GHC/Data Parallel Haskell

From HaskellWiki
Revision as of 16:19, 2 December 2008 by Malcolm (talk | contribs) (module DotP should export dotp_wrapper)

Jump to: navigation, search

Data Parallel Haskell

Data Parallel Haskell is the codename for an extension to the Glasgow Haskell Compiler and its libraries to support nested data parallelism with a focus to utilise multicore CPUs. Nested data parallelism extends the programming model of flat data parallelism, as known from parallel Fortran dialects, to irregular parallel computations (such as divide-and-conquer algorithms) and irregular data structures (such as sparse matrices and tree structures). An introduction to nested data parallelism in Haskell, including some examples, can be found in the paper Nepal – Nested Data-Parallelism in Haskell.

Project status

A first technology preview of Data Parallel Haskell (DPH) is included in the 6.10.1 release of GHC. All major components of DPH are implemented, including code vectorisation and parallel execution on multicore systems. However, the implementation has many limitations and probably also many bugs. Major limitations include the inability to mix vectorised and non-vectorised code in a single Haskell module, the need to use a feature-deprived, special-purpose Prelude in vectorised code, and a lack of optimisations (leading to poor performance).

The purpose of this technology preview is twofold. Firstly, it gives interested early adopters the opportunity to see where the project is headed and enables them to experiment with simple DPH programs. Secondly, we hope to get user feedback that helps us to guide the project and prioritise those features that our users are most interested in.

Disclaimer: Data Parallel Haskell is very much work in progress. Some components are already usable, and we explain here how to use them. However, please be aware that APIs are still in flux and functionality may change during development.

Where to get it

Currently, we recommend to use the implementation in GHC 6.10.1. It is available in source and binary form for many architectures. (Please use the version in the HEAD repository of GHC only if you are a GHC developer or a very experienced GHC user and if you know the current status of the DPH code – intermediate versions may well be broken while we implement major changes.)


From a user's point of view, Data Parallel Haskell adds a new data type to Haskell –namely, parallel arrays– as well as operations on parallel arrays. Syntactically, parallel arrays are like lists, only that instead of square brackets [ and ], parallel arrays use square brackets with a colon [: and :]. In particular, [:e:] is the type of parallel arrays with elements of type e; the expression [:x, y, z:] denotes a three element parallel array with elements x, y, and z; and [:x + 1 | x <- xs:] represents a simple array comprehension. More sophisticated array comprehensions (including the equivalent of parallel list comprehensions) as well as enumerations and pattern matching proceed in an analog manner. Moreover, the array library of DPH defines analogs of most list operations from the Haskell prelude and the standard List library (e.g., we have lengthP, sumP, mapP, and so on).

The two main differences between lists and parallel arrays are that (1) parallel arrays are a strict data structure and (2) that they are not inductively defined. Parallel arrays are strict in that by demanding a single element, all elements of an array are demanded. Hence, all elements of a parallel array might be evaluated in parallel. To facilitate such parallel evaluation, operations on parallel arrays should treat arrays as aggregate structures that are manipulated in their entirety (instead of the inductive, element-wise processing that is the foundation of all Haskell list functions.)

As a consequence, parallel arrays are always finite, and standard functions that yield infinite lists, such as enumFrom and repeat, have no corresponding array operation. Moreover, parallel arrays only have an undirected fold function foldP that requires an associative function as an argument – such a fold function has a parallel step complexity of O(log n) for arrays of length n. Parallel arrays also come with some aggregate operations that absent from the standard list library, such as permuteP.

A simple example

As a simple example of a DPH program, consider the following code that computes the dot product of two vectors given as parallel arrays:

dotp :: Num a => [:a:] -> [:a:] -> a
dotp xs ys = sumP [:x * y | x <- xs | y <- ys:]

This code uses an array variant of parallel list comprehensions, which could alternatively be written as [:x * y | (x, y) <- zipP xs ys:], but should otherwise be self-explanatory to any Haskell programmer.

Running DPH programs

Unfortunately, we cannot use the above implementation of dotp directly in the current preliminary implementation of DPH. In the following, we will discuss how the code needs to be modified and how it needs to be compiled and run for parallel execution. GHC applies an elaborate transformation to DPH code called vectorisation that turns nested into flat data parallelism. This transformation is only useful for code that is executed in parallel (i.e., code that manipulates parallel arrays), but there it raises the level of expressiveness dramatically.

No type classes

Unfortunately, vectorisation does not properly handle type classes at the moment. Hence, we currently need to avoid overloaded operations in parallel code. To account for that limitation, we specialise dotp on doubles.

dotp_double :: [:Double:] -> [:Double:] -> Double
dotp_double xs ys = sumP [:x * y | x <- xs | y <- ys:]

Special Prelude

As the current implementation of vectorisation cannot handle some language constructs, we cannot use it to vectorise those parts of the standard Prelude that might be used in parallel code (such as arithmetic operations). Instead, DPH comes with its own (rather limited) Prelude in Data.Array.Parallel.Prelude plus three extra modules to support one numeric type each Data.Array.Parallel.Prelude.Double, Data.Array.Parallel.Prelude.Int, and Data.Array.Parallel.Prelude.Word8. These three modules support the same functions (on different types) and if a program needs to use more than one, they need to be imported qualified (as we cannot use type classes in vectorised code in the current version). If your code needs any other numeric types or functions that are not implemented in the these Prelude modules, you currently need to implement and vectorise that functionality yourself.

To compile dotp_double, we add the following three import statements:

import qualified Prelude
import Data.Array.Parallel.Prelude
import Data.Array.Parallel.Prelude.Double

Impedance matching

Special care is needed at the interface between vectorised and non-vectorised code. Currently, only simple types can be passed between these different kinds of code. In particular, parallel arrays cannot be passed. Instead, we can pass flat arrays of type PArray. This type is exported by our special-purpose Prelude together with a conversion function fromPArrayP (which is specific to the element type due to the lack of type classes in vectorised code).

Using this conversion function, we define a wrapper function for dotp_double that we export and use from non-vectorised code.

dotp_wrapper :: PArray Double -> PArray Double -> Double
{-# NOINLINE dotp_wrapper #-}
dotp_wrapper v w = dotp_double (fromPArrayP v) (fromPArrayP w)

It is important to mark this function as NOINLINE as we don't want it to be inlined into non-vectorised code.

Compiling vectorised code

The definition of dotp_double requires two language extensions, namely PArr to enable the syntax of parallel arrays and ParallelListComp for the parallel comprehension. Furthermore, we need to explicitly tell GHC which modules we want to vectorise.

Currently, GHC either vectorises all code in a module or none. This can be inconvenient as some parts of a program cannot be vectorised – for example, code in the IO monad (the radical re-ordering of computations performed by the vectorisation transformation is only valid for pure code). As a consequence, the programmer currently needs to partition vectorised and non-vectorised code carefully over different modules.

The compiler option to enable vectorisation is -fvectorise. Overall, we get the following complete module definition for the dot-product code:

{-# LANGUAGE PArr, ParallelListComp #-}
{-# OPTIONS -fvectorise #-}

module DotP (dotp_double,dotp_wrapper)

import qualified Prelude
import Data.Array.Parallel.Prelude
import Data.Array.Parallel.Prelude.Double

dotp_double :: [:Double:] -> [:Double:] -> Double
dotp_double xs ys = sumP [:x * y | x <- xs | y <- ys:]

dotp_wrapper :: PArray Double -> PArray Double -> Double
{-# NOINLINE dotp_wrapper #-}
dotp_wrapper v w = dotp_double (fromPArrayP v) (fromPArrayP w)

Assuming the module is in a file DotP.hs, we compile it was follows:

ghc -c -Odph -fcpr-off -fdph-seq DotP.hs

The option -Odph enables a predefined set of GHC optimisation options that is geared at optimising DPH code. Moreover, we use -fcpr-off as GHC's CPR phase doesn't play nice with type families at the moment, which in turn are heavily used in the DPH library. We shall discuss -fdph-seq below.

Using vectorised code

Finally, we need a wrapper module that calls the vectorised code, but is itself not vectorised. In this simple example, this is just the Main module that generates two random vectors and computes their dot product:

import System.Random (newStdGen)
import Data.Array.Parallel.PArray (PArray, randomRs)

import DotP (dotp_wrapper)  -- import vectorised code

main :: IO ()
  = do 
      gen1 <- newStdGen
      gen2 <- newStdGen
      let v = randomRs n range gen1
          w = randomRs n range gen2
      print $ dotp_wrapper v w   -- invoke vectorised code and print the result
    n     = 10000        -- vector length
    range = (-100, 100)  -- range of vector elements

We compile this module with

ghc -c -O -fdph-seq Main.hs

and finally link with

ghc -o dotp -fdph-seq -threaded DotP.o Main.o

NOTE: The code as presented is unsuitable for benchmarking as we wouldn't want to measure the purely sequential random number generation (that dominates this simple program). For benchmarking, we would want to guarantee that the generated vectors are fully evaluated before taking the time. The module Data.Array.Parallel.PArray exports the function nf for this purpose.

Parallel execution

The array library of DPH comes in two flavours: dph-seq and dph-par. The former supports the whole DPH stack, but only executes on a single core. In contrast, dph-par implements multi-threaded code.

In the above compiler invocations, we used the option -fdph-seq to select the dph-seq flavour. We can as well compile with -fdph-par to generate multi-threaded code. By invoking ./dotp +RTS -N2, we use two OS threads to execute the program. A beautiful property of DPH is that the number of threads used to execute a program only affects its performance, but not the result. So, it is fine to do all debugging concerning correctness with dph-seq and to switch to dph-par only for performance debugging.

Data Parallel Haskell –and more generally, GHC's multi-threading support– currently only aims at multicore processors or uniform memory access (UMA) multi-processors. Performance on non-uniform memory access (NUMA) machines is generally bad as GHC's runtime makes no effort at optimising placement. Some people have reported that the parallel garbage collector (as included in GHC 6.10.1) should not be used with parallel programs; i.e., it is advisable to start parallel programs with my_program +RTS -N2 -g1 to run on two cores (and different arguments to -N for other core counts). This problem has been addressed in the development version of GHC.

Further examples

Further examples are available in the examples directory of the package dph source. In addition to code using vectorisation (as described above), these examples also contain code that directly targets the two array libraries contained in -package dph-seq and -package dph-par, respectively. For more complex programs, targeting the DPH array libraries directly can lead to much faster code than using vectorisation, as GHC currently doesn't optimise vectorised code very well. However, code targeting the DPH libraries directly can only use flat data parallelism.

The interfaces of the various components of the DPH library are specified in GHC's hierarchical libraries documentation.

Designing parallel programs

Data Parallel Haskell is a high-level language to code parallel algorithms. Like plain Haskell, DPH frees the programmer from many low-level operational considerations (such as thread creation, thread synchronisation, critical sections, and deadlock avoidance). Nevertheless, the full responsibility for parallel algorithm design and many performance considerations (such as when does a computation have sufficient parallelism to make it worthwhile to exploit that parallelism) are still with the programmer.

DPH encourages a data-driven style of parallel programming and, in good Haskell tradition, puts the choice of data types first. Specifically, the choice between using lists or parallel arrays for a data structure determines whether operations on the structure will be executed sequentially or in parallel. In addition to suitably combining standard lists and parallel arrays, it is often also useful to embed parallel arrays in a user-defined inductive structure, such as the following definition of parallel rose trees:

data RTree a = RNode [:RTree a:]

The tree is inductively defined; hence, tree traversals will proceed sequentially, level by level. However, the children of each node are held in parallel arrays, and hence, may be traversed in parallel. This structure is, for example, useful in parallel adaptive algorithms based on a hierarchical decomposition, such as the Barnes-Hut algorithm for solving the N-body problem as discussed in more detail in the paper Harnessing the Multicores: Nested Data Parallelism in Haskell.

For a general introduction to nested data parallelism and its cost model, see Blelloch's Programming Parallel Algorithms.

Further reading and information on the implementation

DPH has two major components: (1) the vectorisation transformation and (2) the generic DPH library for flat parallel arrays. The vectorisation transformation turns nested into flat data-parallelism and is described in detail in the paper Harnessing the Multicores: Nested Data Parallelism in Haskell. The generic array library maps flat data-parallelism to GHC's multi-threaded multicore support and is described in the paper Data Parallel Haskell: a status report. The same topics are also covered in the slides for the two talks Nested data parallelism in Haskell and Compiling nested data parallelism by program transformation.

For further reading, consult this collection of background papers, and pointers to other people's work. If you are really curious and like to know implementation details and the internals of the Data Parallel Haskell project, much of it is described on the GHC developer wiki on the pages covering data parallelism and type families.


Please file bug reports at GHC's bug tracker. Moreover, comments and suggestions are very welcome. Please post them to the GHC user's mailing list, or contact the DPH developers directly: