Fault-tolerant Distributed Systems

Distributed systems have two key advantages: they increase the quality of our interaction with computers and the quantity of resources available to us. The quality increases as components of a distributed system hide failures from users and render a service more reliable (safer from loss) andmore available (less downtime). Increases in quantity, such as computational power and storage capacity, enable users to overcome fundamental limits of scalability of a single node and solve problems larger in size and solve them faster. Although these two advantages are obvious in principle, reaping the benefits is often hard in practice. I have done research in both of these directions: most recently on fault-tolerant distributed systems and in the past on high-performance parallel systems. All projects that I have worked on share a unifying thread: they enable novel distributed solutions on existing platforms—in a sense, working around the constraints of the deployed infrastructure. For example, instead of inventing a new transport protocol, I have built a system that hides server failures by relying on standard TCP semantics. The strength of this type of research is that it makes no assumptions about future technological trends, but works with what already exists. I believe that accommodating entrenched standards and technologies is crucial for any system’s viability and vitality.