Cloud control: voluntary admission control for intranet traffic management

The Internet backbone of many corporations carries two kinds of traffic: urgent and delayable. Shifting some traffic from peak periods to valleys reduces capacity requirements. We consider the case of managing the delayable traffic by an admission control (AC) system. AC gets link utilization feedback every τ seconds. Delayable opt-in sources obtain permission from AC to transmit for up to τ seconds at a rate not exceeding a limit imposed by AC, renewing permission as needed. Urgent traffic bypasses AC. AC must allocate bandwidth to competing delayable traffic sources. We prove that among all throttling transformations of flows that achieve a desired mean aggregate flow, rate limits on flows minimize the variance of their sum. Furthermore, a single rate limit common to all flows achieves the optimum. Thus, for a single link, AC must decide on a single rate limit for all delayable sources in each τ-second cycle. We evaluate different policies that set the rate limit dynamically in an empirical setting using netflow records on a link on the backbone of Yahoo!. Using historical data, we also derive the best possible reduction in capacity of this link using a closed-form solution to an assignment problem. We show that AC can achieve capacity reduction close to the best possible reduction.