(Jun 26) Towards Latency Guarantees in Datacenters

A large portion of computation is now carried out in datacenter. A single datacenter hosts several 100s or 1000s of applications which share common compute and network resources in a datacenter. Isolating each application’s performance, i.e., ensuring that its performance is predictable regardless of the behavior of other applications is essential to developing and deploying data center applications, since otherwise developers need to account for co-located applications, which increases development and deployment complexity. Despite its importance current approaches to performance isolation are incomplete, and focus mostly on isolating computational resources. In this talk I present two schemes for isolation network performance. The first, Silo, takes a resource allocation based approach and implements mechanisms for guaranteeing an application’s network latency and throughput. The second, Express Pass, takes a congestion control based approach and fairly partitions network resources across applications.  Both approaches require no hardware (ASIC) changes, and can be deployed in today’s datacenter. This shows that full application performance isolation is achievable today.