SNOW: A Distributed Dynamic System for Metacomputing


In a non-dedicated network environment, computers are privately owned. Individual owners do not want to see their systems being saturated by others when they need them. This means the privately owned machines may only be used for parallel processing on an ``availability'' basis. In addition, a non-dedicated network of computers is more likely to be heterogeneous than is a dedicated system. ``Availability'' and ``heterogeneity'' are new issues of distributed network computing, which do not arise in tightly coupled parallel systems. Competition for computing resources does not lead to guaranteed high performance. To simultaneously utilizing idle machines and maintaining high capabilities for local computations, network process migration mechanism is proposed as a solution. The simple idea underlying this mechanism is that when the imbalanced workload of a distributed system occurs, parallel processes residing on overloaded machines are migrated to other available machines.

We have introduced a high-level mechanism and its associated methodologys to support efficient process migration in a non-dedicated, heterogeneous network computing environment. The newly proposed network process-migration mechanism is general. It can be applied to any distributed network environment. In particular, based on this mechanism, a software system named MpPVM is designed and implemented to support efficient process migration for PVM application programs. We have studied the interaction between process migration and resource management and proposed modifications of pvmd and pvmlib of PVM to maintain reliable data communications among processes in a migration environment. Our implementation and experimental results confirm the applicability and potential of the proposed mechanism in a non-dedicated heterogeneous environment. Experimental results indicate the MpPVM software system is efficient and is scalable in the sense that it can carry out a process migration in tenth of milliseconds and the migration costs becomes less notable when the problem and ensemble increase.




Xian-He Sun
Department of Computer Science
Illinois Institute of Technology
Chicago, IL 60616

Back to SCS Home Page