LSBATCH: A Distributed Load Sharing Batch System

Batch processing, a primary mode of computing in mainframes and supercomputers, is becoming important for networked systems as the computing environments become more and more distributed. In this paper, we discuss the architectural and design considerations , and some important implementation issues of Lsbatch, a distributed batch system for large{scale, heterogeneous computer systems. Lsbatch supports batched submission and execution of parallel as well as sequential jobs in a system of up to several thousand hosts with possibly diierent architectures, Unix operating system varieties, and power. Implemented on top of the Utopia network operating system 5] as a distributed utility, Lsbatch takes advantage of the rich set of resource and load information services and eecient remote execution mechanisms of Utopia and provides scheduling algorithms to fully utilize the computing resources scattered around a distributed system. To the user, Lsbatch appears very much like a exible single host batch system; the complexities of distributed and heterogeneous computers and system failures are hidden in the implementation. To satisfy real world constraints to resource sharing, a variety of connguration mechanisms are provided in Lsbatch to allow sophisticated site management policies to be established and enforced.