Performance Modeling of Multiprocessor Systems With Paging

97 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1980.The rapid advancement in semiconductor technology continues to change the environment in which computers are designed. As hardware costs decline, systems with multiple processors become an interesting alternative to conventional single processor systems. An analytic model has been developed to describe the performance of a wide range of multiprocessor system configurations and workloads. This model deals specifically with P tightly-coupled, identical processors with shared primary and secondary memory. Secondary memory consists of a paging drum with S sectors. The workload consists of J independent, identically-distributed jobs whose faulting or I/O behavior is described by both a mean ((lamda) faults/sector-time) and a squared coefficient of variation (K). In addition, the processing overhead for each I/O request is added to a job's execution time at a processor (C sector-times/fault).The model developed provides an estimate of system throughput for various numbers of processors, jobs, and drum sectors, and for various workloads. Throughput, the average number of processors doing useful work, is given by(DIAGRAM, TABLE OR GRAPHIC OMITTED...PLEASE SEE DAI)This model is based on a deterministic scheduling model for the system, and known models which describe the sub-parts of the system are embedded. The accuracy of the model is assessed by comparison to a large number of runs of a simulator using exponential and hyperexponential fault time distributions. For a wide range of values of the parameters, the formula provides a very good estimate of throughput (the average relative error for 195 simulation runs is only 3.0%.). Even though the squared coefficient of the fault distribution varied from 1 to 16 in the simulation runs, the model fit quite well without using K. This suggests that the mean fault rate is perhaps a sufficient measure of the faulting process. Further evidence for this conclusion is the fact that the model's prediction only improves slightly when a drum queue wait model is employed which includes K.The model can be used to examine the behavior of multiprocessor systems, including the sensitivity of system throughput to each of the system parameters and parameter trade-offs related to system performance. In particular, in a system with a fixed amount of memory, the addition of jobs to the system causes a change in the memory allocation for each job and thus modifies each job's faulting behavior. The above formula for throughput is useful to examine the desirability of adding or subtracting jobs in such a system.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD