Active storage revisited: the case for power and performance benefits for unstructured data processing applications

  • Authors:

  • Clinton Wills Smullen, IV;Shahrukh Rohinton Tarapore;Sudhanva Gurumurthi;Parthasarathy Ranganathan;Mustafa Uysal

  • Affiliations:

  • University of Virginia, Charlottesville, VA, USA;Lockheed Martin, Cherry Hill, NJ, USA;University of Virginia, Charlottesville, VA, USA;Hewlett Packard Labs, Palo Alto, CA, USA;Hewlett Packard Labs, Palo Alto, CA, USA

  • Venue:
  • Proceedings of the 5th conference on Computing frontiers
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

The proliferation of digital data has resulted in a mushrooming of data-intensive applications, especially in the area of unstructured data processing. Given the growing popularity of unstructured data processing applications (e.g., FlickrTM, Google MapsTM), it is important to rethink system architectures to efficiently run these applications, from both the performance and power viewpoints. In this paper, we revisit active storage, which proposed offloading computation to disk drive processors, as a possible system architecture for these applications. Unlike previous work, we evaluate the microarchitectural aspects of active storage and perform an in-depth examination of the design of the offload processors. Using a set of unstructured data processing benchmarks, we examine two choices along the I/O path where the computation can be offloaded in existing system architectures -- a disk drive processor and a disk array controller. Our evaluation demonstrates that there are interesting tradeoffs in the choice of each location and that microarchitectural enhancements to these processors can provide significant performance boosts. We show that active storage architectures can provide large power savings, by using lower-power processors along the I/O path, while exploiting the data-level parallelism on the storage side of the system.