Measurement of memory access contentions in multiple vector processor systems
The purpose of this paper is to study delays encountered when several vector processors access a common memory. Such delays may be caused by access conflicts at the memory or by contentions within the network connection processors to common memory. Because such delays are potentially the limiting performance bottleneck of common memory architectures, it is important to understand how they depend on design parameters, in particular the number of processors serviced by the common memory, the degree of memory interleaving, and the speed of the memory chips. Such understanding will make it possible to assess the performance of systems with increasing number of processors and new memory technologies. It is essential for any assessment of the future potential of common memory systems. There are three possible approaches to the study of this problem, analytical models, simulations, and measurements on existing machines. All three approaches are necessary because each has advantages and limitations. Analytical models owe their effectiveness to simplifying assumptions that need to be checked by the other two methods. In particular, stochastic methods and queueing theory offer not techniques to deal with regular memory access patterns that are common in scientific and engineering applications. Simulations can generally approximate reality more closely, however, even simulations do not replace real measurements. They are often very slow and contain assumptions that need to be tested. This paper presents a set of measurements of memory access delay performed on several computers of the Cray family under carefully controlled conditions. 9 refs., 12 figs., 1 tab.
- Research Organization:
- Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
- Sponsoring Organization:
- USDOE; USDOE, Washington, DC (United States)
- DOE Contract Number:
- W-7405-ENG-36
- OSTI ID:
- 5369921
- Report Number(s):
- LA-UR-91-1665; CONF-911103-6; ON: DE91013205
- Resource Relation:
- Conference: Conference on supercomputing '91, Albuquerque, NM (United States), 18-22 Nov 1991
- Country of Publication:
- United States
- Language:
- English
Similar Records
Access ordering and coherence in shared-memory multi-processors
Vector access performance in parallel memories using a skewed storage scheme