This paper describes a memory management study motivated by the PASM reconfigurable parallel processing system being designed and developed at Purdue University. The general problem of determining the optimal (maximum processor utilization) page service request rate for the common secondary storage of a reconfigurable system is described, and the PASM restrictions on this general problem are discussed. A queueing network model is developed for the memory hierarchy of the multiple control units in PASM. The model is analyzed to determine the optimal page service request rate, and simulation results verifying various aspects of the model are presented. The model is parameterized, and the parameters can be varied in order to determine the system performance impact of such changes. The techniques used are also related to the general model for a reconfigurable parallel processing system.
The PASM-based model involves the partitioning of a large number of processing elements into multiple virtual machines, each controlled by one or more control units. The multiple control units share a common secondary storage and use paging to transfer programs into their primary memory.
In addition to the limitations imposed on the general case by the PASM system, several simplifying assumptions were made about average execution time, task page fault rates, and the distribution of the number of controller groups required by a job (all of these assumptions can be easily changed, and new simulations can be run to assess their impact).
The simulation results support the analytical results for the average number of independent instruction streams and the worst-case controller idle time. The optimal service rate was also found to be a function of the average number of preloaded pages per job when the System Control Unit performed preloading.
The paper is splendidly organized and very well written. It presents an excellent overview of the general memory management problem for reconfigurable parallel processors, and it provides a clear and concise description of the PASM architecture and virtual memory system. The analysis presentation is also done nicely, preceded by a discussion of the terminology to be used and a brief description of the two-station cyclic network model that was used in the analysis. The Bibliography is extensive and cites a number of classic papers dating back almost 20 years. This paper is well worth reading not only in terms of the results, but also for the general statement of the problem and the PASM system overview. There is material here for parallel architecture experts and for those interested simply in an overview of current problems and research efforts.