SGI Power Challenge Research Articles

Renumbering unstructured grids to improve the performance of codes on hierarchical memory machines

The performance of unstructured grid codes on workstations and distributed memory parallel computers is substantially affected by the efficiency of the memory hierarchy. This efficiency essentially depends on the order of computation and numbering of the grid. Most grid generators do not take into account the effect of the memory hierarchy when producing grids so application programmers must renumber grids to improve the performance of their codes. To design a good renumbering scheme a detailed runtime analysis of the data movement in an application code is needed. Thus, a memory hierarchy simulator has been developed to analyse the effect of existing renumbering schemes such as bandwidth reduction, the Greedy method, colouring, random numbering and the original numbering produced by the grid generator. The renumbering is applied to either vertices, edges, faces or cells and two algorithms are proposed to consistently renumber the other entities used in the solver. The simulated and actual timings show that bandwidth reduction and Greedy methods give the best performance on IBM RS/6000, SGI Indy, SGI Indigo and SGI Power Challenge machines for three-dimensional Poissons's, Maxwell's and the Euler equations solvers. The improvement in performance is over a factor of two for applications with large grids and a high ratio of memory-accesses to computation. This factor is even higher for memory hierarchies with small caches.

Non-linear Optimization on a Parallel Intel i860 RISC Based Architecture

A description of a parallel implementation of the Genetic Algorithm for non-linear optimization is presented. The system is implemented with four Intel i860 RISC processors using INMOS transputers for communications. The performance is compared to the Cray Y-MP, Cray J916, an SGI Power Challenge L R8000 75 MHZ workstation, an SGI Challenge L R4400 100 MHZ work-station, an SGI Indy SC R4400 200 MHZ workstation, and both parastation and Paramid parallel I860 machines, each with one to four participating I860s. Comparisons were made for neural network optimization of a limited dependent variable problem and a chaotic time series. A complex rational expectations model was also estimated. The I860 based machines were able to achieve performances comparable to the supercomputers.

A parallel molecular dynamics simulation code for dialkyl cationic surfactants

We have developed a new simulation code, COMFORT, for the study of assemblies of flexible surfactant molecules, structured for parallel execution and specialised to surfactants with dialkyl chain geometry. The approach is a hybrid domain-decomposition and systolic-loop algorithm which is suitable for systems composed of long chain molecules and with tens of thousands of atoms in total. The algorithm uses a modified Ewald technique for two dimensionally periodic systems which has been successfully parallelized. The code was designed to be highly portable between machines of different architectures. The code has been tested on a number of platforms including the Intel iPSC/860, the IBM SP1, the CRAY T3D, a SGI Power Challenge and a number of workstation clusters. We demonstrate that scalable parallel computing technology, combined with appropriate software, can provide a commercially viable simulation system for use in the exploration and development of surfactant assemblies.

Parallel benchmarks of turbulence in complex geometries

In this paper we present benchmark results from the parallel implementation of the three-dimensional Navier-Stokes solver Prism on different parallel platforms of current interest: IBM SP2 (all three types of processors), SGI Power Challenge XL and Cray C90. The numerical method is based on mixed spectral element-Fourier expansions in ( x − y) and z-directions, respectively. Each one (or a group) of the Fourier modes can be computed on a separate processor as the linear contributions in Navier-Stokes equations (Helmholtz solvers) are completely uncoupled. Coupling is obtained via the non-linear contributions (advection terms) and requires a global transpose of the data and one-dimensional multiple-point FFTs. We first analyze the computational complexity of Prism identifying basic computational kernels and communication bottlenecks. These kernels are benchmarked separately on several computer architectures. More specifically, we obtain timings for BLAS routines on all aforementioned processors but also on the SP1, Cray J90, the Intel Paragon, and the DEC AlphaServer 8400 5/300. A two-dimensional version of Prism is also benchmarked on most processors testing both direct and iterative methods of solution. Next, we define two three-dimensional benchmark flow problems in prototype complex geometries, and measure parallel scalability and performance using different message-passing libraries. Special emphasis is placed on runtime data processing, e.g. turbulence statistics computed during the simulation. Our results provide a measure of MFlop s sustained performance of individual processors, and indicate the limitations of the current communications software for typical problems in computational mechanics based on spectral or finite element discretizations.

Galaxies Collide On the I-Way: an Example of Heterogeneous Wide-Area Collaborative Supercomputing

As a part of the Supercomputing '95 High Performance Computing Challenge, the authors have used the re sources of three of the National Science Foundation (NSF) supercomputing centers as a single, distributed, heteroge neous metacomputer to carry out the largest simulation of colliding galaxies yet attempted. The metacomputer con sisted of the following machines: TMC CM-5, Cray C90/T3D, IBM SP-2, and SGI Power Challenge. This paper describes the scalable parallel numerical algorithms used to carry out the hybrid N-body/gas dynamical simulation, as well as the inter-MPP communication system the authors developed to connect them. The system, called sclib, is simple yet flexible, and consists of a network of communicating objects, or agents. Each communication agent is a process that resides on the front-end processor of a parallel supercomputer and brokers the interactions between the various components of the distributed com putation. The design is independent of host architecture and execution model yet takes advantage of scheduling resources local to the compute servers. As such, it serves as a useful prototype for future research in distributed, heterogeneous, high-performance computing.

Radio Synthesis Imaging: a Grand Challenge HPCC Project

This paper describes the technique of radio synthesis imaging and its computational requirements. The radio synthesis imaging project at the National Center for Su percomputing Applications (NCSA) is addressing these requirements in four areas: real-time transfer of data from remotely located telescopes to NCSA and archiving of these raw data; production of software for calibration, image formation, and image improvement on NCSA par allel processor supercomputers; development of a system for archiving of final processed images in a digital library; and production of sophisticated tools for image visualiza tion and analysis. At Supercomputing '95 in San Diego the authors demonstrated an observation of the center of the Milky Way galaxy, data processing on the remote NCSA SGI Power Challenge Array, accessing complementary data in the NCSA Digital Image Library, and visualization of the galactic center data sets.

A parallel computing approach to creating engineering concept spaces for semantic retrieval: the Illinois Digital Library Initiative project

This research presents preliminary results generated from the semantic retrieval research component of the Illinois Digital Library Initiative (DLI) project. Using a variation of the automatic thesaurus generation techniques, to which we refer to as the concept space approach, we aimed to create graphs of domain-specific concepts (terms) and their weighted co-occurrence relationships for all major engineering domains. Merging these concept spaces and providing traversal paths across different concept spaces could potentially help alleviate the vocabulary (difference) problem evident in large-scale information retrieval. In order to address the scalability issue related to large-scale information retrieval and analysis for the current Illinois DLI project, we conducted experiments using the concept space approach on parallel supercomputers. Our test collection included computer science and electrical engineering abstracts extracted from the INSPEC database. The concept space approach called for extensive textual and statistical analysis (a form of knowledge discovery) based on automatic indexing and co-occurrence analysis algorithms, both previously tested in the biology domain. Initial testing results using a 512-node CM-5 and a 16-processor SGI Power Challenge were promising.

1984 International Conference On Parallel Processing

This paper describes a set of microbench- marks fc'r measuring the performance characteristics of MPI impkmentations. We explain the rationale behind the bench- marks and present the benchmark results on the Intel Para- gon, IBM SP2, and SGI Power Challenge. Our measurements reveal how the hardware architecture and the underlying MPI implementation afject the message passing performance on the three platforms.