Encyclopaedia Index

Parallel Computing: A PHOENICS "first"

See also: How to run Parallel PHOENICS 3.1
and An earlier report on comparative performance

Contents

  1. Introduction
  2. Benefits
  3. Features
  4. Benchmarking
  5. Platforms

1. Introduction

PHOENICS was the first of the general-purpose CFD codes to be adapted for parallel-arcitecture computers, the work being conducted within EC-funded projects PASHA and EUROPORT.

* CHAM has developed a generic parallel version of the PHOENICS CFD code which can easily be ported to any message-passing MIMD machine providing a balance between achieving very good performance and retaining ease of use and code maintainability.

* Parallel PHOENICS has already been used by major industrial, research and academic organisations, e.g., STATOIL, Honda R&D and DRA. Parallel computing can significantly improve the competiveness of business in many different ways:

A parallel-PHOENICS simulation of flow around a ship's hull: the pressure field

CFD is well suited to parallel processing and parallel computers have become the main platform for the computation of complex and time-consuming problems.

To take advantage of the increased power of the latest generation of parallel computers, CHAM have developed a generic parallel version of its general purpose CFD code, PHOENICS, to run on most multi-processor computers, work-station clusters and recently PC-clusters.

2. The Benefits of Parallel CFD

  1. Accuracy: For CFD predictions to be realistic, the use of fine grids is essential requiring powerful computers with large memory resources.
  2. Speed: For CFD techniques to be used routinely in design and production cycles, fast computers are required to achieve reasonable iterative times.
  3. Scalability: As computing demands increase one can upgrade the parallel system by adding more processors, thereby enhancing rather than repeating on the initial investment.
  4. Economy: Parallel computers give better price/performance ratios than vector supercomputers.

3. Features

PHOENICS was the first general-purpose CFD code to be ported generically to massively-parallel computers. Porting is based on domain decomposition, where the computational domain is divided into sub-domains.

The computational work related to each sub-domain is then assigned to its own processor. A modified version of the PHOENICS solver (EARTH) is replicated over all available processors and runs in parallel, exchanging boundary data at appropriate times.

The PHOENICS pre- and post-processors (SATELLITE and PHOTON) run in sequential mode.

Rapid convergence of the solver is achieved due to the efficient sub-domain coupling, hence no additional sweeps are required.

The processors communicate using the standard message-passing protocols (i.e. PVM or MPI) as used on all major parallel platforms, therefore the code can run on any parallel machine that supports PVM or MPI.

Another ship-simulation picture; the grid

4. Benchmarking

Demonstration platforms with 4, 8, 16 and 32 processors have been used to benchmark parallel PHOENICS.

The efficiency depends on the size and complexity of the case (the more complex the better). Even with simple cases with increasing grid size, speed-ups of 5, 8, 14 and 23 (respectively) are achieved compared wit the same cases run on high-end serial workstations.

The simulation results matched those on serial runs, and from the user's viewpoint, the procedure is no different from running standard PHOENICS.

Benchmarking of Parallel PHOENICS has been completed on three classes of parallel platforms.

  1. PC Clusters, using fast ethernet ;
  2. Small SMP Servers and workstation clusters (SGI Octane, Sun Sparc, HP-K240, etc) can now simply run parallel PHOENICS using MPI or PVM. Here the opportunity exists for network users to optimise the use of their hardware resources by running clusters within the network as if it were a parallel device.
  3. Finally larger machines such as the SGI Origin 2000, IBM RS6000/SP, Digital 8400, CRAY T3E and many more parallel supercomputers have been benchmarked with impressive results.

Detailed performance figures are demonstrated in the simulations of the flow around a ship hull for different platforms. In the following figures is shown:

1. Reduction of elapsed time with increasing number of processors

2. Performance is almost linear with the number of processors

5. Platforms

PHOENICS has now been ported and benchmarked on a wide range of parallel machines, in fact all MPP and SMP computers that support message passing (MPI or PVM) can now run parallel PHOENICS.

Systems Benchmarked:

Parallel PHOENICS is also available on the following systems: