On Design and Implementation of a Large Virtual NVRAM Cache for Software RAID

In order to overcome the small write problem in RAID5, especially software RAID5, we have designed and implemented a software RAID with a large virtual NVRAM cache under the Linux kernel. Because no additional hardware is needed to implement our write cache, we named it Virtual NVRAM Cache or VC-RAID for short. The main idea is to use a combination of a small portion of the system RAM and a log disk to form a hierarchical cache. The log disk can be either a dedicated physical disk or several partitions of disks in the RAID. Since the log disk quickly absorbs write data from the small RAM, this hierarchical cache appears to the host as a large nonvolatile RAM. A prototype VC-RAID implemented under the Linux kernel has been tested for an extended period of time to show that it functions properly. Performance measurements have been carried out using typical programs and popular benchmarks such as ServerBench 4.1, PostMark and Bonnie. Our measurements show that the VC-RAID has superb performance advantages over the built-in software RAID shipped with the Linux package. Depending on the workload characteristics, performance gains due to VC-RAID range from 67.7% to an order of magnitude. For applications that have data locality, we observed up to a factor of 16 performance improvements in terms of user response time. In many situations, VC-RAID achieves similar performance as RAID0 and some time better than RAID0 indicating that VC-RAID realizes the maximum potential to hide small write problems in RAID5.