Abstract

Meeting tail latency Service Level Objectives (SLOs) in shared datacenter networks is known to be an important and challenging problem. The main challenge is in determining limits on the multi-tenancy such that SLOs are met. This requires calculating latency guarantees, which is a difficult problem, especially when tenants exhibit bursty behavior as is common in production environments. Nevertheless, recent papers in the past two years have shown techniques for calculating latency based on a branch of mathematical modeling called Deterministic Network Calculus (DNC). The DNC theory is designed for adversarial worst-case conditions, which is useful in some scenarios, but is often overly conservative. Typical tenants do not require strict worst-case guarantees, but are only looking for SLOs at lower percentiles (e.g., 99th, 99.9th). By considering SLOs at lower percentiles, it is possible to pack together many more tenants while still meeting tail latency SLOs. In this talk, I’ll present a brand new technique for calculating tail latency based on a probabilistic theory called Stochastic Network Calculus (SNC). SNC is a new theory that is actively being developed by the theory community to overcome the limitations of DNC , and we are the first to bring this theory to practice in a real computer system. In experiments on our cluster, we demonstrate that our system can support twice as many tenants as the state-of-the-art while meeting tail latency SLOs.