Reducing traffic generated by conflict misses in caches

Off-chip memory accesses are a major source of power consumption in embedded processors. In order to reduce the amount of traffic between the processor and the off-chip memory as well as to hide the memory latency, nearly all embedded processors have a cache on the same die as the processor core. Because small caches dissipate less power and are cheaper than large caches, a small cache is preferable to a large cache. Furthermore, because set-associative caches consume more power than direct-mapped caches, a direct-mapped cache is preferable to a set-associative one. Small, direct-mapped caches generally incur many conflict misses, however. In this paper we propose and evaluate a structure called the Conflict Detection Table (CDT). This table can be used to determine if a memory access is expected to hit the cache. If a hit is expected and a miss occurs, then a conflict is detected and appropriate action can be taken. In addition, we propose two cache structures that employ this technique: the Bypass in Case of Conflict (BCC) cache and the Sub-block in Case of Conflict (SCC) cache. The BCC cache bypasses the cache when a conflict is detected, whereas the SCC cache fetches a sub-block of the missing cache block in such a case. Experimental results using several embedded workloads show that the BCC and SCC cache reduce the amount of traffic significantly in many cases. Furthermore, overall they incur the same number of cache misses as the direct-mapped cache. This shows that the BCC and SCC cache reduce the amount of power consumed with a negligible reduction in performance.