Managing Resources across Geo-Distributed Datacenters for Big Data Analytics and Mobile Cloud Applications

点击翻译按钮获取中文摘要

Both big data analytics and mobile cloud applications are increasingly geographically distributed, relying upon services from multiple datacenters around the world. Yet, wide-area inter-datacenter links are costly to build and to upgrade, and inter-datacenter bandwidth is bound to become a scarce resource in the near future. Traditional data analytics need to first move all raw data to one datacenter before they can be processed, which adversely affect their performance and are bandwidth intensive. Real-time streaming analytics and interactive mobile cloud applications such as video conferencing and messaging, on the other hand, require low latencies. As such, there is a pressing need to manage and allocate inter-datacenter resources at a fine granularity to meet the diverse performance targets from both data analytics and mobile cloud applications. In this program, we propose to design and implement a new software-defined inter-datacenter infrastructure, in order to optimize the utilization of inter-datacenter bandwidth. Operating in the application layer, our infrastructure provides a substrate and a public application programming interface (API) to support existing data analytic frameworks and interactive mobile cloud applications. It schedules data analytic tasks, routes data streams, and manages bandwidth allocations to application flows at a fine granularity, with the objective of utilizing scarce inter-datacenter bandwidth with optimal efficiency. Following the principles of software-defined networking, our infrastructure fully decouples the control and data planes, and makes task scheduling, flow routing, and traffic engineering decisions based on live analytics of network conditions. We will propose new algorithms to make these decisions in the controller, catering for the specific needs of each category of applications. For batch and streaming analytics applications, our new algorithms are designed to optimize task scheduling and assignment in multiple datacenters, so that both the volume and latency of inter-datacenter traffic can be optimized without prior knowledge of the communication patterns of data analytics jobs. For interactive mobile cloud applications that are bandwidth-intensive and latency-sensitive, our algorithms are designed to optimize resource allocations according to the delay sensitivity of competing application flows. To be more expressive, our infrastructure uses executable scripts to implement control plane decisions in the data plane. We plan to implement our new infrastructure in a real-world inter-datacenter testbed involving multiple datacenters worldwide. This program will lead to fundamentally new ways of using a large number of datacenters around the world to support big data analytics and mobile applications, both of which will soon to be critical and indispensable to businesses and consumers alike in Canada.