Conclusions

It is a challenging problem to continuously devise new methods to hide the memory latency in view of the ever-increasing performance memory gap. In this study, we focused on whether it is possible to alleviate this problem either by having a good history-based cache management strategy or by intelligently controlling the cache parameters of a L2 cache. We found that no single scheme is good for all workloads, which points to some sort of an adaptive scheme or a profiling-based more cache-conscious data placement for programs with large running times. The hotornot scheme seems promising but it needs to be tested on more workloads. Also, in some cases, the difference in cache misses among policies was too small to warrant shifting to a more complicated strategy. Increasing the cache size mostly helps but it again depends on the active data set of the executing program. Similarly, increasing associativity mostly helps, but that needs to be balanced against increased hit time. In particular, we found that increasing associativity could potentially increase cache misses. As an aside, we also quantify the trade-off between cache misses and dirty writes in this study and try to incorporate the problem of dirty writes in the cache management algorithm.