Microsoft Research335 тыс
Следующее
Опубликовано 21 июня 2016, 21:25
Abstract: Cloud network infrastructures, such as datacenter and inter-datacenter networks, are increasingly expected to provide vital support for cloud applications such as online services, big data processing, large graph/matrix computation, and distributed storage. The large variety of cloud applications have demanded a diverse range of network service requirements such as minimizing flow completion times, optimizing network utilization, meeting latency targets, and satisfying fairness constraints across tenants. However, network architectures used today are far from optimal for meeting these requirements. This mismatch hurts customer experience and ultimately leads to a loss of revenue for cloud service providers. In this talk, we argue for a programmable network abstraction — A software defined transport (SDT) architecture for optimizing cloud network infrastructure. By providing a centralized programmable platform to define network transport mechanisms in software, network operators can easily schedule network resources to optimize for service requirements in real time. However, explicit network resource scheduling at the central controller would raise a serious latency and scalability concern. To solve this issue, we proposed resource allocation algorithms that can scale SDT to many of today's cloud network infrastructures. Further, centralized network resource allocation can cause severe, transient congestion because network devices may apply changes at different times. We develop a novel technique that leverages a small amount of scratch capacity on links to apply changes in a provably congestion-free manner.
Свежие видео
Случайные видео