Infiniband's Role in Next-Gen Data Centers

I received one of my favorite embedded systems trade rags in the mail yesterday. The theme of this month's issue is "InfiniBand, the next generation of networking." Infiniband seems to have been christened the preferred datacenter interconnect by the networking industry. It works around an architecture called RDMA (remote direct memory access) -- which basically allows the physical components of nodes in a fabric to talk directly, without having to go through the overhead of the TCP stack or bog down the operating system.

Insiders sort of pigeonhole Infiniband's value into high performance computing environments. But I believe that as enterprise continues to evolve to "scale-out" environments (commodity hardware virtualized to behave like big SMP or mainframe systems) -- these discussions around the communication (and latency, I/O throughput, performance issues) between networked servers within a datacenter will become more mainstream. Infiniband will certainly play a huge role in Grid environments (and already does, in e-Science Grids).

So virtualization functionality is rapidly moving down the stack into the network. These initial discussions about data center interconnect and physical network performance are just the tip of the iceberg in terms of networking discussions for Grid.

The networking players aren't just debating the preferred interconnect between servers within a datacenter -- they're debating the preferred network topology for connecting Grids.