Delay versus Stickiness Violation Trade-offs for Load Balancing in Large-Scale Data Centers

Abstract

Most load balancing techniques implemented in current data centers tend to rely on a mapping from packets to server IP addresses through a hash value calculated from the flow five-tuple. The hash calculation allows extremely fast packet forwarding and provides flow `stickiness', meaning that all packets belonging to the same flow get dispatched to the same server. Unfortunately, such static hashing may not yield an optimal degree of load balancing, e.g., due to variations in server processing speeds or traffic patterns. On the other hand, dynamic schemes, such as the Join-the-Shortest-Queue (JSQ) scheme, provide a natural way to mitigate load imbalances, but at the expense of stickiness violation. In the present paper we examine the fundamental trade-off between stickiness violation and packet-level latency performance in large-scale data centers. We establish that stringent flow stickiness carries a significant performance penalty in terms of packet-level delay. Moreover, relaxing the stickiness requirement by a minuscule amount is highly effective in clipping the tail of the latency distribution. We further propose a bin-based load balancing scheme that achieves a good balance among scalability, stickiness violation and packet-level delay performance. Extensive simulation experiments corroborate the analytical results and validate the effectiveness of the bin-based load balancing scheme.