The three-year grant project, called Big Data and Optical Lightpaths-Driven Networked Systems Research Infrastructure (BOLD), is led by Eugene Ng, who said he wants to eliminate a bottleneck experienced even by researchers using advanced supercomputing clusters like the one at Rice. Ng, an associate professor of computer science and electrical and computer engineering, worked with his team to establish preliminary findings that point toward the possibility of big strides in networking architecture.
“I’m very optimistic,” Ng said. “The early simulations we have done in past work have shown basically that these kinds of networks can get nearly optimal performance without the complexity and cost of using conventional network technologies.”
Optical and electrical networks each have their strong points, and Ng’s research will attempt to marry the two, eliminating a common bottleneck in computing. It’s always possible to throw more money at computer networks to increase their capacity, Ng said, but increasing capacity becomes less efficient and more difficult as demand increases, so they’re looking to create a more scalable, efficient solution.
Electric packet switches have the benefit of low latency, while optical networks have the benefit of carrying very high bandwidth streams. But the problem comes when it’s time to convert optical data into electrical signals that computers can use. Such exchanges are costly, both in terms of time and physical energy, Ng said.
“The purpose [of our research] is to meet the data demands of research where large amounts of data must be moved from measurement instruments to compute centers where they’re stored in file systems," Ng said. "While you’re processing [the data], they need to be repeatedly extracted from the file system into the compute node, and then newly generated data from the processing could also be exchanged with the compute node and be stored back to the file system.” No matter how fast a computing cluster is, processing speed is typically limited by the bottleneck created by the network’s ability to move the data.
Ng said his research will focus on evolutions in hardware as well as the software and middleware that will make using the new network architecture user-friendly. “The user essentially specifies simple functions that they want to apply to the processing of the data. And the middleware deals with how to move the data, how to store the data, how to schedule that computation to run, and utilize the compute resources in a cluster as efficiently as possible,” Ng said. Once this technology becomes available, Ng said he’d like to see researchers focusing more on their science and algorithms and worrying less about their computers and waiting for jobs to finish.
“The infrastructure is going to accelerate their work and make it more efficient,” he said. “Oftentimes the CPU may sit idle doing no computation because they’re waiting for the data to arrive, so all those CPU cycles are wasted if the network can’t keep up with it.”
The project’s first milestone will be a small prototype system built in the Rice campus lab where the team can begin experimenting, Ng said. The next step will be to integrate the system with Rice’s existing supercomputer cluster so that anyone using that cluster for research can begin reaping the rewards of increased speed and efficiency.
Though the grant supplied three years of funding for the project, Ng said he expects what they produce to stay in use much longer than that.
