Improving Oblivious Reconfigurable Networks with High Probability
share
Oblivious Reconfigurable Networks (ORNs) use rapidly reconfiguring switches to create a dynamic time-varying topology. Prior theoretical work on ORNs has focused on the tradeoff between maximum latency and guaranteed throughput. This work shows that by relaxing the notion of guaranteed throughput to an achievable rate with high probability, one can achieve a significant improvement in the latency/throughput tradeoff. For a fixed maximum latency, we show that almost twice the maximum possible guaranteed throughput rate can be achieved with high probability. Alternatively for a fixed throughput value, relaxing to achievement with high probability decreases the maximum latency to almost the square root of the latency required to guarantee the throughput rate. We first give a lower bound on the best maximum latency possible given an achieved throughput rate with high probability. This is done using an LP duality style argument. We then give a family of ORN designs which achieves these tradeoffs. The connection schedule is based on the Vandermonde Basis Scheme of Amir, Wilson, Shrivastav, Weatherspoon, Kleinberg, and Agarwal, although the period and routing scheme differ significantly. We prove achievable throughput with high probability by interpreting the amount of flow on each edge as a sum of negatively associated variables, and applying a Chernoff bound. This gives us a design with maximum latency that is tight with our lower bound (up to a log factor) for almost all constant throughput values.
READ FULL TEXT