Software-Oriented Distributed Shared Cache Management for Chip Multiprocessors

This thesis proposes a software-oriented distributed shared cache management approach for chip multiprocessors (CMPs). Unlike hardware-based schemes, our approach offloads the cache management task to trace analysis phase, allowing flexible management strategies. For single-threaded programs, a static 2D page coloring scheme is proposed to utilize oracle trace information to derive an optimal data placement schema for a program. In addition, a dynamic 2D page coloring scheme is proposed as a practical solution, which tries to ap- proach the performance of the static scheme. The evaluation results show that the static scheme achieves 44.7% performance improvement over the conventional shared cache scheme on average while the dynamic scheme performs 32.3% better than the shared cache scheme. For latency-oriented multithreaded programs, a pattern recognition algorithm based on the K-means clustering method is introduced. The algorithm tries to identify data access pat- terns that can be utilized to guide the placement of private data and the replication of shared data. The experimental results show that data placement and replication based on these access patterns lead to 19% performance improvement over the shared cache scheme. The reduced remote cache accesses and aggregated cache miss rate result in much lower bandwidth requirements for the on-chip network and the off-chip main memory bus. Lastly, for throughput-oriented multithreaded programs, we propose a hint-guided data replication scheme to identify memory instructions of a target program that access data with a high reuse property. The derived hints are then used to guide data replication at run time. By balancing the amount of data replication and local cache pressure, the proposed scheme has the potential to help achieve comparable performance to best existing hardware-based schemes.Our proposed software-oriented shared cache management approach is an effective way to manage program performance on CMPs. This approach provides an alternative direction to the research of the distributed cache management problem. Given the known difficulties (e.g., scalability and design complexity) we face with hardware-based schemes, this software- oriented approach may receive a serious consideration from researchers in the future. In this perspective, the thesis provides valuable contributions to the computer architecture research society