The impact of skewed associativity on the replacement traffic in COMA multiprocessors

Skewed associative caches use different hash functions for their different banks. This property causes this kind of cache to reduce the rate of conflict misses. It has been shown that exploiting skewed associative caches to build the attraction memory of a COMA (Cache Only Memory Architecture) multiprocessor improves the utilization of the attraction memory space and allows data items to be replicated more freely. The obvious result is that the use of these memories can reduce the replacement traffic by reducing the rate of conflict misses in the attraction memory. However, there is a need for a precise analytical model that can predict the direct impact of skewed associative attraction memories on the replacement traffic. Such a model can help to predict and compare the replacement traffic in the absence or presence of the mentioned memories. This paper analytically models the direct impact of skewed associative attraction memories on the replacement traffic in COMA multiprocessors and then uses simulation results to verify the correctness of the derived model. The analytical model and simulation results both show that the replacement traffic is reduced significantly if the attraction memory is built based on the skewed associative architecture.