Files in this item
|(no description provided)|
|Title:||Circuit-switched multicomputers and heuristic load placement|
|Author(s):||Grunwald, Dirk Claus|
|Doctoral Committee Chair(s):||Reed, Daniel|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||In this thesis, we examine two areas germane to ensemble, or multicomputer, systems: communication between individual computers and the placement of the individual components of a parallel program in the ensemble. These topics are related. Programs that execute on ensemble systems must be partitioned, and those components must communicate to be effective. We assume the raison d'etre of ensemble systems is performance; the methods used to solve problems, and the implementation of those methods on an ensemble system depend on the communication and computational performance of the ensemble system.
However, such performance has a price; the programming model for ensemble systems differs greatly from the shared memory paradigm to which we have grown accustomed. Programming languages and tools that simplify the process of partitioning and placing individual parts of a computation are needed. In this thesis, we show that simple strategies used to share the resources of an ensemble system between individual program components, or processes, are very effective.
We measure the performance of existing ensemble computer architectures and compare existing network architectures. We analyse several routing algorithms for circuit switched ensembles. These algorithms are simple enough to allow hardware implementation with switching speeds less than five hundred nanoseconds. We compare these networks to existing and proposed networks.
We use the network model to examine process placement on an ensemble system. We characterize the observable behavior of processes; this is used to generate synthetic workloads for a simulation of process placement. We also capture the observational behavior of parallel prolog programs and two search tree programs. The conclusions from our study show that, with an appropriate network model, random process distribution is sufficient; in fact, process distribution using knowledge of the current system state is generally penalized due to stale information and information overflow.
|Rights Information:||Copyright 1989 Grunwald, Dirk Claus|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9010869|