A Highly Available Parallel Environment for PVM/MPI

Du Shu, Wang Dongsheng
Department of Computer Science, Tsinghua University, 
Beijing 100084, P.R.China

	Recently, the Cluster of Computers (COC) has been used to run large 
parallel programs increasingly. Task migration is a desirable and useful 
facility to implement Load-Balance and High Availability in COCs. This report 
presents a quick migration protocol for PVM/MPI tasks, which allows non-migrating 
tasks to execute during most of the time of migration. Message buffering and 
process table updating are key mechanisms of this protocol. Because MPI 
implementation does not make provisions for tasks migration, this report also 
describes the work we have done to modify an MPICH P4 implementation to allow 
task migration. Processes management mechanism and deadlock avoidance technique 
are the key issues we discuss. At last we introduce the high availability 
characters of our system that is completed grounded on these ideas and techniques. 
This system can not only recover from the process-error and node-failure, but 
also implements a simple load-balancing algorithm to automatically take advantage 
of the new available computing nodes.

Keywords: Cluster of Computers (COCs), High Availability, MPI, Process Migration, 
	Checkpointing