What's Next for Datacenter Load Balancers - A Comparative Performance Study of the State-of-the-Art
Lakshmi Krishnaswamy, Katia Obraczka, Abdul Kabbani · 2024
Network-level load balancing algorithms have been proposed to help provide application flows with adequate quality-of-service (QoS) by assigning traffic to different network paths to avoid congestion as well as improve overall network utilization. As datacenters scale up and become more geographically distributed, wide-area network (WAN) datacenter traffic, which typically consists of data-heavy tasks, has become increasingly prevelant. Some of the noteworthy challenges raised by the coexistence and interaction between WAN- and intra-datacenter (DC) traffic are the differences in their QoS requirements, link utilizations and round-trip times. To the best of our knowledge, these challenges have not yet been addressed by existing datacenter load balancers. To highlight this gap, in this paper, we conduct a comparative performance study of state-of-the-art datacenter load balancers considering datacenter topologies and workloads that reflect how datacenters have evolved to keep up with their driving applications. Through extensive simulations, we study state-of-the-art load balancers under different network topologies and workloads, including DC-, WAN-, as well as DC-WAN mixed workloads. Our study shows that current load balancers are not able to adequately distribute load under WAN and DCWAN mixed traffic. Additionally, we also observe that load balancer performance is highly sensitive to its parametric choices, especially when subject to WAN and DC-WAN mixed workloads. Our study also highlights that the interplay between transport-level congestion control and load balancing mechanisms plays a significant role and must be considered in order to meet desired performance goals. We also share our experience engineering our experimental methodology by (1) providing recommendations on how to set load balancer parameters and (2) introducing a datacenter workload modeling framework to represent both synthetic and more realistic datacenter workloads, including mixed DC and WAN traffic.