Expoiting Critical Data Regions to Reduce Data Cache Energy Consumption

Ananda Vardhan, Y.N. Srikant

In this report, we address the problem of power consumption in a data cache by focusing on the latency tolerance of data regions. Latency tolerance of a data region indicates how critical the data region to the overall performance of the program. We use a profile- based technique focusing on context-as well as path-sensitive analysis to identify critical and non-critical data regions. We deploy criticality analysis to drive a power-aware optimization technique which allocated a split data cache, operating in normal and drowsy modes, to critical and non-critical data regions respectively. This technique saves around 30% of total power and 20% of leakage power in the data cache without any significant performance penalty.

pdf