Decreasing Queue Length in Data Centers
Most every time you visit a website, a host of servers somewhere have to run the right operations to gather the data you need, and then send it to you. As some websites use large data centers, with the servers working for you spread out, the latency between the servers can impact performance. Currently decentralized communication protocols are used to manage communication in a data center, but researchers at MIT have recently designed a new, centralized system that can offer better performance.
Decentralized protocols enable each node in a network to send and receive information without instruction. Provided the routers transmitting the information do not get overwhelmed, this approach can work well, but in some data centers they are being overwhelmed, causing large queues to form, leading to congestion. One would not expect a solution to come from sending requests to a central arbiter server, which takes 40 microseconds, but the MIT, Fastpass system not only reduces the congestion, but does so by a very significant amount. The Fastpass system takes advantage of parallel programming to divide the work of scheduling communication across multiple cores. The first core looks at the pending requests, schedules one for a slot, and then passes to the next core all of the requests involving either the source or destination node of the schedule request.
The researchers found that by using this approach, the Fastpass system is able to handle a network transmitting at 2.2 terabits per second, with just eight cores. In experiments to be presented in August, the researchers will show that Fastpass cut the average queue length in a Facebook data center by 99.6%, and the average latency from 3.56 microseconds to 0.23 microseconds.