Distributed systems such as clusters of PCs are low-cost alternatives for running parallel rendering systems, but they have high communication overhead, and limited memory capacity on each processing node. In this paper, we focus on the strategy for distributing the parallel rendering work among the PCs. A good distribution strategy provides better load balance, and avoids the need for replicating data on the relatively small memory of each PC. Our goal is to study different distribution strategies on the scope of the Parallel ZSweep algorithm, introducing in PZSweep another work distribution strategy: work stealing. This strategy allows a decentralized control of the work to be done, and provides a dynamic load redistribution. We propose two different algorithms to select the processor that will be "stolen" and show that the simplest one, nearest neighbor, was the most efficient. We also showed that the load redistribution schemes strongly depended on the initial load distribution, with an interleaved assignment, our systems could outperform the original Parallel ZSweep algorithm. We conclude that for running large datasets on a cluster of PCs, Parallel ZSweep requires dynamic load distribution strategy.

Este artigo apresenta a arquitetura e os resultados da avaliação de desempenho do supercomputador Netuno, um cluster de alto desempenho recentemente instalado na UFRJ. São apresentados detalhes tanto de sua arquitetura como dos softwares básicos e de middleware utilizados na sua construção. Os resultados de avaliação obtidos registram um desempenho de 16,2 Tflops sustentados para o benchmark HPL (High Performance Linpack), o que colocou o supercomputador Netuno na 138ª posição na lista Top500 de junho de 2008. Atualmente, o supercomputador Netuno atende diversas instituições de ensino e pesquisa no Brasil, participantes das redes temáticas de pesquisa de Geofísica Aplicada e de Oceanografia (REMO), patrocinadas pala Petrobras.

No abstract

This paper presents a description and the evaluation of the Netuno supercomputer, a high-performance cluster installed at Federal University of Rio de Janeiro in Brazil.The results for the High Performance Linpack (HPL) benchmark and two real applications are reported. Since building a high-performance cluster for running a wide range of applications is a non-trivial task, some lessons learned from assembling and operating this cluster, such as the excelent performance of the OpenMPI library, and the importance of the use an efficient parallel file system over the traditional NFS system, can be useful knowledge to support the design of new systems. Currently, Netuno is being heavily used to run large scale simulations in the areas of ocean modeling, meteorology, engineering, physics, and geophysics.