End-to-End Congestion Management for Non-Blocking, Multi-Stage Switching Fabrics using Commodity Switches

We propose and evaluate end-to-end, proactive congestion management schemes for non-blocking multi-stage switching fabrics. All control functions are delegated to network adapters, which coordinate among each other through a variable-grain request-grant protocol. Switching elements simply forward packets, regardless to whether these correspond to payload data or control messages. We examine two different schemes. The first one merely regulates the long-term injection rate towards each output. For this scheme we show pathological states that may appear in practice, and conclude that it is stable only if some internal speedup is employed. The second scheme reserves buffer space inside the some switching elements before granting segments. Using a fluid model, we show that this scheme eliminates congestion in Clos/Benes fabrics for any number of stages, and detailed computer simulations demonstrate its robust operation without requiring speedup. Finally, for this distributed set-up, we show how to reduce the size of reassembly buffers, so as to fit them into SRAM buffers at the network adapters.

An shortened version of this paper has appeared in: Proc. 6th ACM/IEEE Symp. on Architectures for Networking and Communications Systems "ANCS 2010,' La Jolla, CA (ACM, October 2010)