While InfiniBand was heralded a number of years ago as a means of creating seamless server and storage switched fabrics, the InfiniBand protocol has lost its initial traction in the high-performance-computing space as a means of lashing together nodes in massively parallel Linux clusters.
The OpenIB Alliance members say they will now focus on developing upper-level InfiniBand protocols and APIs. The organization will be hosting a developers’ workshop from February 6 through 9 in Sonoma, California, to hammer out the proposed OpenIB software architecture, most notably including Remote Direct Memory Access (RDMA) over InfiniBand, which would link the main memories of nodes in a cluster directly to each other – rather than passing through cache and I/O buses – and thereby more tightly coupling nodes together while reducing latencies in data transfers between nodes.
