Gibraltar: GPU-Accelerated Software RAID for HPC
Anthony Skjellum, University of Alabama at Birmingham
Date: Thursday, Feb 2
Time: 2:30 PM
Location: 442 Dana, ECE Dept.
In this talk, we present the jointly developed UAB-Sandia Gibraltar library, that commoditizes RAID by first choosing software RAID, then performs middleware acceleration on that RAID using NVIDIA-based GPUs. The ability tocompute parity stripes at full rate with 2,3, 4, or more parity disks allowsstorage to become survivable for long periods of time, and particularly can address triple-disk failure and batch correlated failures. Gibraltar RAID is discussed, and the error rates that motivates this work are discussed as well. We are continuing to develop and extend this technology, which initially was designed to serve blocks via iSCSI and iSER. Current and future plans,including local striping for multi-cloud-vendor storage are mentioned. Current estimates are that practical petabytes of storage can be constructed for under $200,000, including a large amount of excess computational power, to enable "computation close to the storage." The general area of middleware acceleration through GP-GPU computing is also discussed.