Tolerating First Level Memory Access Latency In High-Performance Systems

In order to improve performance, future parallel systems will continue to increase the processing power of each node in a system. As node processors, though, can execute more instructions concurrently, they become more sensitive to the first level memory access latency. This paper presents a set of hardware and software techniques, collectively referred to as register preloading, to effectively tolerate long first level memory access latency. The techniques include speculative execution, loop unrolling, dynamic memory disambiguation, and strip-mining. Results show that register preloading provides excellent tolerance to first level memory access latency up to 16 cycles for an issue 4 node processor. INTRODUCTION The objective of designing a high-performance system is to speed up the execution of application programs. An important approach to achieve this objective is to exploit program parallelism at both the instruction level and the multiprocessor level. For example, the Alliant FX/..