Associativity Revisited -a Study of Set, Column, and Skewed Associative Tlbs Using Spec95

This paper describes the application of two small-set cache polices to hardware and software managed TLBs. The column-and 2-way skewed-associative policies as applied to TLBs are introduced which disperse cache accesses to avoid connict misses in an eeort to maximise the lifetime of useful entries. Long length trace driven simulation using all 18 SPEC95 benchmarks has shown that while both policies utilise a two entry set, the column policy had on average a 35% improvement in miss rate over the 2-way set-associative. Some benchmarks showed a 90% improvement. The average improvement represents 75% of the average improvement shown by 4-way set-associative over 2-way. The 2-way skewed policy on average outperformed 4-way set-associative by 13.9% where the top six benchmarks tested averaged 75% and ranged from 50% to 90% improvement. In addition, these policies are as space-eecient as 2-way set-associative methods. A number of researchers argue that large size (number of TLB entries) small associative TLBs should be built in preference to smaller fully-associative designs. We have conducted extensive simulations which compare a fully-associative policy with the introduced column-and skewed-associative TLB polices. The comparison trades associativity for entries to maintain similar TLB silicon resource requirements. We show that these placement policies can achieve improvements on miss rates of up to and over 99% of the fully associative TLBs. On average the column policy showed 26% and the skewed policy showed 45% improvement for the 4:1 entry ratio simulated. The performance advantages of the set dispersion techniques studied show a clear case for their consideration for small-set TLB management but is it applicable to software secondary TLB management. We discuss the issues in using both policies for software secondary TLBs and conclude the skewed-associative policy is certainly applicable.