The Shared Cache Dynamic Partitioning Algorithm And Vlsi Implementation Study

With the rapid development of information technology, the application field of computer becomes more and more wide, which is now covering almost every aspect of people daily life. However, the development pace of the performance of uniprocessor has been slowed down gradually due to the technical difficulties in power consumption, instruction-level parallelism (ILP) and the memory latency. Currently, multiprocessors are replacing uniprocessors and play a major role. Sustained and quick advances in performance of processors also make the gap between memory and processors even big. Therefore, a more effectively memory subsystem can significantly enhance the performance of the whole processor.In the multi-processors architectures, the last level cache is usually shared among several cores. In this kind of architecture, one core request may evict the block that frequently accessed by the other cores in the shared cache, which results in the system performance degradation. To improve the system performance, dynamic cache partitioning is used in the shared cache, which can avoid the interaction between different applications.In addition, the LRU method is not convenient for hardware implementation, so the caches in many processors use the pseudo-LRU replacement policies at present, which have lower hardware overhead than the LRU policy. Because the multi-core interaction still exists in the shared cache which uses pseudo-LRU method. Therefore, a novel dynamic cache partitioning strategy that based on the pseudo-LRU policies (PLRU-SCP) is proposed in this paper.The presented partitioning strategy has a novel statistical method based on the binary tree in the analysis circuit. To meet the properties of this method, a non-exhaustive algorithm is introduced in the partitioning circuit. The partitioning circuit controls the shared cache intermittently, avoiding the disadvantage of continuous control. The shared cache adopts a new structure applying to the dynamic partitioning. Furthermore, a novel replacement policy which is based on the Not Recently Used (NRU) policy is put forward for this structure.The evaluation, by running SPEC CPU2000benchmarks, shows that on average the shared cache with PLRU-SCP strategy presented in this paper can improve the performance by11.20%over the cache without the partitioning strategy using LRU policy (LRU-NP) and6.93%over the cache with the utility-based partitioning strategy using LRU method (LRU-UCP). Finally, the PLRU-SCP and LRU-UCP strategies are implemented in hardware. The synthesis results indicate that under the PLRU-SCP strategy, the area of analysis, partitioning and shared cache circuit can obtain17.46%,27.05%and12.54%reduction compared to their respective area under the LRU-UCP strategy.