Research On LLC Replacement Policy For Heterogeneous Chip Multi-processors

The progresing integrated-circuit technique makes it possible that multiple processors can be integrated on chip. Compared with single-core processor, Chip Multi-Processors(CMP) enables the simultaneous execution of multi-threads, showing the obvious advantages of low complexity and low-power. Owning to the numerous computing units integrated in GPU, GPU is better at parallel computing ability. In order to give full play to the advantages of CPU and GPU, they are integrated on chip and constitute the CPU-GPU heterogeneous system, realizing their cooperation and acceleration.In the heterogeneous Chip multi-processors, the serious problem of “memory wall” becomes more prominent. Constructing the chip storage structure, especially the Last Level Cache(LLC) is an important technique method to offset the effect of “memory wall”. Moreover, how to effectively manage LLC resources is the key factor in determining the processor performance.Cache replacement policy is an important issue concerning the system structure research, the quality of the replacement policy will influence the overall system performance. Large amount of research on homogeneous CMP replacement policy have been done at home and abroad, but small on heterogeneous CMP replacement policy. Results achieved in this paper are as follows:The threads’ number of SPEC CPU2006 application and 22 kinds of GPU application is accumulated in this paper, as well as their number of access per-kilo instructions. Results are obtained after the comparison, which are as follows: on average, GPU application has a higher access rate of off-chip storage systems than CPU application; the access rate of GPU application is not directly in proportion to the number of threads. Based on the mentioned results, newly conclusions are reached via analysis on the heterogeneous CMP with LRU replacement policy: The traditional LRU replacement policy will become invalid in heterogeneous CMP.Based on conclusions above, this paper proposes the dynamic LLC replacement policy named DIPP(Dynamic Insertion/ Promotion Policy), which is suitable for heterogeneous CMP, according to the differences between GPU application and CPU application. DIPP dynamically changed the insertion policy and promotion policy of the GPU core and the CPU core. Via limiting Cache resources obtained by the GPU core, DIPP reduces the thread interference between applications and reaches the overall performance improvement. DIPP consists of the monitor OHAI(Online Heterogeneous Applications Identification) and the decision-making unit. OHIA is based on UMON, which enables the on-line identification and classification. While, OHIA makes DIPP become possible. On the basis of the OHIA classification, the decision-making is to pick out the most suitable replacement policy for this fragment application, realizing the best performance.This paper makes experiments in the Mac Sim simulator. Each heterogeneous CMP performance of DIPP and LRU has been contrasted and 10 GPU applications and 27 SPEC CPU2006 applications have been chosen. The result is that in the heterogenerous CMP with DIPP, the average performance of cache-friendly applications(arithmetic mean value) improved by 23.29%, large working-set applications improved by 13.95%, compute intensive applications improved by 9.66% and stream applications improved by 3.8%.It turns out that the heterogeneous CMP with DIPP replacement policy achieves better performance than heterogeneous CMP with LRU.