Verification of performance improvement of multi-plane operation in SSDs

Solid state drives (SSDs) provide significantly higher average performance than hard disk based storage by implementing various levels of parallelism internally. As a result, it quickly replaces hard disks in the server storage market. Parallel processing can be done in a channel level by simultaneously transmitting data through multiple channels, in a chip level by striping I/O requests to multiple chips on a page basis, in a die level by interleaving requests on multiple dies of a chip, and in a plane level by processing requests in multiple planes of a die at the same time. Although various policies have been proposed to maximize the parallel processing capability of SSDs, researches on the effect of the plane level parallel processing have not been sufficiently performed. This is because, unlike the channel level, chip level, and die level parallelism, in order to perform multi-plane operation, the types of operations must be the same, and the target pages of the planes must be located at the same position. Therefore, this study aims at verifying whether the multi-plane operation improves the overall performance of the SSD. Trace-driven simulation shows the followings. First, the wise policy that performs multi-plane operation only when the constraints of the multi-plane operation are satisfied always shows better performance than the policy that does not perform multi-plane operation. Second, the greed policy which performs the multi-plane operation with wasting clean pages even when the constraints of the multi-plane operation are not satisfied has excellent performance in the traces having the large requests. However, in the trace with a relatively small request size, its performance is lower than that of not performing multi-plane operation.