Daniel Savarese

Communication overhead for space science applications on the Beowulf parallel workstation

The Beowulf parallel workstation combines 16 PC-compatible processing subsystems and disk drives using dual Ethernet networks to provide a single-user environment with 1 Gops peak performance, half a Gbyte of disk storage, and up to 8 times the disk I/O bandwidth of conventional workstations. The Beowulf architecture establishes a new operating point in price-performance for single-user environments requiring high disk capacity and bandwidth. The Beowulf research project is investigating the feasibility of exploiting mass market commodity computing elements in support of Earth and space science requirements for large data-set browsing and visualization, simulation of natural physical processes, and assimilation of remote sensing data. This paper reports the findings from a series of experiments for characterizing the Beowulf dual channel communication over-head. It is shown that dual networks can sustain 70% greater throughput than a single network alone but that bandwidth achieved is more highly sensitive to message size than to the number of messages at peak demand. While overhead is shown to be high for global synchronization, its overall impact on scalability of real world applications for computational fluid dynamics and N-body gravitational simulation is shown to be modest.

A Coming of Age for Beowulf-Class Computing

Beowulf-class systems along with other forms of PC clustered systems have matured to the point that they are becoming the strategy of choice for some areas of high performance applications. A Beowulf system is a cluster of mass market COTS personal computers interconnected by means of widely available local area network (LAN) technology. Beowulf software is based on open source code Unix-like operating systems that, in a majority of cases, is Linux. The API for Beowulf is based on message passing semantics and mechanisms including explicit models such as PVM and MPI or implicit models such as BSP of HPF. Since its introduction in 1994, Beowulf-class computing has gone through five generations of PCs from multiple microprocessor vendors including the Intel x86 family, DEC’s Alpha, and the PowerPC from IBM and Motorola. Originally, Beowulfs were implemented as small clusters in the range of 4 to 32 nodes. Larger clusters of 48 to 96 processors were deployed two and a half years ago. Today there are many systems of 100 to 300 processors with systems of over a thousand processors in the planning stage for implementation over the next year.

Improving Application Performance on HP/Convex Exemplar

The Earth and space sciences community faces rich computational challenges ranging from static, regular, and embarrassingly parallel to dynamic, unstructured, and tightly coupled. This problem domain requires highly scalable systems exhibiting broad generality, efficiency, and programmability. These capabilities are appearing in the emerging scalable shared memory cache-coherent architectures like that of the HP/Convex Exemplar SPP-1000. The goal of this class of architecture is to make scientific programming as easy and efficient as it is on vector supercomputers. The authors describe the Exemplars architecture, whose global system organization comprises up to 16 multiprocessors interconnected by four SCI (Scalable Coherent Interface) ring networks. They then present the findings from four applications: the piecewise parabolic method, a finite-element method for unstructured meshes, a tree code for the n-body problem, and a particle-in-cell code. The authors present application performance data derived after the Exemplar at Goddard Space Flight Center went into production use. These studies expose the operational properties of the Exemplar and determine its suitability for Earth and space sciences applications. The testing reveals that global cache coherence can be used effectively to simplify programming and data migration. However, the basic problem of locality sensitivity still demands direct programmer involvement to achieve effective system behavior. The question of whether message-passing or shared memory programming models are better remains open.

An initial evaluation of the Convex SPP-1000 for earth and space science applications

The Convex SPP-1000, the most recent SPC, is distinguished by a true global shared memory capability based on the first commercial version of directory based cache coherence mechanisms and SCI protocol. The system was evaluated at NASA/GSFC in the Beta-test environment using three classes of operational experiments targeting earth and space science applications. A multiple program workload tested job-stream level parallelism. Synthetic programs measured overhead costs of barrier, fork-join, and message passing synchronization primitives. An efficient tree-code version of an N-body simulation revealed scaling properties and measured the overall efficiency. This paper presents the results of this study and provides the earliest published evaluation of this new scalable architecture. >

From Toys to Teraflops: Bridging the Beowulf Gap

Do-it-yourself supercomputing has emerged as a solution to cost-effectively sustain the computational demands of the scientific research community. Despite some of the successes of this approach, represented by Beowulf-class computing, it has limitations that need to be recognized as well as problems that need to be resolved in order to extend its scope of applicability. While the performance of hardware incorporated into these systems has continued to improve at a remarkable rate, enabling the execution of steadily larger and more compute-intensive applications, the software environment of the machines has seen little to no improvement or evolution. The authors find that this gap between the rates of development of hardware and software is a crucial obstacle to the full exploitation of these systems into and beyond the Teraflops realm. They provide suggestions as to how this gap might be narrowed within the context of Beowulf-class computing.