Study Of Hardware Adptive Prefetch Technoligy Based On Application Pragram Memory Access Pattern

With unbalance speed of development between CPU and storage systems in modern microprocessor, execution of application program wait too long time for data to return, the performance of the storage system has become a bottleneck and degrades the computer system performance. Prefetching techniques can be inferred in advance of the next level of data loaded into the current level early in hierarchical memory system. Prefetching, to a certain extent, reduce the application memory access miss rate, improve the performance of the storage system, so it is widespread concern in the industry.Prefetching by implementation of points can be divided into software prefetch and hardware prefetch. Because the software prefetch has extend and promotion restrictions, this article focuses on the hardware prefetch, the main work is as follows:1. We analyze the application program in point of prefetching technique view. Discuss the cycle structure in application program that closely relate to memory access patterns. To SPEC CPU 2006, for example, the overall execution of the test program were analyzed, and for a few specific test program, we analysis of their working set features, two level Cache memory access, the data characteristics in loop that called by application program most and basic memory access pattern mode.2. L1 D Cache prefetch unit is based on stream prefetch. We use Stream Length History Table to real-time record the length of stream memory access imformation that can determine the number of prefetched data and timely stop prefetching. This structure can filter out some unnecessary prefetch.3. L2 Cache prefetch unit is baased on stride prefetch. We use Memroy Access Map to record application progam memory access state imformation in space, and use these kinds of information to tolerate interference of disorder execution and calculate prefetch data address.4. L2 Cache Prefetch real-time monitor memory system on if prefetch is successful, the number of resource consumption and resource congestion, to determine the iuuse of the prefetched data.Finally, we implement such prefetch structure in CMP$im simulator, using SPEC CPU 2006 test this prefetch structure. Through experiments, we found that this adaptive prefetch structure utilizes more real-time information than traditional prefetch structure; hardware overhead is not high in the premise than traditional prefetch structure advantage in performance.