The case for fine-grained traffic engineering in data centers

In recent years, several techniques have been suggested for routing and traffic engineering in data centers. However, not much is known about how these techniques per-formrelative to each other under realistic data center traffic patterns. Our preliminary study reveals that existing techniques can only achieve 80% to 85% of the ideal solution in terms of the number of bytes delivered. We find that these techniques suffer due to their inability to utilize global knowledge of the properties of traffic flows and their inability to make coordinated decision for scheduling flows at fine time-scales. Even recent traffic engineering techniques such as COPE fail in data centers despite their proven ability to adapt to dynamic variations, because they are designed to operate at longer time scales (on the order of hours, at least). In contrast, data centers, due to the bursty nature inherent to their traffic, require adaptation at much finer times scales. To this end, we define a set of requirements that a data center-oriented traffic engineering technique must possess in order to successfully mitigate congestion. In this paper, we present the design for a strawman framework that fulfills these requirements.