When a parallel computer is used to run several parallel jobs concurrently, there are well-known performance benefits to be obtained if the process scheduling is coordinated so that all the processes of each parallel job are run at the same time. Currently the most popular scheme for coordinated scheduling is gang scheduling. However, there are many problems associated with conventional gang scheduling, such as inscalability and inefficiency in resource utilisation and job performance, which have hindered its widespread adoption. We have developed techniques which solve these problems and thus allow multiple parallel programs to run concurrently and share the machine in a scalable and efficient fashion.
Bing Bing Zhou
Paul Mackerras
David Walsh