Standards-based hybrid probe instrumentation for multiprocessor computing systems

Hybrid performance monitoring provides an unobtrusive and relatively inexpensive method of evaluating and tuning parallel hardware and software performance. In this paper, the design of standards-based hybrid performance monitoring hardware is examined for physical nodes containing a cluster of four processors. The design incorporates a NIST developed MultiKron probe acquisition chip and an interface design is presented which is suitable for multiprocessor systems and efficiently avoids probe consistency problems. It provides a queue in separate hardware and provides a dedicated hardware arbiter to eliminate the need for semaphores. This reduces the software overhead of sending performance data and provides a deadlock-free method of communication