Scalable
Event Handling for GHC" paper by Bryan O'Sullivan and Johan Tibell.

Greetings,
  Peter Wortmann