|
Tera benchmarked a code called TB2 written by Dr. Timothy J. Barth, a scientist at the NASA Ames Research Center. TB2 is a computational fluid dynamics code that is used to benchmark how well competing computer systems can handle unstructured, dynamically changing problems that typify aircraft or space shuttle aerodynamic simulations. In the past, this problem has run most effectively on one processor of a vector supercomputer. Tera believes its time on this benchmark problem is the fastest ever recorded. Extremely difficult issues of memory latency (the clinical term for the Latency Bug) and parallel task synchronization have prevented this code from being run effectively in parallel on competing systems. Though designed for a vector uniprocessor, TB2 required few code changes to run well on a multiprocessor MTA, Tera claims. The most significant change was to unplug the standard library solver, replacing it by a parallel direct solver developed at Tera. This change is transparent to the application in that the computed numerical results are identical. According to the benchmark results, A T90 processor can perform as many as 1.7 billion arithmetic operations and access as many as 2.5 billion memory locations per second. A single MTA processor is limited to less than half this arithmetic rate and only about one-tenth the T90's memory rate. Yet just one MTA processor alone surpasses a T90 in performance when solving TB2's linear systems. Four MTA processors together achieve a 2.9-fold speed-up for the series of compute-intensive Newton-solve iterations performed on this small benchmark input.
|
|