Title: Using GPUs to Enable Highly Reliable RAID
Speaker:
Matthew Curry, University of Alabama at Birmingham
Date/Time: Tuesday, May 18, 2010, 1:00 pm
Location: CSRI Building/Room 279 (Sandia NM)
Brief Abstract: Creation of large, highly available RAID arrays is a difficult and expensive proposition. The challenges are numerous, but two main factors hinder the scalability of today's storage resources. First, hardware-based RAID controllers are expensive, increasing the cost of storage per gigabyte dramatically to ensure reliability and performance.
Second, while RAID 6 is the industry standard for reliability, other RAID configurations with increased fault tolerance are within the capabilities of today's hardware. Vendors generally do not provide these options, instead choosing to support more traditional hierarchical RAID organizations that increase hardware requirements and only marginally increase reliability. This decision reduces the amount of usable space that can be supported by a single controller.
This talk outlines the benefits and challenges of a prototype RAID system which uses a GPU-based implementation of Reed-Solomon coding to provide erasure correction. By exploiting certain aspects of GPU architectures, this implementation has been shown to support RAID 6 (and other enhanced RAID levels) at line speed for large disk arrays. The increased flexibility of a software implementation is able to provide non-traditional RAID organizations (including multi-level RAID) among several storage servers, or lend erasure correcting capabilities to other related non-RAID applications.
CSRI POC: Lee Ward, 505-844-9545 |