Hakim Weatherspoon finds ways to make large, distributed computer systems work better. “I utilize machines that communicate through the network to provide some service,” he says. “We start out with the theory then we have to create some artifact.” 

As part of his dissertation, Weatherspoon developed a software layer that allows many data centers to provide backup storage for each other. “All the nodes in this large federated system cooperate to maintain a significantly higher degree of data reliability and the cost to an individual institution is incremental,” he says. “They have to put in an extra server disk.”

Unlike the distributed systems used in cloud computing, the machines are in different administrative domains, and so are mutually distrustful of each other, even while they share resources. “You need some mechanisms to ensure the bits are the same when you read them some time later, that the bits exist when you attempt to read them, and that the bits will be accessible,” explains Weatherspoon.

Extra storage is cheap, but keeping those disks spinning takes energy, so Weatherspoon is creating a way for data centers to separate more frequently accessed data from little used data. “Now you can talk about saving energy on a global scale by moving data around accordingly and trading off which disk in which data center is spun up and which one is spun down to optimize performance as well as energy consumption,” he says.

Another project solves the dilemma of whether to back up data in real time, and slow down the system, or wait until the system isn’t so busy, and risk losing some data if disaster strikes. Performance suffers with the former because the sender waits for acknowledgement that the last batch of data was received before sending the next. Over fiber-optic networks, the response time is limited by the speed of light. Those fiber-optic networks, however, are ultra reliable. By adding a little redundancy to enable any lost data to be reconstructed, Weatherspoon has found they can be just as reliable as the server disks the data is eventually stored on.

“Why not consider it stably stored while in transit and move on to the next operation?” he asks. “We call it the smoke and mirrors file system. It’s a different way of approaching the problem.”

Weatherspoon finds that applications that involve data help focus his research, but he’s also published papers in networking and operating systems. “In order to enable this smoke and mirrors process, adding redundancy into the network actually required a new networking protocol,” he says. “It also required some operating systems research in order to process packets at line speed.”

When he got his Ph.D., Weatherspoon wasn’t sure he wanted to join academia; most computer science doctorates go into industry. So he chose a less common career path—a two-year postdoc right here at Cornell with Ken Birman’s group. “I wanted to see if I wanted to teach, advise students, and write research proposals,” he says. “I found out that I love teaching and advising students, and doing research. I like the cooperation of working in groups and leading a team. I also like the competitive nature of writing a proposal; you are challenged to do the best that you can.”