Yee, Ming Chung (2019) Reused Frequency-Based Replacement Policy With Program Counter Predictor On Various Memory Access Types For Last Level Cache Memory. PhD thesis, Universiti Sains Malaysia.
|
PDF
Download (983kB) | Preview |
Abstract
Currently, the main trend of microprocessor cache research is on the last level cache (LLC) memory. The typical LLC memory size is between 2 MB to 20 MB. The researches focus on improving the performance of LLC memory in terms of instruction per cycle (IPC) by using replacement policy. Static Re-reference Interval Prediction Hit Priority (SRRIP-HP) replacement policy can evict cache line that is frequently used and might be reused very soon. In this thesis, a new replacement which uses reused frequency and Re-reference Interval Value (RRPV) is proposed to mitigate the problem of SRRIP-HP with the aim to improve performance in terms of IPC. This replacement policy is called as Reused Frequency Promotion (RFP). However, the proposed RFP evicts newly inserted lines too early on recency-friendly benchmarks. This is because RFP inserts the reused frequency and RRPV in static manner without knowing which lines need to be kept longer. Therefore, a new dead block predictor called Program Counter Hit Miss (PCHM) Predictor is use to modify the insertion decision of RFP. The PCHM Predictor allows line that is predicted to be reused in future very soon to be stored longer in LLC memory. This can increase its chance of being reused and improve performance. The combination of RFP and PCHM Predictor is called RFP-PCHM. However, RFP-PCHM performs worse on some benchmark due to different type of memory access are updating the PCHM Predictor. Each type of memory access namely load/RFO, prefetching and writeback has different reused behavior. Therefore, the updating and insertion decision of PCHM predictor from RFP-PCHM is modified; and the promotion decision of RFP-PCHM is also modified based on the type of memory access namely load/RFO, prefetching and writeback. This replacement policy is called RFP-PCHM-I. In the performance evaluations, the proposed RFP, RFP-PCHM and RFP-PCHM-I are tested at single-thread, multi-thread and multi-program evaluation using SPEC CPU 2006 and Cloudsuite benchmarks on ChampSim simulator. Least Recently Used (LRU) replacement policy is used as baseline. LRU evicts cache line that is the least recently used in LLC memory. The RFP, RFP-PCHM and RFP-PCHM-I outperform LRU with 0.72%, 0.77% and 2.52% respectively in terms of weighted geometric mean of normalized IPC. RFP-PCHM-I also overall outperforms the Signature-based hit predictor ++ (SHiP++) and Leeway. In conclusion, the reused frequency of RFP have further improved the performance of SRRIP-HP. In addition, the PCHM predictor of RFP-PCHM have further improved the performance of RFP. The modification of updating and insertion decision of PCHM predictor; and promotion decision of RFP-PCHM have further improve the performance of RFP-PCHM.
| Item Type: | Thesis (PhD) |
|---|---|
| Subjects: | T Technology T Technology > TK Electrical Engineering. Electronics. Nuclear Engineering |
| Divisions: | Kampus Kejuruteraan (Engineering Campus) > Pusat Pengajian Kejuruteraaan Elektrik & Elektronik (School of Electrical & Electronic Engineering) > Thesis |
| Depositing User: | Mr Mohamed Yunus Mat Yusof |
| Date Deposited: | 19 Jan 2022 07:00 |
| Last Modified: | 19 Jan 2022 07:03 |
| URI: | http://eprints.usm.my/id/eprint/51226 |
Actions (login required)
 |
View Item |
