Dynamic and Load-Aware Flowlet for Load-Balancing in Data Center Networks

Load balancing is critical for fully utilizing network bandwidth in data center networks. Although simple and relying on only local information, the widely-used methods such as ECMP and Flowlet have their limitations and drawbacks. We design DLAF, a novel hardware-based load-balancing architecture and algorithm, to address their issues and achieve better load-balancing performance. DLAF applies the multi-hashing scheme to dynamically generate flowlets without needing to configure a static timeout threshold, and sends each flowlet to the least-loaded equal-cost path based on local counting. We simulate DLAF and the other methods using a real data center packet trace to reveal their performance. We further implement the algorithm using P4 and evaluate it using iperf on a fat-tree topology in Mininet. The comparisons with ECMP and Flowlet show that, with similar memory consumption to that of Flowlet, the new algorithm achieves not only perfect load balancing but also the best flow throughput performance. Therefore, it is ideal to be deployed on programmable switches as the default multi-path load-balancing solution in data center networks.