Vote, a virtual shared-memory system and an extension to the Peace parallel operating system, provides architectural transparency and efficiency to effectively solve HPC problems.
A major concern with high-performance general-purpose work-stations is to speed up the execution of commands, uniprocess applications, and multiprocess applications with coarse- to medium-grain parallelism. To that end, a simple extension of a uniprocessor machine such as a shared-bus, shared-memory architecture can be employed. Both kinds of machines generally use the same OS model, and the same application can execute on these machines without recoding. However, an intrinsic limitation of the shared-bus architecture is the low number of processors that can be connected to the shared bus. When this number exceeds a critical value, the system's global performance drops drastically because of bus saturation. When two or more processors store a copy of the same memory block in their respective caches and one of them performs a write operation on a location in that block, a set of bus actions is necessary to guarantee that every subsequent read operation by any processor can get the up-to-date value of the modified location. Typically, researchers use simulation to investigate how to improve the performance of such machines. In particular, trace-driven simulation offers a good trade-off between speed, accuracy, and flexibility. A key point of this methodology is to find traces that both represent typical operating conditions and include all information potentially needed for an accurate simulation of the system. The authors have developed a methodology and a set of tools (called Trace Factory) to generate traces for the performance evaluation of shared-bus, shared-memory multiprocessor systems. Trace Factory is particularly useful for evaluating a multi-processor architecture's performance related to different work-loads and to most of the influencing activities of the operating system. The designer can evaluate and tune architectural solutions for coherence protocol, cache structure, bus, and memory.
Los Alamos National Laboratory and the Digital Equipment Corporation have developed a workstation cluster that provides visualization power that they claim could be as good as or better than many supercomputers. Workstation clusters are becoming the computational resource of choice for many researchers: their power rivals that of supercomputers, and they're cheaper. However, they run into problems when they are used for visualization: clusters have not been able to seamlessly transmit visualization data to a screen. The authors have overcome this limitation with the Distributed Data Imaging System described herein.
Volumetric data rendering is CPU- and memory-intensive. Technological innovations in data acquisition and data generation will probably continue to grow at a rate such that the data produced are orders of magnitude too big to be handled, analyzed, or projected (rendered) on an affordable desktop computer. In addressing these demands, the San Diego Supercomputer Center Networked Volume Renderer efficiently manages and provides access to several direct-volume-rendering algorithms running on advanced-computing rendering engines. SDSC_NetV's most innovative aspects are the easy-to-use graphical interface, which hides all notion of a network, and the server manager, which assigns advanced computing servers based on the user's rendering request and on the set of currently available resources. The server manager, acting as a render broker for big rendering jobs, works in unison with the graphical interface to give the user several of the most effective features from turnkey, application-builder, and advanced-architecture volume-rendering systems.
Designing a parallel machine would be much easier if one inter-connection network were "best" for all applications and all operating environments (including hardware, software, and financial factors). Unfortunately, no such network exists. Furthermore, even for a fixed application domain and a fixed operating environment, selecting the best network may be difficult because many cost and performance metrics could be used. Suppose someone asked you to select the best animal. What features would you use to compare, say, an alligator and an armadillo? In some ways, the two are very similar: both have four legs, a rugged exterior, and sharp claws. However, in other ways the two are very different: one prefers a marshy environment, the other dry land; one has a long tail, the other a short tail; one is a reptile, the other a mammal. Which of the two, then, is a better animal and what makes it better? These questions apply to interconnection networks as well. Suppose you are comparing the average message delay, for a given set of traffic conditions, for a hypercube network and a mesh network. do you incorporate that hypercube networks may require more complex hardware? Should the total path width of all the links the networks employ be the same? Or should the two networks require the same number of transistors per switch? This article explores the problems of determining which metrics or weighted set of metrics designers should use to compare networks and how they should apply these metrics to yield meaningful information. The authors also look at problems in conducting fair and scientific evaluations.
HIPPI networks, which are popular as connections among supercomputers and high-end workstations, have not yet become widespread in distributed computing because of their distance limitations. This article compares two schemes that overcome this restriction and provide high bandwidth utilization: HIPPI tunneling and IP routing.
In exploring the performance of remote tasking and distributed interprocess communication, these authors find that concurrency introduced at the lowest levels of software is vital to exploiting concurrency at higher levels.
To speed up the consistent-state restoration of distributed systems, Complete Process Recovery uses vector time to address unorthodox message-handling issues and overlapping failures.
This article describes the findings and analysis following the European Commission-funded survey of High-Performance Computing and Networking (HPCN) software tools. The study's goal was to ascertain which forces are shaping the European HPCN software-tool market. The survey used detailed interviews with 25 European tool-development centers to define a model for the European software-tool market. Analysis of the market structure indicates that where software tools contain technologies linked to the underlying hardware, the policies of the hardware vendors are dominant forces. The influence of hardware vendors is less apparent where the tools are more closely associated with software applications.
The rise of the Internet and World Wide Web (WWW) might be breaking down barriers between the President and the public. Since Bill Clinton's first presidential campaign in 1992, his staff has used the Internet as a communications conduit. With the help of research scientists from the Massachusetts Institute of Technology's (MIT's) Artificial Intelligence Laboratory, the White House's Comlink syste...
The Trapper software-engineering environment for parallel and distributed systems comprises tools for software design, hardware configuration, mapping, monitoring, visualization, and performance tuning of parallel applications and systems. This article's authors used Trapper as the main software-development tool to realize a complex industrial research application within the Prometheus collision-avoidance project at Daimler-Benz. In this project, they outfitted a Mercedes 500 SEL research vehicle with 18 cameras and approximately 60 computing nodes for various image-processing and control algorithms such as lane following, obstacle detection, and traffic-sign recognition. This article seeks to analyze the behavior and improve the performance of the most time-consuming task of the whole software: traffic-sign recognition. The authors show how Trapper effectively detected and eliminated performance bottlenecks of this parallel application.
