Cost-Sensitive Cache Replacement Algorithms

Cache replacement algorithms originally developed in the context of simple uniprocessor systems are aimed at reducing the aggregate miss count. However, in modern systems, cache misses have different costs. The cost may be latency, penalty, power consumption, bandwidth consumption, or any other ad-hoc property attached to a miss. Thus, in many practical cases, it is desirable to inject the cost of a miss into the replacement policy in addition to access locality. In this paper, we propose several extensions of LRU which account for non-uniform miss costs. These LRU extensions have simple implementations, yet they are very effective in various situations. We evaluate our extended algorithms using trace-driven and execution-driven simulations. We first explore the simple case of two static miss costs using trace-driven simulations to understand when cost-sensitive replacements are effective. We show that very large improvements of the cost function over LRU are possible in many practical cases. Then, as an example of their effectiveness, we apply the algorithms to the second level cache of a multiprocessor with ILP processors, using the miss latency as the cost function. We show that by applying our simple replacement policies sensitive to the latency of misses we can improve the execution time of parallel applications by up to 18% using execution-driven simulations with RSIM.