In her sixth year as a PhD student in the computer science department at Rice University, Yiting Xia was struck by an untraditional idea for improving the speed of data center transmissions. “Back in March, I felt so sleepy as I was laying down one night,” she said, “but this idea kept hitting me until I finally got up, got a piece of paper and drew the layout of the network and thought ‘I’m done.’ It took like an hour.”
But her rest remained disrupted. Xia said, “I tried to go to bed a second time and another idea hit me. Again, I was so sleepy, but I couldn’t fall asleep. I went back to the paper and put down some more notes.”
By November, Xia was presenting her new ideas in the 15th ACM Workshop on Hot Topics in Networks (HotNets 2016) in Atlanta, Georgia. Although her progress may seem rapid, she attributes her success to years of focused study and research. “I don’t believe in overnight success,” she said. “You accumulate your knowledge and then that inspires your bright idea one day.”
Previously, Xia’s research focused on leveraging optical devices to improve network performance. She devised improvements for the physical layer using optical devices like power splitters to boost data transmissions in big data applications. Splitters are common components in traditional infrastructure plans, duplicating input signals N-way to multiple outputs. Xia said, “It is exciting to find its new use in data centers: relaying data in multicast, the one-to-many communication pattern widely observed in machine learning applications.”
“We were able to reduce overall data transmission time by several orders of magnitude,” she said as she described the team’s novel approach. “We successfully transformed improvements in the hardware into accelerations of real applications, which is usually considered hard because they have a huge gap as the bottom-most and the top-most layers in the network.”
But Xia’s sleep-interrupting idea challenged traditional data center construction. She said, “People never think of changing the original network topology, but different arrangements of network devices can have different futures and different benefits. How could we combine the benefits of these different topologies for both good performance and easy construction?”
Most people choose optical devices like switches to enable faster transmissions to send traffic. On a Houston freeway, this would be similar to designating certain lanes for high occupancy vehicles when commuter traffic is heaviest. Xia proposed not adapting existing freeway lanes, but temporarily borrowing less utilized lanes from other parts of the grid. “By liberating very smart devices to change the cabling, we’re leveraging existing infrastructure for better performance,” she said.
Recent developments have already helped minimize the labor costs associated with rearranging data center configurations. Data center managers previously addressed reconfigurations by hiring a lot of people to go around and unplug, move, and reconnect servers and switches. But companies like HP have introduced pre-fabricated units like the Performance Optimized Datacenter (POD). These pre-constructed units arrive with their internal server clusters ready to begin interfacing with the rest of the data center after only a few external connections are made.
CS professor Eugene Ng, Xia’s adviser, encouraged her to apply the temporary road re-assignment concept to the new pre-fabricated data center units. “We can automatically hit a button to move a ‘road’ and then when the congestion is solved, hit a button and move it back or even turn off a road or two, to save energy. If you have a fixed number of roads, the way you arrange them can change your capacity without adding roads. We can maximize the existing infrastructure to save on budget.”
She’s had internships with both Google and HP Labs, and as she draws closer to her dissertation, she’s considering what a career in industry might be like. “I still love research,” she said, “but I think I am ready to move on, to do something else. My research is very practical. We build networks over automation, boost transmissions to have better experience for your applications. So my new passion is to work in the industry and apply my skills to a real product.”
Xia points out that the apps most people use on their phones or to surf the Internet are not built on super advanced technologies and she feels there are still gaps between research and users’ end experience. “How well you build your application or product to leverage the most advantage of technology… that would be a very cool thing to do, to help with that transition,” she said. “It’s out of my comfort zone, but I like to challenge myself. It is safe to stay in a research lab, but I want to do the right thing for me and chase my own passion.”
Yiting Xia completed her Ph.D. in CS in 2018.