GpH -- Glasgow Parallel Haskell

Activity Report May 2004

Report by: Phil Trinder

The Team

Phil Trinder, Kevin Hammond, Hans-Wolfgang Loidl, Abyd Al Zain, Jost Berthold, Xiao Yan Deng, Murray Gross, Steffen Priebe, Andre Rauber du Bois, Leonid Timochouk, Yang Yang.

Status

A complete, GHC-based implementation of the parallel Haskell extension GpH and of evaluation strategies is available. Extensions of the runtime-system and language to improve performance and support new platforms are under development.

System Evaluation and Enhancement

The first 3 items are linked by a British Council/DAAD project.

We have ported GUM to computational GRIDs, replacing the current PVM communications layer with MPICH-G2. Measurements show that standard GUM gives good performance on a single cluster and on multiple clusters with low-latency interconnect. However standard GUM gives poor performance on multiple clusters connected with high-latency interconnect and we are developing GRIDGum - an implementation with specific adaption techniques for this shared, heterogeneous and hierarchical architecture.
We are designing a generic parallel runtime environment encompassing both the Eden Eden and GpH runtime environments
We are investigating cost models for distributed computations designing an integrated runtime environment for both Eden and GpH.
In separate work GpH is being used as a vehicle for investigating scheduling on the GRID.
We are teaching parallelism to undergraduates using GpH at Heriot-Watt and Phillips Universitat Marburg.

GpH Applications

GpH is being used to parallelise the GAP mathematical library in EPSRC project (GR/R91298).

Implementations

The GUM implementation of GpH is available in two development branches.

The stable branch (GUM-4.06, based on GHC-4.06) is available for RedHat-based Linux machines: binary snapshot (see installation instructions). The stable branch is available from the GHC CVS repository via tag gum-4-06.
The unstable branch (GUM-5.02, based on GHC-5.02) is currently being tested on a Beowulf cluster. The unstable branch is available from the GHC CVS repository via tag gum-5-02-3.

Our main hardware platform are Intel-based Beowulf clusters. Work on ports to other architectures is also moving on (and available on request):

A port to a Sun-Solaris shared-memory machine exists but currently suffers from performance problems.
A port to a Mosix cluster has been built in the Metis project at Brooklyn College, with a first version available on request from Murray Gross.

3.3.2.1 Further reading:

GpH Home Page