April 22, 2015
Shingled Magnetic Recording (SMR) drives have been introduced to the market in the last two quarters, and based on industry projections and the applicability of SMR to future recording technologies (i.e. HAMR, BPM) it is likely that magnetic drives over 10TB in capacity will exclusively use SMR for the foreseeable future. Under these circumstances, storage vendors must learn how to modify existing and future systems to the new performance characteristics and interfaces of these devices. This work proposes the following two lines of investigation: 1) Modeling and prediction of SMR drive performance from workload characteristics. The goal of this work will be to develop (a) drive models based on previously-developed characterization methodology and (b) workload metrics and measurement techniques which may be used together in order to predict performance of existing workloads on shingled drives. 2) Storage stack optimization for SMR. Investigate modifications at layers ranging from the STL to the file system for maximizing the performance of SMR-based storage systems. A key goal of this work is the investigation of the differential between host-managed and drive-managed device operation: what are the conditions for performance in the two cases to be equivalent, and absent those conditions, what is the difference in achievable performance between the two?