Proactive Congestion Avoidance under Network Dynamics

Опубликовано 21 июня 2016, 18:40

Latency sensitive applications, such as search, online games and video conferences, are rapidly growing and increasingly rely on cloud network infrastructures to deliver their traffic smoothly. However, the traffic distributions on cloud networks are constantly evolving due to dynamics like updates in network devices (e.g. firmware upgrades), changes in applications (e.g. VM migrations) and various kinds of faults (e.g. configuration/link/device failures), resulting in a significant risk of congestion. Congestion hurts customer experience and ultimately leads to a loss of revenues for cloud service providers. In this talk, I will present two concepts in traffic management for proactively protecting the dynamic networks from congestion: Smooth traffic Distribution Transition (SDT) and Forward Fault Correction (FFC). SDT supports updating the routing rules in a network to achieve a traffic distribution that satisfies requirements from operators without causing congestion. For example, before rebooting a switch, an operator will drain the switch first. SDT provides a common functionality that is needed during diverse types of network updates. The key challenge to realize SDT is from the inherent difficulty in synchronizing the changes to many devices, which may lead to unforeseen transient link load spikes or even congestion. I present one primitive, zUpdate, which performs SDT via multi-step and progressive network re-configuration planning. While SDT addresses predictable updates, cloud networks also face unpredicted challenges such as link failures and high switch-configuration delays that can cause heavy congestion and packet loss. To address this challenge, I propose FFC which spreads network traffic such that freedom from congestion is guaranteed under arbitrary combinations of up to k faults. FFC encodes the constraints that arise from the large number of possible fault cases and solves them efficiently using a "sorting network".