Impact of hypervisor cache locking on performance

Abstract

Server virtualization has become the modern trend in various industries. Industries are resorting to the deployment of virtualized servers in order to cut the cost of additional, expensive hardware and consolidate the servers on minimal hardware for easier management and maintenance. Virtualized servers are often seen connected to disk arrays in many industries ranging from small to medium business to large data centers. In such a setup, assuring a low latency of data access from the disk array plays an important role in improving the performance and robustness of the overall system. Caching techniques have been researched and used in the past on traditional processors to reduce the number of memory accesses and have proven benefits in alleviating the response times of applications. The research done in this paper explores caching on the hypervisor and analyzes the performance of data cache locking technique in hypervisor caches. The research aims at reducing the Input / Output (I/O) latency in a server virtualized Storage Area Network (SAN) setup, which thereby increases the performance of applications running on the virtualized servers. The authors introduce a miss table that can determine the blocks of data in the hypervisor cache that need to be locked. Way cache locking method is used for locking, such that only selected lines of cache are locked (not the entire cache). The proposed cache locking technique is later evaluated with the introduction of a small victim buffer cache and probability based cache replacement algorithm. Valgrind simulation tool was used to generate memory traces by virtual machines (VMs). Experimental results show an improved cache hit percentage and a considerable reduction in the I/O response time due to the proposed cache locking technique when compared to the results without hypervisor cache locking.