Predictable Network Computing

Clusters of networked commercial, off-the-shelf (COTS) workstations are presently used for computation intensive tasks that were typically assigned to parallel computers in the past. However, it is hardly possible to predict the timing behavior of such systems or to give guarantees about execution times. In this paper we show how our SONiC(Shared Objects Net-interconnected Computer) system can control timing and partitioning of a workstation as a step towards a distributed real-time system built from COTS components. SONiC provides a class-based programming interface for creation of replicated shared objects of arbitrary , user-defined sizes. Weak consistency protocols are employed to improve system's performance. Our Scheduling Service ensures the requested interactive behavior of a workstation while simultaneously giving a specified number of CPU cycles to parallel tasks. Using off-line scheduling methods we are able to implement real-time Guaranteed Services on COTS workstations.