XFabric: reconfigurable network topologies at rack scale
Wednesday, April 1, 2015Sergey Legtchenko, Microsoft Research Cambridge
Systems like the AMD SeaMicro, HP Moonshot and Boston Viridis are commercially available high density clusters with networking, compute and storage in a single chassis. They demonstrate techniques that will be used in datacenters to scale to a thousand of servers per rack, as opposed to about 40 today. In particular, the network is scaled by replacing the traditional Top of Rack switch by a distributed switching fabric in which each CPU is co-located with a (packet) switch, at the extreme on the processor silicon as a system-on-a-chip (SoC). These switches are interconnected creating multi-hop topologies (e.g., a 3D Torus).
Distributed fabrics are not fully provisioned: they do not have full bi-section bandwidth and the latency between servers is a function of the number of hops. There is no optimal topology for all workloads and the static topology design impacts workload performance. The talk will describe XFabric, a reconfigurable rack-scale network for SoCs with embedded packet switches. XFabric is a packet-switched network running over a physical circuit-switched network. The circuit-switched network allows the physical topology of the network to be dynamically changed to adapt to workload traffic.
Orateur : Sergey Legtchenko, Microsoft Research Cambridge
More details here …
marc.shapiro (at) nullacm.org