Regression in concurrency performance from ghc 6.6 to 6.8

While most Great Language Shootout benchmarks are running noticeably (10-20+ %) faster with GHC 6.8.1, benchmarks for concurrency seem to consistently be running slower, on a variety of architectures.

Attached are some example programs from the shootout where performance has worsened.

The thread-ring benchmark:

import Control.Monad
import Control.Concurrent
import System.Environment
 
ring = 503
 
new l i = do
  r <- newEmptyMVar
  forkIO (thread i l r)
  return r
 
thread :: Int -> MVar Int -> MVar Int -> IO ()
thread i l r = go
  where go = do
          m <- takeMVar l
          when (m == 1) (print i)
          putMVar r $! m - 1
          when (m > 0) go
 
main = do
  a <- newMVar . read . head =<< getArgs
  z <- foldM new a [2..ring]
  thread 1 z a

We can benchmark this as follows:

$ ghc-6.8.1 -O2 threadring.hs -o ghc68 -no-recomp
$ ghc-6.6.1 -O2 threadring.hs -o ghc66 -no-recomp

$ time ./ghc68 30000000 ; time ./ghc66 30000000
75
./ghc68 30000000  4.98s user 0.05s system 100% cpu 5.030 total
75
./ghc66 30000000  4.58s user 0.04s system 99% cpu 4.643 total

'chameneos' is another that has got slower.

http://shootout.alioth.debian.org/gp4/benchmark.php?test=chameneos&lang=ghc&id=2

$ time ./ghc66 7000000 ; time ./ghc68 7000000 
14000000
./ghc66 7000000  3.36s user 0.05s system 99% cpu 3.425 total
14000000
./ghc68 7000000  3.75s user 0.06s system 99% cpu 3.815 total

Trac metadata

Trac field	Value
Version	6.8.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Runtime System
Test case
Differential revisions
BlockedBy
Related
Blocking
CC	dons@galois.com, sjanssen@cse.unl.edu
Operating system	Multiple
Architecture	Multiple