CCoW: Optimizing Copy-on-Write Considering the Spatial Locality in Workloads

Abstract

Copy-on-Write (CoW) is one of the most essential memory management techniques enabling efficient page sharing between processes. Specifically, combined CoW with the fork system call, applications, even with a huge memory footprint, can take a snapshot of the current in-memory data at low overhead. However, since the CoW takes place per page in the page fault handler, each time the page fault occurs, the operating system should get involved. This leads to significant performance degradation for write-intensive workloads. This paper proposes coverage-based copy-on-write (CCoW), an optimized CoW scheme considering the locality in memory accesses to mitigate the problem of CoW. CCoW measures the spatial locality in process address spaces with the concept of coverage. While processing CoW, CCoW copies multiple pages in advance for high-locality memory regions, thereby minimizing the involvement of OS for write-intensive workloads. We explain the challenges for measuring the locality and provide the optimization to implement the concept. Evaluation with a prototype demonstrates that this approach can improve the overall performance of applications by up to 10% with a small amount of memory overhead.