Performance of load balancing algorithms in Clos packet switches

The size of a single-hop crossbar fabric is still limited by the technology, and the fabrics available on the market do not exceed terabit capacity. A multihop fabric, such as the Clos network, provides higher capacity by using smaller switching elements. When the traffic load is balanced over the switches in a middle stage, all the traffic can get through the fabric, as long as the switch outputs are not overloaded. However, the delay that packets experience through the Clos switch depends on the granularity of flows that are balanced. We examine the maximum fabric utilization under which a tolerable delay is provided for various load balancing algorithms, and derive its general formula in terms of the number of flows that are balanced. We show that the algorithms which balance flows with sufficiently coarse granularity provide both high fabric utilization and delay guarantees to the most sensitive applications.