Hybrid Caching for Chip Multiprocessors Using Compiler-Based Data Classification
Yong Li · D-Scholarship@Pitt (University of Pittsburgh) · 2011
The high performance delivered by modern computer system keeps scaling with an increasingnumber of processors connected using distributed network on-chip. As a result, memory accesslatency, largely dominated by remote data cache access and inter-processor communication, is becoming a critical performance bottleneck. To release this problem, it is necessary to localize data access as much as possible while keep efficient on-chip cache memory utilization. Achieving this however, is application dependent and needs a keen insight into the memory access characteristics of the applications. This thesis demonstrates how using fairly simple thus inexpensive compiler analysis memory accesses can be classified into private data access and shared data access. In addition, we introduce a third classification named probably private access and demonstrate the impact of this category compared to traditional private and shared memory classification. The memory access classification information from the compiler analysis is then provided to the runtime system through a modified memory allocator and page table to facilitate a hybrid private-shared caching technique. The hybrid cache mechanism is aware of different data access classification and adopts appropriate placement and search policies accordingly to improve performance. Our analysis demonstrates that many applications have a significant amount of both private and shared data and that compiler analysis can identify the private data effectively for many applications. Experimentsresults show that the implemented hybrid caching scheme achieves 4.03% performance improvement over state of the art NUCA-base caching.