Massively Parallel Processing Systems Research Articles

Hybrid checkpointing using emerging nonvolatile memories for future exascale systems

The scalability of future Massively Parallel Processing (MPP) systems is being severely challenged by high failure rates. Current centralized Hard Disk Drive (HDD) checkpointing results in overhead of 25% or more at petascale. Since systems become more vulnerable as the node count keeps increasing, novel techniques that enable fast and frequent checkpointing are critical to the future exascale system implementation. In this work, we first introduce one of the emerging nonvolatile memory technologies, Phase-Change Random Access Memory (PCRAM), as a proper candidate of the fast checkpointing device. After a thorough analysis of MPP systems, failure rates and failure sources, we propose a PCRAM-based hybrid local/global checkpointing mechanism which not only provides a faster checkpoint storage, but also boosts the effectiveness of other orthogonal techniques such as incremental checkpointing and background checkpointing. Three variant implementations of the PCRAM-based hybrid checkpointing are designed to be adopted at different stages and to offer a smooth transition from the conventional in-disk checkpointing to the instant in-memory approach. Analyzing the overhead by using a hybrid checkpointing performance model, we show the proposed approach only incurs less than 3% performance overhead on a projected exascale system.

Using cluster computers in bioinformatics research

In the last ten years, high-performance and massively parallel computing technology comes into a high speed developing phase and is used in all fields. The cluster computer systems are also being widely used for their low cost and high performance. In bioinformatics research, solving a problem with computer usually takes hours even days. To speed up research, high-performance cluster computers are considered to be a good platform. Moving into the new MPP (massively parallel processing) system, the original algorithm should be parallelized in a proper way. In this paper, a new parallelizing method of useful sequence alignment algorithm (Smith-Waterman) is designed based on its optimizing algorithm already exists. The result is gratifying.

Wavelength-division multiplexing free-space optical interconnect networks for massively parallel processing systems

Wavelength-division multiplexing (WDM) techniques provide many advantages for building optical interconnect networks for massively parallel processing (MPP) systems. A design for a 1024-channel network for MPP systems based on the interconnection-cached network with vertical-cavity surface-emitting laser (VCSEL) arrays with one wavelength is described. We then show how a WDM version with four different wavelengths can increase the channel density. We also show how a WDM system can reduce the fan-in loss by a factor of 4. All the VCSEL's in each array are of the same wavelength, while different arrays use different wavelengths. We describe our experimental WDM subsystem containing four VCSEL arrays, operating at wavelengths of 843, 950, 970, and 980 nm, and three different WDM filters for multiplexing-demultiplexing. We present the operational results of the subsystem at 1 Gbit/s per channel.

Modeling and evaluating design alternatives for an on-line instrumentation system: a case study

This paper demonstrates the use of a model-based evaluation approach for instrumentation systems (ISs). The overall objective of this study is to provide early feedback to tool developers regarding IS overhead and performance; such feedback helps developers make appropriate design decisions about alternative system configurations and task scheduling policies. We consider three types of system architectures: network of workstations (NOW), symmetric multiprocessors (SMP), and massively parallel processing (MPP) systems. We develop a Resource OCCupancy (ROCC) model for an on-line IS for an existing tool and parameterize it for an IBM SP-2 platform. This model is simulated to answer several what if questions regarding two policies to schedule instrumentation data forwarding: collect-and-forward (CF) and batch-and-forward (BF). In addition, this study investigates two alternatives for forwarding the instrumentation data: direct and binary tree forwarding for an MPP system. Simulation results indicate that the BF policy can significantly reduce the overhead and that the tree forwarding configuration exhibits desirable scalability characteristics for MPP systems. Initial measurement-based testing results indicate more than 60 percent reduction in the direct IS overhead when the BF policy was added to Paradyn parallel performance measurement tool.

Performance of Massively Parallel Computers for Spectral Atmospheric Models

Abstract Massively parallel processing (MPP) computer systems use high-speed interconnection networks to link hundreds or thousands of RISC microprocessors. With each microprocessor having a peak performance of 100 or more megaflops per second, there is at least the possibility of achieving very high performance. However, the question of exactly how to achieve this performance remains unanswered. MPP systems and vector multi-processors require very different coding styles. Different MPP systems have widely varying architectures and performance characteristics. For most problems, a range of different parallel implementations is possible, again with varying performance characteristics. In this paper, we provide a detailed evaluation of MPP performance for a spectral transform kernel as used in weather and climate modeling applications. Using a specially designed spectral transform code, the authors study performance on three different MPP systems: Intel Paragon, IBM SP2, and Cray T3D. Great care is taken to...

The SUPRENUM and GENESIS projects

This paper provides an overview of two major European projects involving massively parallel processing (MPP) GENESIS and SUPRENUM. The two projects were closely related and in fact overlapped in time. This issue is entirely devoted to papers dealing with these projects. In this paper we will briefly introduce each project and then provide an introduction and associated context for each of the subsequent papers related to that project. The SUPRENUM project was established as a consortium of German companies, university departments and research centers, with the goal of designing a Gflops MPP supercomputer for scientific applications. It resulted in the Suprenum-1 prototype computer which had a peak rate of over 5 Gflops. The GENESIS ESPRIT project was established in order to design a second generation European MPP system. It utilized the developments already available from SUPRENUM, Supernode and several other European projects. It resulted in major advances in network interconnections, node design, MPP operating systems and languages, and benchmarking. SUPRENUM, while strictly a German supercomputer project, led naturally to the broader European GENESIS project. The projects were in fact closely related and many of the key participants in SUPRENUM were also participants in