DAWGS - A Distributed Compute Server Utilizing Idle Workstations

Abstract A collection of powerful workstations interconnected by a local area network forms a large computing resource. The problem of locating and efficiently using this resource has been the subject of much study. When the system is composed of workstations, an attractive technique may be employed to make use of workstations left idle by their owners. The Distributed Automated Workload balancinG System (DAWGS) is designed to allow users to utilize this networked computing power for their programs. Essentially, DAWGS is an interface between the user and the kernel which allows users to submit batch-type or interactive-type processes or jobs for execution on an idle workstation somewhere on a local area network. DAWGS uses a distributed scheduler based on a bidding scheme which resolves many of the problems with bidding to determine which machine to run a process on. It properly redirects all I/O from the remotely running process back to the machine from whence the process came. DAWGS is capable of checkpointing processes and restarting any type of process, including interactive ones, even when the restart is on a machine different than the one the process was previously running on. We show that running processes remotely on idle workstations can result in significantly lower execution times, particularly for processes with a large execution time. Our method is different from previous work in that it is fault-tolerant, maintains total remote execution transparency for the user, and is fully distributed.