Emerging datacenter applications, such as distributed Machine Learning training, rely on predictable networking to achieve high performance. Therefore, cloud providers should efficiently multiplex datacenter network bandwidth, in addition to compute, memory and other datacenter resources, across multiple applications. We develop heuristic and constraint-solver based algorithms to handle datacenter network bandwidth allocation.

Resource allocation has two main objectives: high datacenter utilization and low-latency decision making. We design algorithms that provide high datacenter utilization with practical resource allocation latency. We prototype our solution with OpenStack to evaluate the practicality of our approach.