Hardware Performance Monitoring in Multiprocessors

Multiprocessors are often quoted as being capable of a ‘peak performance,’ but in practise it is difficult to utilize this potential. Consequently, software applications must be well tuned to run efficiently. In this thesis, factors affecting performance in cache-coherent multiprocessors, particularly those which use a sequentially-consistent memory model, are analyzed. Cost-effective hardware features are suggested to measure these factors without intruding on software performance. Many of the proposed features can be readily incorporated into future processor designs, and others are easy to implement in external hardware. The measurements can be used to help tune a program’s performance as well as other purposes such as simulation verification, workload characterization, or runtime performance decision support. To demonstrate cost feasibility and other implementation concerns, the processor card hardware performance monitor developed for the NUMAchine multiprocessor is also described.