Postcard: Minimizing Costs on Inter-Datacenter Traffic with Store-and-Forward

It is typical for cloud providers to operate a number of geographically distributed data centers, where inter-data center traffic constitutes a large portion of a cloud provider's traffic demand over the Internet. Though such inter-data center traffic incurs substantial operational costs that are charged by ISPs, it varies significantly across different overlay links, raising the opportunity to optimize both routing and scheduling of inter-data center traffic to minimize costs. As data is being transmitted multiple sources to their respective destinations, an intermediate data center is also able to store data and forward them at a later time, so that peak traffic demand can be reduced. The cost minimization problem with such store-and-forward is challenging to solve in the general case, due to a large number of variables that determine how data can be sent from a source to a destination, and when should they be "paused" (stored) at an intermediate node, subject to a required maximum transfer time. In this paper, we present Postcard, an online optimization problem carefully formulated to minimize operational costs on inter-data center traffic with store-and-forward at intermediate nodes. To solve the optimization problem with an acceptable number of variables in Postcard, we have simplified the general problem by restricting data transmission to a time-slotted model, such that the problem can be modelled on a time-expanded graph. With extensive simulations, we compare results from solving Postcard to those from solving a flow-based problem without store-and-forward, and present the advantages and drawbacks of store-and-forward when it comes to minimizing costs on inter-data center traffic.