Haskell for multicores

GHC Haskell comes with a large set of libraries and tools for building programs that exploit multicore architectures.

This site attempts to document all our available information on exploiting such hardware with Haskell.

Throughout, we focus on exploiting shared-memory SMP systems, with aim of lowering absolute wall clock times. The machines we target are typical 2x to 32x desktop multicore machine, on which vanilla GHC will run.

Introduction

To get an idea of what we aim to do -- reduce running times by exploiting more cores -- here's a naive "hello, world" of parallel programs: parallel, naive fib. It simply tells us whether or not the SMP runtime is working:

import Control.Parallel
import Control.Monad
import Text.Printf

cutoff = 35

fib' :: Int -> Integer
fib' 0 = 0
fib' 1 = 1
fib' n = fib' (n-1) + fib' (n-2)

fib :: Int -> Integer
fib n | n < cutoff = fib' n
      | otherwise  = r `par` (l `pseq` l + r)
 where
    l = fib (n-1)
    r = fib (n-2)

main = forM_ [0..45] $ \i ->
            printf "n=%d => %d\n" i (fib i)

We compile it with the `-threaded` flag:

   $ ghc -O2 -threaded --make fib.hs
   [1 of 1] Compiling Main             ( fib.hs, fib.o )
   Linking fib ...

And run it with:

   +RTS -Nx

where 'x' is the number of cores you have (or a slightly higher value). Here, on a quad core linux system:

   ./fib +RTS -N4  76.81s user 0.75s system 351% cpu 22.059 total

So we were able to use 3.5/4 of the available cpu time. And this is typical, most problems aren't easily scalable, and we must trade off work on more cores, for more overhead with communication.

Examples

• Riemann's zeta function

Further reading

• GHC's multiprocessor guide
• runtime options to enable SMP parallelism
• API documentation for paralell strategies
• Real World Haskell: Concurrent and Parallel Programming
• Blog posts about parallelism

Thread primitives

Control.Concurrent Control.Concurrent

For explicit concurrency and/or parallelism, Haskell implementations have a light-weight thread system that schedules logical threads on the available operating system threads. These light and cheap threads can be created with forkIO. (We won't discuss full OS threads which are created via forkOS, as they have significantly higher overhead and are only useful in a few situations like in FFIs.)

    forkIO :: IO () -> IO ThreadId

Lets take a simple Haskell application that hashes two files and prints the result:

import Data.Digest.Pure.MD5 (md5)
import qualified Data.ByteString.Lazy as L
import System.Environment (getArgs)

main = do
    [fileA, fileB] <- getArgs
    hashAndPrint fileA
    hashAndPrint fileB

hashAndPrint f = L.readFile f >>= return . md5 >>= \h -> putStrLn (f ++ ": " ++ show h)

This is a straight forward solution that hashs the files one at a time printing the resulting hash to the screen. What if we wanted to use more than one processor to hash the files in parallel?

One solution is to start a new thread, hash in parallel, and print the answers as they are computed:

import Control.Concurrent (forkIO)
import Data.Digest.Pure.MD5 (md5)
import qualified Data.ByteString.Lazy as L
import System.Environment (getArgs)

main = do
    [fileA,fileB] <- getArgs
    forkIO $ hashAndPrint fileA
    hashAndPrint fileB

hashAndPrint f = L.readFile f >>= return . md5 >>= \h -> putStrLn (f ++ ": " ++ show h)

Now we have a rough program with great performance boost - which is expected given the trivially parallel computation.

But wait! You say there is a bug? Two, actually. One is that if the main thread is finished hashing fileB first, the program will exit before the child thread is done with fileA. The second is a potential for garbled output due to two threads writing to stdout. Both these problems can be solved using some inter-thread communication - we'll pick this example up in the MVar section.

Further reading

• A concurrent port scanner
• Research papers on concurrency in Haskell
• Research papers on parallel Haskell

Synchronisation with locks

Control.Concurrent.MVar

Locking mutable variables (MVars) can be used to great effect not only for communicating values (such as the resulting string for a single function to print) but it is also common for programmers to use their locking features as a signaling mechanism.

MVars are a polymorphic mutable variables that might or might not contain a value at any given time. Common functions include:

    newMVar :: a -> IO (MVar a)
    newEmptyMVar :: IO (MVar a)
    takeMVar :: MVar a -> IO a
    putMVar :: MVar a -> a -> IO ()
    isEmptyMVar :: MVar a -> IO Bool

While they are fairly self-explanitory it should be noted that takeMVar will block until the MVar is non-empty and putMVar will block until the current MVar is empty. Taking an MVar will leave the MVar empty when returning the value.

In the forkIO example we developed a program to hash two files in parallel and ended with a couple small bugs because the program terminated prematurely (the main thread would exit when done). A second issue was that threads can conflict with each others use of stdout.

Lets now generalize the example to operate on any number of files, block until the hashing is complete, and print all the results from just one thread so no stdout garbling occurs.

    {-# LANGUAGE BangPatterns #-}
    import Data.Digest.Pure.MD5
    import qualified Data.ByteString.Lazy as L
    import System.Environment
    import Control.Concurrent
    import Control.Monad (replicateM_)

    main = do
        files <- getArgs
        str <- newEmptyMVar
        mapM_ (forkIO . hashAndPrint str) files
        printNrResults (length files) str

    printNrResults i var = replicateM_ i (takeMVar var >>= putStrLn)

    hashAndPrint str f = do
        bs <- L.readFile f
        let !h = show $ md5 bs
        putMVar str (f ++ ": " ++ h)

We define a new variable, str, as an empty MVar. After the hashing, the result is reported with putMVar - remember this function blocks when the MVar is already full so no hashes are dropped on account of the mutable memory. Similarly, printNrResults uses the takeMVar function which will block until the MVar is full - or once the next file is done being hashed in this case.

Note how the value is evaluated before the putMVar call. If the argument is an unevaluated thunk then printNrResults will have to evaluate the thunks before it prints the result and our efforts would have been worthless.

Knowing the str MVar will be filled 'length files' times we can let the main thread exit after printing the given number of results, thus terminating the program.

$ ghc exMVar.hs -o exMVar-threaded --make -O2 -threaded
$ time ./exMVar-threaded +RTS -N2 -RTS 2GB 2GB 2GB 2GB 
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb

  real    0m40.524s


$ time ./exMVar-threaded +RTS -N1 -RTS 2GB 2GB 2GB 2GB
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb
  2GB: b8f1f1faa6dda5426abffb3a7811c1fb

  real    1m8.170s

Further reading

Message passing channels

Control.Concurrent.Chan

For streaming data it is hard to beat the performance of channels. After declaring a channel (newChan), you can pipe data between threads (writeChan, readChan) and tee data to separate readers (dupChan). The flexibility of channels makes them useful for a wide range of communications.

Continuing with our hashing example, lets say the names of the files needing hashed are coming available, or need streaming for other reasons. We can fork a set of worker threads and feed them the filenames through a channel. For consistancy, the program has also been modified to communicate the result from worker to printer via a channel.

{-# LANGUAGE BangPatterns #-}
import Data.Digest.Pure.MD5
import qualified Data.ByteString.Lazy as L
import System.Environment
import Control.Concurrent
import Control.Concurrent.Chan
import Control.Monad (forever, forM_, replicateM_)

nrWorkers = 2

main = do
    files <- getArgs
    str <- newChan
    fileChan <- newChan
    forM_ [1..nrWorkers] (\_ -> forkIO $ worker str fileChan)
    forM_ files (writeChan fileChan)
    printNrResults (length files) str

printNrResults i var = replicateM_ i (readChan var >>= putStrLn)

worker :: Chan String -> Chan String -> IO ()
worker str fileChan = forever (readChan fileChan >>= hashAndPrint str)

hashAndPrint str f = do
        bs <- L.readFile f
        let !h = show $ md5 bs
        writeChan str (f ++ ": " ++ h)

Notice that this has advantages: results are available incrementally, and the performance has improved with the number of parallel hash operations matching the number of cores.

Examples

• Implementing a chat server
• Two reader threads

Further reading

Lock-free synchronisation

Software Transactional Memory

• STM

Todo

Further reading

• Research papers on transactional memory

Asynchronous messages

Control.Exception:asynchronous

• Async exceptions

Todo

Examples

• Graceful exit

Further reading

Parallelism strategies

Control.Parallel

• Parallel, pure strategies

Todo

Further reading

• Parallel binary trees benchmark

Data parallel arrays

Data Parallel Arrays

Todo

Further reading

Foreign languages calls and concurrency

Non-blocking foreign calls in concurrent threads.

Profiling and measurement

   +RTS -sstderr

Further reading

Todo