An HTMT Performance Prediction Case Study: Implementing Cannon's Dense Matrix Multiply Algorithm

This technical memo presents a case study of performance prediction for the Hybrid Technology Multi-Threaded architecture. We use a dense matrix multiply to introduce an analytical methodology to predict the performance of the percolation model on the HTMT architecture. HTMT introduces a percolation program and execution model that (1) is explicitly multi-threaded; (2) incorporates global memory address space; and (3) explicitly exposes the HTMT memory hierarchy to the programmer. The percolation model extends dynamic prefetching to allow the management of contexts that include data, program instructions, and control states. An analytical study of our algorithm and the percolation process is used to determine the number of operations that are performed in each memory region and the amount of data that is exchanged between regions. Current estimates for the processing power, network performance and storage capacity in each memory region are injected into the analytic study to predict th..