作者: Xian-He Sun , Dawei Wang
关键词: Shared memory 、 Computer engineering 、 Distributed memory 、 Flat memory model 、 Computer science 、 Uniform memory access 、 Memory management 、 Computing with Memory 、 Registered memory 、 Average memory access time 、 FLOPS 、 Parallel computing 、 Distributed shared memory 、 Memory map 、 Conventional memory 、 Cache-only memory architecture 、 Memory hierarchy 、 Interleaved memory 、 Extended memory 、 Computer Networks and Communications 、 Hardware and Architecture 、 Software
摘要: Due to the infamous "memory wall" problem and a drastic increase in number of data intensive applications, memory rather than processor has become leading performance bottleneck modern computing systems. Evaluating understanding system is increasingly becoming core high-end computing. Conventional metrics, such as miss ratio, average latency, access time, etc., are designed measure given parameter, do not reflect overall system. On other hand, widely used measurement IPC Flops CPU performance, directly performance. In this paper, we proposed novel metric, Access Per Cycle (APC), with consideration complexity A unique contribution APC its separation evaluation from evaluation; therefore, it provides quantitative "data-intensiveness" an application. The concept introduced; constructive investigation counting accesses cycles at differing levels hierarchy conducted; finally some important usages presented. Simulation results show that significantly more appropriate existing metrics evaluating