Non-Vital Loads

It is a well-known fact that cache hits are vital to superscalar processor performance. The successful execution of a load allows resources to be freed up and it also allows dependent instructions to proceed to execution. As processor cores operating frequencies rise rapidly, the cache hierarchy is struggling to keep up in terms of speed while still trying to provide good hit rates, because the two technologies are progressing at vastly different paces. The problem could be partly alleviated by not blindly putting data into the cache. We find that not all loads are vital to the performance of the processor. In this work we present a theoretical study of various possible categories of loads and scrutinize their vitality. We term these loads as non-vital loads. The data corresponding to these non-vital loads could be removed from the group competing for vital cache space. We found that 75% of all dynamic loads are nonvital and not critical to the performance of the processor for the Spec95 Suite. We followed this by increasing all 75% of their latencies and witnessed negligible performance degradation.