Bridging the gap between applications and networks in data centers

Modern data centers host tens (if not hundreds) of thousands of servers and are used by companies such as Amazon, Google, and Microsoft to provide online services to millions of individuals distributed across the Internet. They use commodity hardware and their network infrastructure adopts principles evolved from enterprise and Internet networking. Applications use UDP datagrams or TCP sockets as the primary interface to other applications running inside the data center. This effectively isolates the network from the end-systems, which then have little control over how the network handles packets. Likewise, the network has limited visibility on the application logic. An application injects a packet with a destination address and the network just delivers the packet. Network and applications effectively treat each other as black-boxes. This strict separation between applications and networks (also referred to as dumb network) is a direct outcome of the so-called end-to-end argument [49] and has arguably been one of the main reasons why the Internet has been capable of evolving from a small research project to planetary scale, supporting a multitude of different hardware and network technologies as well as a slew of very diverse applications, and using networks owned by competing ISPs. Despite being so instrumental in the success of the Internet, this black-box design is also one of the root causes of inefficiencies in large-scale data centers. Given the little control and visibility over network resources, applications need to use low-level hacks, e.g., to extract network properties (e.g., using traceroute and IP addresses to infer the network topology) and to prioritize traffic (e.g., increasing the number of TCP flows used by an application to increase its bandwidth share). Further, a simple functionality like multicast or anycast routing is not available and developers must resort to application-level overlays. This, however, leads to inefficiencies as typically multiple logical links are mapped to the same physical link, significantly reducing application throughput. Even with perfect knowledge of the underlying topology, there is still the constraint that servers