Algorithmic support for commodity-based parallel computing systems.

Abstract

Follows 3 AbstractThe Computational Plant or Cplant is a commodity-based distributedmemory supercomputer under development at Sandia National Laboratories. Distributed-memory supercomputers run many parallel programs simultaneously. Users submit their programs to a job queue. When a job is scheduled to run, it is assigned to a set of available processors. Job runtime depends not only on the number of processors but also on the particular set of processors assigned to it. Jobs should be allocated to localized clusters of processors to minimize communication costs and to avoid bandwidth contention caused by overlapping jobs.This report introduces new allocation strategies and performance metrics based on space-filling curves and one dimensional allocation strategies. These algorithms are general and simple. Preliminary simulations and Cplant experiments indicate that both space-filling curves and one-dimensional packing improve processor locality compared to the sorted free list strategy previously used on Cplant. These new allocation strategies are implemented in Release 2.0 of the Cplant System Software that was phased into the Cplant systems at Sandia by May 2002.Experimental results then demonstrated that the average number of communication hops between the processors allocated to a job strongly correlates with the job's completion time. This report also gives processor-allocation algorithms for minimizing the average number of communication hops between the assigned processors for grid architectures.The associated clustering problem is as follows: Given n points in d , find k points that minimize their average pairwise L 1 distance. Exact and approximate algorithms are given for these optimization problems. One of these algorithms has been implemented on Cplant and will be included in Cplant System Software, Version 2.1, to be released. In more preliminary work, we suggest improvements to the scheduler separate from the allocator.

Communication-Aware Processor Allocation for Supercomputers: Finding Point Sets of Small Average Distance

“…Our work also led to a linear-time dynamic programming algorithm for the 1-dimensional problem of points on a line or ring; see Leung et al [17] for details.…”

A framework for routing and resource allocation in network virtualization

“…[25]. However, we are not aware of other strategies than UDFlex [26] that utilize a topology agnostic routing algorithm to achieve routing-containment for irregular partitions.…”

Communication patterns and allocation strategies.

“…Since job placement affects the system's network contention, allocation is a major factor in determining system performance, particularly in commodity-based supercomputers such as Cplant, which typically have higher communication latencies and lower bandwidth than supercomputers with custom networks. Experiments on Cplant showed that poor allocation could increase the running time of a pair of highcommunication jobs by as much as a factor of two [17]. Other researchers have shown in a variety of studies that interprocessor communication can reduce throughput [2,20,21,24].…”