Application Directed Explicit Management for Advanced Cache Architectures

In this paper, we demonstrate the effectiveness of application directed explicit cache management. We define the generalized split temporal/spatial cache architecture as an abstraction of several advanced cache architectures. We analyze individual problems, identify the inefficiencies in the memory hierarchy and develop explicit cache management algorithms. In our algorithms, the application software controls hardware mechanisms directly. To illustrate various optimizations, problems are chosen from regular, sparse, data structure and graph applications. Analytical performance estimations are derived for several problems. Simulations show reduced memory traffic and improved average memory access time. For example, in the sparse matrix vector multiplication problem, the average memory access time can be reduced by 21% to 62% over a broad range of cache configurations.