Parallel application scheduling on networks of workstations

Parallel application support is one of the ways that have been recently proposed for exploiting the idle computing capacity of workstation networks. However, it has been unclear how to most effectively schedule the processors among different job requests. The distributed memory nature of such environments and the structure of the corresponding applications cause many solutions that were successful for shared-memory machines to be inadequate in the new environment. In this thesis, we investigate how knowledge of system load and application char-acteristics can be used in scheduling decisions. We propose a new algorithm that, with proper exploitation of both the information types above, manages to improve the performance of non- preemptive scheduling relative to other rules. Thus, we show that with the appropriate support in the operating system, the application devel-opers can be left free to use the programming model best suited to the individual application needs. i