Characterizing the Performance Space of Shared Memory Computers Using Micro-Benchmarks

We have been investigating alternatives to conventional benchmarking that will allow users and machine designers to characterize the performance space of the memory hierarchy and the interconnection network of different shared memory machines in a uniform way, while providing a level of detail that is usually associated with hardware monitors. Our approach uses micro benchmarks (small benchmarks each measuring a particular aspect of the system under controlled conditions) and the concept of a Multiprocessor Memory Reference Pattern (M-MRP). A M-MRP is a family of memory reference sequences generated by a subset of processors. Each M-MRP is characterized by 5 different attributes: 1) the number of processors generating addresses; 2) the region of space covered by each processor; 3) the distance between consecutive addresses; 4) the degree of contention on the data; and 5) the data dependencies between the processors. By varying each of these five dimensions we can exercise the system and the individual components in a specific and predictable way. We then measure the performance response of the system to a particular micro benchmark. Using our methodology we have been able to obtain detailed measurements about the performance characteristics of the memory hierarchies and interconnection networks of the KSR1 and Stanford DASH under many different conditions. Our results clearly show for each machine how the network latency changes in response to different communication patterns and degree of data contention.