Difference between revisions of "Parallelism"

Revision as of 12:05, 20 April 2011

Parallel Programming in Haskell

Parallelism is about speeding up a program by using multiple processors.

In Haskell we provide two ways to achieve parallelism:

 - Concurrency, which can be used for parallelising IO.
 - Pure parallelism, which can be used to speed up pure (non-IO)
   parts of the program.

Concurrency (Control.Concurrent):

   Multiple threads of control that execute "at the same time".
   - Threads are in the IO monad
   - IO operations from multiple threads are interleaved
     non-deterministically
   - communication between threads must be explicitly programmed
   - Threads may execute on multiple processors simultaneously
   - dangers: race conditions and deadlocks

Pure Parallelism (Control.Parallel):

   Speeding up a pure computation using multiple processors. 
   - Pure parallelism has these advantages:
     - guaranteed deterministic (same result every time)
     - no race conditions or deadlocks

Rule of thumb: use Pure Parallelism if you can, Concurrency otherwise.

Starting points

Control.Parallel. The first thing to start with parallel programming in Haskell is the use of par/pseq from the parallel library.

Nested Data Parallelism. For an approach to exploiting the implicit parallelism in array programs for multiprocessors, see Data Parallel Haskell (work in progress).

Multicore GHC

Since 2004, GHC supports running programs in parallel on an SMP or multi-core machine. How to do it:

Download a recent GHC.

Compile your program using the -threaded switch.

Run the program with +RTS -N2 to use 2 threads, for example (RTS stands for runtime system; see the GHC users' guide). You should use a -N value equal to the number of CPU cores on your machine (not including Hyper-threading cores). As of GHC v6.12, you can leave off the number of cores and all available cores will be used (you still need to pass -N however, like so: +RTS -N).

Concurrent threads (forkIO) will run in parallel, and you can also use the par combinator and Strategies from the Control.Parallel.Strategies module to create parallelism.

Use +RTS -sstderr for timing stats.

To debug parallel program performance, use ThreadScope.

Alternative approaches

Nested data parallelism: a parallel programming model based on bulk data parallelism, in the form of the DPH and Repa libraries for transparently parallel arrays.
Intel Concurrent Collections for Haskell: a graph-oriented parallel programming model.

Related work

Parallel portal
Ongoing research in Parallel Haskell
The Sun project to improve http://ghcsparc.blogspot.com/ GHC performance on Sparc]
A Microsoft project to improve industrial applications of GHC parallelism.

@@ Line 1: / Line 1: @@
-== Parallel Programming in GHC ==
+== Parallel Programming in Haskell ==
+Parallelism is about speeding up a program by using multiple processors.
-This page contains notes and information about how to use parallelism in GHC to speed up pure functions in your program.
+In Haskell we provide two ways to achieve parallelism:
-You may be interested in [[GHC/Concurrency|concurrency]] instead, which would allow you to manage simultaneous IO actions.
+  - Concurrency, which can be used for parallelising IO.
+  - Pure parallelism, which can be used to speed up pure (non-IO)
+    parts of the program.
+[[Concurrency]] (Control.Concurrent):
-GHC provides multi-scale support for parallel programming, from very fine-grained, small "sparks", to coarse-grained explicit threads and locks (using concurrency), along with other models of parallel programming.
+    Multiple threads of control that execute "at the same time".
+    - Threads are in the IO monad
+    - IO operations from multiple threads are interleaved
+      non-deterministically
+    - communication between threads must be explicitly programmed
+    - Threads may execute on multiple processors simultaneously
+    - dangers: [[race conditions]] and [[deadlocks]]
+Pure Parallelism (Control.Parallel):
-The parallel programming models in GHC can be divided into the following forms:
+    Speeding up a pure computation using multiple processors.
+    - Pure parallelism has these advantages:
+      - guaranteed deterministic (same result every time)
+      - no [[race conditions]] or [[deadlocks]]
+Rule of thumb: use Pure Parallelism if you can, Concurrency otherwise.
-* Very fine grained: parallel sparks and futures, as described in the paper "[http://www.haskell.org/~simonmar/bib/multicore-ghc-09_abstract.html Runtime Support for Multicore Haskell]"
-* Nested data parallelism: a parallel programming model based on bulk data parallelism, in the form of the [http://www.haskell.org/haskellwiki/GHC/Data_Parallel_Haskell  DPH] and [http://hackage.haskell.org/package/repa Repa] libraries for transparently parallel arrays.
-* Intel [http://software.intel.com/en-us/blogs/2010/05/27/announcing-intel-concurrent-collections-for-haskell-01/ Concurrent Collections for Haskell]: a graph-oriented parallel programming model.
-The most important (as of 2010) to get to know are implicit parallelism via sparks.  If you're interested in scientific programming specifically, you may also be interested in current research on nested data parallelism in Haskell.
 === Starting points ===
@@ Line 24: / Line 34: @@
 {{GHC/Multicore}}
+=== Alternative approaches ===
+* Nested data parallelism: a parallel programming model based on bulk data parallelism, in the form of the [http://www.haskell.org/haskellwiki/GHC/Data_Parallel_Haskell  DPH] and [http://hackage.haskell.org/package/repa Repa] libraries for transparently parallel arrays.
+* Intel [http://software.intel.com/en-us/blogs/2010/05/27/announcing-intel-concurrent-collections-for-haskell-01/ Concurrent Collections for Haskell]: a graph-oriented parallel programming model.
 === Related work ===