Hardware-Assisted Multiprocessor Performance Measurements

摘要

The report describes the implementation and use of a hardware-assisted trace measurement system (TRAMS) used to obtain performance measurements of parallel cooperating processes executing on a multiprocessor computer. The benefit of TRAMS is that the overhead required to obtain timing information is approximately two orders of magnitude better than the standard system call, thus providing more accurate results with minimum perturbation to the measured processes. The level of accuracy allows measurement of fine-grain portions of these parallel processes which cannot be reasonably measured using standard techniques, and are therefore usually presented as negligible. Some measurements that have been obtained using TRAMS on a tightly-coupled, shared-memory parallel processor are reported here and include basic programming constructs, process creation, process synchronization, and shared memory allocation.