Ambassadors in an orbital super-cluster

Abstract

Past attempts to exploit the unused capacity of clusters of users’ desktop workstations to run super-computing jobs have proven less successful than was expected. One solution to the problems of past approaches is an orbital supercluster. This is a loose federation of clusters of workstations, these clusters being distributed across the world. It has the following properties: Jobs submitted to the super-cluster are started only on clusters in parts of the world where it is night – when such clusters are free from use (and interference) by desktop users. Long running jobs are migrated periodically to keep the execution in nighttime. To implement an orbital super-cluster, the following are required: Techniques for efficient interaction over the very large latencies between an executing job and that job's client. Techniques for periodic migration of otherwise stationary computational elements of jobs. This paper analyses the reasons for the failure of past approaches, explains the approach and how it avoids past problems, and describes the techniques for interaction over very large communication latencies that will underpin implementation of the approach.