Resource-Controlled Parallel Execution of Real-Time Applications in NOWs

We propose to address problems of abstraction, mechanism, and policy involved in the use of networks of workstations (NOWs) to improve application performance and system capacity through opportunistic parallel processing:

• We intend to focus on applications whose user interfaces require updates at regular intervals, such as interactive computer graphics, rather than the long running, response time driven supercomputing applications that are the target of most NOW work.

• We intend to exploit the characteristics of these applications to allow NOW scheduling at a much finer granularity than has been considered previously, in times measured in seconds rather than minutes or hours. This fine scheduling granularity will allow us to take advantage of the unused resources on currently active workstations, since we rely on them for only short intervals at a time. Developing facilities to estimate available resources on active workstations, and to choose a set on which to host a parallel application, is another goal of this work.

• We propose to develop cooperative operating system and application level facilities that will allow the amount of resource consumed by a remotely executing application to be limited (thus protecting the performance delivered to the local user of the workstation) while still providing adequate application performance.

• Because we cannot rely on resource availability over the short interval of time between the display of successive frames, we plan to replicate the distributed tasks. A major goal of this proposal is to develop replication schemes that are both effective (i.e., lead to on-time completion of all work with very high probability) and efficient (i.e., consume only very modestly more resources in the normal case than would an unreplicated computation).

Our work will involve both modelling (during system design) and experimentation (during prototype evaluation). The test bed for our prototype will be a cluster of high-performance Pentium-based PCs employing two significant research prototypes: the SPIN extensible operating system and an ``open architecture'' high-speed switched network. As a side benefit of the work, it will serve as a significant test of the facilities these two independent research efforts provide to high-performance applications.