Raymond M. Bryant

The MVA priority approximation

A Mean Value Analysis (MVA) approximation is presented for computing the average performance measures of closed-, open-, and mixed-type multiclass queuing networks containing Preemptive Resume (PR) and nonpreemptive Head-Of-Line (HOL) priority service centers. The approximation has essentially the same storage and computational requirements as MVA, thus allowing computationally efficient solutions of large priority queuing networks. The accuracy of the MVA approximation is systematically investigated and presented. It is shown that the approximation can compute the average performance measures of priority networks to within an accuracy of 5 percent for a large range of network parameter values. Accuracy of the method is shown to be superior to that of Sevciks shadow approximation.

The IBM external user interface for scalable parallel systems

Abstract The IBM External User Interface (EUI) for scalable parallel systems is a parallel programming library designed for the IBM line of scalable parallel computers. The first computer in this line, the IBM 9076 SP1, was announced in February 1993. In essence, the EUI is a library of coordination and communication routines that can be invoked from within FORTRAN or C application programs. The EUI consists of four main components: task management routines, message passing routines, task group routines, and collective communication routines. This paper examines several aspects of the design and development of the EUI.

The MVA Pre-empt resume priority approximation

A Mean Value Analysis (MVA) approximation is presented for computing the average performance measures of closed multiclass queueing networks containing non pre-emptive Head Of Line (HOL) and Pre-empt Resume (PR) priority centers. The approximation has the same storage and computational requirements as MVA thus allowing computationally efficient solutions of large priority queueing networks. The accuracy of the MVA PR approximation is systematically investigated and presented in terms of error contour diagrams. The contour diagrams reveal that the approximation can compute the average performance measures of priority networks to within an accuracy of 5 percent for a large range of network parameter values. Accuracy of the method is also compared to Sevciks shadow approximation and another MVA approximation recently proposed by Chandy and Lakshmi.

Operating system support for parallel programming on RP3

RP3, the Research Parallel Processing Prototype, was a research vehicle for exploring the hardware and software aspects of highly parallel computation. RP3 was a shared-memory machine that was designed to be scalable to 512 processors; a 64-processor machine was in operation from October 1988 through March 1991. A parallel-programming environment based on the Mach operating system was developed, and a variety of programming models were tested on the machine. To help user programs realize the full potential of parallelism on RP3, the RP3 operating system was extended to support such RP3 architectural features as noncoherent caches, local and interleaved storage, and a hardware performance monitor. The system included explicit job-scheduling and processor-allocation facilities, facilities for exploiting the RP3 memory hierarchy, and performance-data collection and logging facilities. This paper describes these components of the RP3 operating system, provides the rationale for the design decisions that were made, and discusses the implementation of these operating system facilities.

Maximum Processing Rates of Memory Bound Systems

Multiresource queuing systems are of particular maportance in modding compnter systems because a job must have access to a processor and main memory sunultaneousIy in order to proceed. Existing methods of determining maximum processing rates for multiresourcequeuing systems are limited to small memory sizes because problem compleraty grows exponentially with increasing memory size. By restneting our attention to a particular scheduling discipline (first-come-first-loaded or FCFL) and treating memory as the limiting resource, methods of calculating maximum processing ~ates of memory bound systems for reahstic mare memory sizes are derived. The distribution of the number of jobs loaded under the FCFL pohey is given m terms of a convolution of the memory request size distribution. The time averaged behavior of the number of loaded jobs is also found. Finally, the framework is extended to allow multiple job classes in the input stream. The results of fins approach allow one to estimate main memory size requirements from a workload characterization given m terms of arrival rate, memory size distribution, and CPU service rate.