A protocol for load sharing among a cluster of heterogeneous Unix workstations

Abstract

In this paper we propose a protocol for load sharing among a cluster of heterogeneous Unix workstations. Our protocol called the distributed process management protocol (DPMP), not only enables load sharing using nonpreemptive process migration but also seamlessly integrates the processes running on a network of machines. Remote processes can be accessed (for signalling, for example), in the same way as local processes making process migration highly transparent to the users and the applications. DPMP also has built in mechanisms to detect and recover from node and network failures. DPMP can be implemented at either the kernel or the user level. We also describe an implementation of DPMP within the Linux kernel. Preliminary performance studies show that the performance gains obtained by using DPMP are substantial.