T. Sterling et al., HPC ARCHITECTURE REQUIREMENTS FOR EARTH AND SPACE SCIENCE APPLICATIONS, Zeitschrift fur angewandte Mathematik und Mechanik, 76, 1996, pp. 445-448
Scalable high performance computer architecture brings substantial pro
cessing, memory, and communications resources to large scientific and
engineering applications while leveraging industrial investment and ra
pid growth in VLSI microprocessor technology. These processor architec
tures, while approaching supercomputer performance, were developed to
support the workstation market and incorporate policies consistent wit
h the needs of that user community. Large parallel computers differ fr
om workstations in that they exhibit high diameter and experience extr
emes in memory access latency. These architectures employ sophisticate
d cache structures to minimize average latency and achieve high appare
nt memory bandwidth. Unfortunately, many large scientific codes do not
exhibit memory access patterns with the necessary locality of referen
ce on which effective use of cache mechanisms relies. This presentatio
n examines the memory access patterns of applications from the Earth a
nd space sciences community and considers their implications for cache
based architecture.