In discussing petascale computing requirements, it is important to distinguish between classes of applications. There exist a limited set of interesting applications that can take advantage of less-expensive approaches to petascale computing. For example, large scale molecular dynamics codes can require relatively little communication per unit computation. For such codes, extrapolations of the Blue Gene architecture will cost effectively provide petascale performance. Likewise, for dense matrix problems and for highly structured sparse matrix problems, inexpensive SIMD or Vector accelerators can be combined with high volume commodity microprocessors in inexpensive low bandwidth networks to good effect in solving petascale problems.
For more general petascale computing, architectural balance is key. Such problems include most realistic engineering simulations, radiation-hydrodynamics and magneto-hydrodynamics for fusion accelerator design, scalable climate simulation, crash analysis, and graph-theoretic problems. These kinds of problems pose communications, load-balancing, memory bandwidth and latency requirements that are beyond the inexpensive architectural approaches mentioned above. In addition, data-centric computing is already driving costs upward as a premium is placed on high-performance I/O strategies.
I will outline one approach to petascale computing that provides high
architectural balance at near optimum cost while providing ease of porting
for the large body of existing MPI code. I will briefly critique proposed
programming models for petascale architectures.
Maintained by: Bernadette Watts
Modified on: February 13, 2006