Server OEM Panel OFA Sonoma Workshop Bob Souza, HP 15 March 2010 www. openfabrics. org 1
Usual Talk 1. Where we use OFED – High-Performance Clusters • From big to small. Mostly Linux. Multiple kernels, distributions, interconnect vendors – Limited use in HP-UX (SRP, RDS) – FSI market data, messaging – Storage (HP SFSg 3, X 9000 NAS) 2. Whine about – Distros – Vendor-specific releases – Support
Last Year • 10 Gb. E on motherboards • Big, flat 10 Gb. E switches • CEE/DCB enabling – Converged Ethernet – 10 Gb. E RDMA alternatives
RDMA Choice • Rich RDMA ecosystem - customer chooses – IB – i. WARP – Ro. CEE, aka ((IBx? )|R|RDMA)o(CE)? E • Buy the performance you need
Nurturing the RDMA Ecosystem • HPC becoming mainstream – Cores, Ethernet • Migrate 1 Gb. E community to 10 Gb. E – The 50% of TOP 500 on 1 Gb. E – The hoi polloi who get 10 Gb. E LOM for ‘free’ – Good 10 Gb. E performance data helps everyone • Make it easy
Nurturing the RDMA Ecosystem Leveraging CNAs – 10 Gb. E – FCo. E + RDMA • Ro. CEE • i. WARP
What’s Important IB Cost of large fabrics can be an issue – Fat-tree alternatives for cost reduction • Pruned trees • Torus • Etc.
What’s important • Customer Choice • Interoperability – More than ever – Vendor responsibility • Simplified customer interactions, support – Drives adoption – More code downstream helps – It should ‘just work’
www. openfabrics. org 9
Скачать презентацию Server OEM Panel OFA Sonoma Workshop Bob Souza
6247bb8e2d7a5b734c3ed57f3573c108.ppt