Utilizing In-Network Buffering for Scheduling and Routing in Data Center Networks

In this paper, we aim to effectively utilize in-network buffering to schedule and route packets with low communication overhead, small delay and throughput optimality in fat-tree networks. While nearly zero in-network queueing can be guaranteed by performing precise time allocation and path assignment at network endpoints as in Fastpass, there is a high communication overhead, and buffer occupancy at the endpoints can become bottlenecks. By spreading scheduling functionalities to different network layers in a fat-tree network, the complexity can be decreased significantly at the cost of moderate buffer occupancy at intermediate switches. Inspired by this observation, we propose a simple dynamic pod scheduling (DPS) scheme, which performs scheduling at the granularity of pod units, each of which is paired to at most one other pod unit to transmit packets at each slot. By doing so, less information is required to arrange the packet transfers and inter-pod traffic will experience less downlink contentions. Through extensive evaluations, we find that DPS outperforms Fastpass in terms of delay while still guaranteeing throughput-optimality.