Abstract
As real-time applications become more and more complicated their demands of processing capacity can hardly be satisfied. Massively parallel computers, such as Intel Paragon, with their scalable architecture and tremendous processing power make them very attractive for the next generation real-time systems. How to support realtime communication on the massively parallel computers becomes an important issue. This project focuses on wormhole networks of Paragon, and study different transmission control approaches. In this kind of network, transmission control at the source node plays an important role in supporting real-time traffic. Traditional system uses an unregulated greedy transmission control. This may result in unfairness of network access and unbounded packet blocking time, making it very difficult to efficiently support real-time applications. To overcome this problem, this project proposes a new regulated-transmission approach in which packet transmission at the source is regulated and hence unnecessary network contention is eliminated. This method is distributed and requires no explicit state information exchange at run time. It is compatible with the current architecture of Paragon and can be easily implemented. The success of this approach relies on assigning properly values to all parameters. A simple heuristic but efficient parameter selection algorithm is developed. The performance of each approach is evaluated in terms of the worst case achievable utilization (WCAU). This metric is important to real-time applications as it is related to predictability and stability of the system. The performance comparisons are also made via simulation in terms of the metric called guarantee probability. From these evaluations, We can see that regulated transmission approach using the parameters selected by our heuristic algorithm performs much better than the unregulated one and has a potential application in practical systems.
Li, Hung (1995). Scheduling of real-time communication network for parallel processing. Master's thesis, Texas A&M University. Available electronically from
https : / /hdl .handle .net /1969 .1 /ETD -TAMU -1995 -THESIS -L552.