Lowering Inter-datacenter Bandwidth Costs via Bulk Data Scheduling

Abstract

Cloud service providers (CSP) of today operate multiple data centers, over which they provide resilient infrastructure, data storage and compute services. The links between data centers have very high capacity, and are typically purchased by the CSPs using established billing practices, such as 95-thpercentile billing or average-usage billing. These links are used to serve both client traffic as well as CSP-specific bulk data traffic, such as backup jobs, etc. Past studies have shown a diurnal pattern of traffic over such links. However, CSPs pay for the peak bandwidth, which implies that they are under-utilizing the capacity for which they have paid for. We propose a scheduling framework that considers various classes of jobs that are encountered over such links, and propose GRESE, an algorithm that attempts to minimize overall bandwidth costs to the CSP, by leveraging the flexible nature of the deadlines of these bulk data jobs. We demonstrate the problem is not a simple extension of any well-known scheduling problems, and show how the GRESE algorithm is effective in curtailing CSP bandwidth costs.