1. Efficacy of GPUs in RAID Parity Calculation 8/8/2007 Matthew Curry and Lee Ward Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000.

2. Disk bit error rates versus capacity –Is this being addressed at the disk lately? Now see 10E16 in new drives Commodity versus custom –Custom FPGA/ASIC or expensive graphics chip? OK, it's just a fun question Motivation

3. Environment 2 X AMD 246 Opteron @ 2 Ghz 2 GB memory NVIDIA GeForce 880 GTX CUDA 1.0 ('07) and 2.0 beta2 ('08)‏ Fedora Core 7 ('07) and Debian Lenny ('08)‏

4. Latency (8/8/07)‏

5. Throughput (8/8/07)‏

6. Issues CUDA is very primitive, very specific –Simultaneous compute/DMA not supported PCI transfer overhead is relatively large –Really need those asynchronous host-card transfers

7. Reimplemented 3+3 (6/26/08)‏

8. 13+3 (6/26/08)‏

9. 14+2 (6/26/08)‏

10. Currently Attempting to leverage GPU with Linux RAID code –CUDA runs only in user space –RAID code in kernel space Working at iSCSI target –Again, CUDA in user space but target in kernel Informal, separately funded, collaboration with Anthony Skjellum at UAB –Has grant supporting the work –Sandia funding requested of ASC/SIO for FY'09

11. Sandia resource and plans Purchased 1 SBOD for the test system listed earlier About to purchase NVIDIA GTX 280 –Did someone mention asynchronous transfers :-)‏ Partially motivated replacement of 8 ancient LSI RAIDs –With 8 SBODs and 4 dual FC4 Qlogic controllers Purchased 16 SATA FLASH-based drives for “battery-free”, reliable controllers investigation –Transactional updates, we're thinking Active-Active, coherent, pairs project via commodity HPC interconnect

12. Conclusion RAID error code compute/check performance is now competitive with top of the line Industry products Exciting, worthwhile, related projects –“Battery-free” RAID –Commodity HPC interconnect for cache coherency –Mitigation of peak power demand by delaying spin- up using FLASH disks? Potential cost savings –If we can manage to treat the core code as product and not a research project