Superways: A Datacenter Topology for Incast-heavy workloads

Several important datacenter applications cause incast congestion, which severely degrades flow completion times of short flows and throughput of long flows. Further, because most flows are short and the incast duration is shorter than typical round-trip times, reactive mechanisms that rely on congestion control are not effective. While modern datacenter topologies provide high bisection bandwidth to support all-to-all traffic, incast is fundamentally a many-to-one traffic pattern, and therefore, requires deep buffers or high bandwidth at the network edge. We propose Superways, a heterogeneous datacenter topology that provides higher bandwidth for some servers to absorb incasts, as incasts occur only at a small number of servers that aggregate responses from other senders. Our design is based on the key observation that a small subset of servers which aggregate responses are likely to be network bound, whereas most other servers that communicate only with random servers are not. Superways can be implemented over many of the existing datacenter topologies and can be expanded flexibly without incurring high cost and cabling complexity. We also provide a heuristic for scheduling jobs in our topology to fully utilize the extra capacity. Using a real CloudLab implementation and using ns-3 simulations, we show that Superways significantly improves flow completion times and throughput over existing datacenter topologies. We also analyze cost and cabling complexity, and discuss how to expand our topology.