add chan, a simple Chan performance benchmark

smp/Makefile

@@ -8,7 +8,8 @@ SUBDIRS = \
 	threads001 \
 	threads003 \
 	threads006 \
-	threads007
+	threads007 \
+	chan
 
 # later:
 #  stm001

smp/chan/Makefile

+TOP = ../..
+include $(TOP)/mk/boilerplate.mk
+
+FAST_OPTS = 100000
+NORM_OPTS = 5000000
+SLOW_OPTS = 50000000
+
+include $(TOP)/mk/target.mk

smp/chan/chan.hs

+-- benchmarks communication on Chan 
+-- 
+-- This is a synthetic benchmark that is sensitive to scheduler
+-- behaviour.  In GHC 6.12 and earlier we triggered a context switch
+-- shortly after waking up a thread, whereas in 6.14 and later we
+-- stopped doing that.  This benchmark performs worse with 6.14
+-- becauuse not doing the context switch allows a lot of data to build
+-- up in the Chan, making GC expensive.
+
+import Control.Concurrent
+import Control.Concurrent.Chan
+import System.Environment
+import Control.Monad
+
+main = do
+  [n] <- fmap (fmap read) getArgs
+  c <- newChan
+  m <- newEmptyMVar
+  a <- forkIO $ forM_ [1..n] $ \i -> writeChan c i
+  b <- forkIO $ do forM_ [1..n] $ \i -> readChan c; putMVar m ()
+  takeMVar m