TRANSPARENT PROCESS MIGRATION PROTOCOL FOR DISTRIBUTED APPLICATIONS IN HETEROGENEOUS UNIX NETWORK

Abstract

In this thesis, analysis and design of a protocol for the transparent process scheduling and process migration in heterogeneous unix networks has been proposed and implemented. Further, an algorithm to measure the load of any machine in distributed heterogeneous Unix network has been devised and it is analyzed for the performance and efficiency. Because a distributed system can be used effectively by its end users only if its software presents a single system image of this physically distributed system to the users, So all the resources of any node should be easily and transparently accessible from any other. A common kind of machine clusters typically used is a set of heterogeneous Unix workstations. While solutions are available for transparent sharing of resources such as files and printers in such a scenario, an important resource that is typically not shared is the CPU. Hence, even though there is an idle machine is the system, a heavily loaded machine cannot share the idle CPU transparently which signifies wastage of CPU cycles. Process migration can reduce the loss of such CPU cycles by migrating processes from a highly loaded machine to the idle (or lightly loaded) one. Remote execution or non-preemptive migration is a form of process migration where the processes are placed in remote nodes during creation. In a heterogeneous environment only non-preemptive process migration can be performed transparently. We have proposed a transparent process scheduling and process migration protocol and a set of conventions to achieve load sharing for a cluster of heterogeneous machines each running some flavors of Unix. In the protocol, the single system image is also preserved by creating unique cluster wise process ids, keeping a traditional filename space, uniform user credentials, etc. The protocol itself do not provide all the features of process migration, but it rely on the underlying Unix network. For example, the protocol assumes that the machines already have a more or less uniform file system view(through NFS) and are uniformly administered for user credentials. These are declared as conventions that all machines in the cluster must ensure. The protocol also does not dictate the policy decisions and can work with any load sharing policy. In the current version of the protocol, there are a fewlimitations. For example, the current version of the protocol does not have support for migration of process from one architecture of Unix to another as, the migration time in this case is significantly high as compared to time required for the completion of process.