Building Energy-Efficient Systems for Sequential I/O Workloads

  • John Davis ,
  • Suzanne Rivoire

MSR-TR-2010-30 |

Solid-state drives (SSDs) have the potential to replace traditional rotational hard disk drives (HDDs) as the main storage media for computer systems. The power and performance characteristics of SSDs differ from those of HDDs, requiring the designers of hardware systems to reevaluate how they build and tune their systems. In this paper, we design an SSD-based system for highly energy-efficient sequential I/O. Using this system, we break the current record for the 10 GB category of the JouleSort energy-efficiency benchmark by more than a factor of 2. Furthermore, a single SSD combined with an ultra-low power processor is able to beat the previous record by 10%. In addition, using a variety of hardware configurations, we contrast the impact of tuning on the power and performance of SSD-based systems with traditional HDD-based systems. We also demonstrate that by trading latency for power, we can achieve similar energy efficiency across a variety of systems, from embedded-class to desktop- and server-class systems. In other words, we show that energy-efficient computing does not require ultra-low-power components. Finally, we use this data to project the characteristics of an ideal energy-efficient data-intensive computing system using currently available components.