QnAs with Nancy A. Lynch

In computer science, distributed systems involve many processors cooperating to complete a certain task—a field with significance for the current age of wireless communications and cloud computing. Nancy Lynch, the NEC Professor of Software Science and Engineering at the Massachusetts Institute of Technology, literally wrote the book on distributed computing. Her textbook Distributed Algorithms is considered a definitive reference work for the field. Her contributions to the field include impossibility results—such as her well-known “FLP” result—that establish theoretical limits on what distributed systems can accomplish. In recognition of her work, Lynch was elected to the National Academy of Engineering in 2001 and the National Academy of Sciences in 2016, among numerous other awards and honors. PNAS spoke with Lynch about her work.

Nancy A. Lynch. Image courtesy of Tony Rinaldo (photographer).

PNAS: Why did you decide to study distributed computing?

Lynch: I got interested in this area back in 1977 or 1978, soon after I joined the faculty at Georgia Tech. I had been working in theoretical computer science, but I thought that it would be more interesting to look for an area that would connect better with practical computer science research. Distributed computing was a new “hot area” at the time, and I figured that this was an area where theoretical techniques could have something to contribute.

PNAS: What is the “FLP” result?

Lynch: “FLP” is just the authors' initials—Fischer, Lynch, and Paterson—for my best-known paper that we wrote back in 1985, which is called “Impossibility of distributed consensus with one faulty process” (1). If you have many processors that want to make a decision, and some have an initial preference of voting “yes,” and some have an initial preference of voting “no,” you want to have them exchange their information and decide one way or the other. This comes up, for example, in the database area where processors are handling or managing data transactions. At the end of the transaction, which has to be executed in many places in the network, the processors all have to agree: Should they accept the results of the transaction or should they abort the whole thing? At the same time, you have to deal with possibilities of some of the processors failing. The FLP result shows that, if you have to contend with the possibility of even one processor failing, then it isn't possible for the processors to reliably reach agreement on a decision. It’s kind of a surprising result. From the practical side, it meant that people designing transaction-processing algorithms had to be careful when they described their work, to say how it circumvented the limitation that’s expressed by this impossibility result.

PNAS: Your Inaugural Article (2) explores how ants estimate local population density. Why did you decide to study this problem?

Lynch: As I worked on distributed algorithms over the years, it became clear that many biological systems are really distributed algorithms. A good example is insect “house hunting.” Ants have to decide on the location of a new nest, and they do this by exploring randomly, finding potential nests, deciding if they like them, and trying to persuade others. But when they reach a certain point, and you have enough ants in one nest, then they say, “okay, this is our new nest,” and then go off and force other ants to join them. But how do they judge when there are enough ants in the nest? They don’t actually count them, they just see if they’re dense enough. And the way they tell if they’re dense enough is they walk around randomly and see how many ants they bump into in a short period. We thought we’d analyze how well that works. It turns out it works pretty well. Just using this random walking technique, they get very close to the actual number.

PNAS: What are some applications for the results of your Inaugural Article?

Lynch: My student Cameron and postdoc Hsin-Hao took this further and used these techniques to estimate sizes of social networks. Our Inaugural Article assumes that you know the area and you could estimate how many ants there are. But conversely, if you know how many agents there are, you could figure out the area, which is basically the size of the social network. In general, the kinds of algorithms that insects perform can also be executed by swarms of robots. For instance, Radhika Nagpal’s group at Harvard has developed swarms of robots that perform algorithms that are very insect-like—the same types of algorithms that an ant colony might perform.

PNAS: What are the next steps in this work?

Lynch: We’re interested in distributed algorithms for the brain. Suppose you have a number of neurons firing, and you want to select only the one that is firing at the fastest rate to continue firing. We tried to figure out what’s the smallest network that can solve this problem and that can cause the decision to be made as fast as possible. Also, I’m very interested in how you can put these algorithms together so that, if you can solve one problem, then you could use that as a sort of subroutine to solve another one. We keep looking for parallels between biological systems and computer network algorithms. The hope is that we help researchers to understand the behavior of biological systems, and we also learn to use some of the techniques of biological systems—the principles that biological systems use—to help design better computer algorithms.