Scalability vs. Performance
Daniel Pressel · 2001
In the ideal world, the performance of a program running on a supercomputer would always be proportional to the peak speed of the system being used. Furthermore, the program would always achieve a high percentage of peak (e.g., 50% or better). In the real world, this is frequently not the case. Therefore, it is important to distinguish between the following five concepts: (1) performance (run time), (2) ideal speedup, (3) hard scalability (fixed problem size speedup), (4) soft scalability (scaled speedup), and (5) throughput (how long it takes to run a collection of jobs). This report addresses these concepts and explains their meanings and differences. Hopefully, this will allow readers to evaluate the behavior of programs and computer systems, and most importantly, to evaluate their own expectations for running a program on a particular system or class of systems. Examples, which demonstrate these concepts, are drawn from a variety of projects and include both problems from multiple computational technology areas (CTAs) and results from outside of the Department of Defense (DOD). In some cases, there will also be theoretical arguments to help better explain the issues.