Abstract

Server storage systems are often built using a large number of disks to meet the performance and capacity demands of data-intensive applications. However, such large storage systems can consume a significant amount of power, thereby increasing the power and cooling costs of a data center. In this talk, I will present a novel disk drive design, called “intra-disk parallelism”, which can facilitate building high-performance, low-power enterprise storage systems. Intra-disk parallelism extends the conventional hard disk drive architecture by: (i) decoupling the way that the spindle and arm-assembly of a disk drive are used to service I/O requests, so that we can overlap disk seeks with rotational delays, and (ii) decoupling the multiplicity of the components within each of these two electro-mechanical systems to further enhance parallelism.

I will first provide a historical retrospective on intra-disk parallelism, discussing the similarities and key differences between our approach and the multi-actuator drives of the past. I will present an overview of the design space of intra-disk parallelism, identifying the locations within a disk drive where parallelism can be incorporated. Using a set of commercial workloads, I will provide an analysis of the performance and power characteristics of a specific design within this space and show that storage arrays built using such drives consume 40%-60% less power while delivering performance that is comparable to arrays built using conventional disk drives. Finally, I will discuss the key engineering and cost issues involved in building intra-disk parallel drives and show that intra-disk parallelism can be a practical approach to build energy-efficient enterprise storage systems.