Measuring congestion-induced performance imbalance in Internet load balancing at scale

Abstract

Internet load balancers typically use congestion-oblivious hashing algorithms to distribute traffic across load-balanced (LB) paths, resulting in imbalanced load distribution. When congestion occurs, LB paths could experience different levels of congestion, called congestion imbalance. Congestion imbalance has been extensively studied in datacenters (DCs). It is, however, still under-explored in the Internet. In this paper, we take a first step towards measuring congestion imbalance among Internet LB paths at scale. We present Congi, our lightweight prober, that leverages support vector machine (SVM) classifiers to efficiently detect congestion imbalance using latency samples. Our experiments show that Congi is capable of detecting congestion imbalance between uncongested and congested LB paths: on average, the uncongested path has 3x greater throughput and 0.4% lower packet loss rate than the congested one. To measure congestion imbalance at scale, we then use Congi to conduct measurement campaigns from worldwide DCs to millions of/24s. We find that most DCs experience significant congestion imbalance to 36%–43% of end hosts. Lastly, we use Congi to direct web page downloads under congestion imbalance from a campus network, and show that the download time can be reduced by 44% on average for Alexa top sites for clients experiencing network congestion under congestion imbalance.