Electrical & Computer Engineering     |     Carnegie Mellon

Thursday, September 14, 12:00-1:00 p.m. HH-1112


Phillip Stanley-Marbell
Technische Universiteit Eindhoven

Programming Unreliable Networks of Computation

It is possible to build systems comprising large numbers of failure-prone, energy- and memory-constrained computing elements, and to utilize them for computational tasks. We consider the mechanisms by which upsets/faults occur in both computation and communication, and how, when not masked, such upsets may lead to errors and failures. The effects of logic upsets on the numeric magnitude of value deviations in programs are studied. These analyses are used to develop methods for encoding values in programs, such that the numeric magnitude of error is kept within bounds specified in the type annotation of variables in programs. For such encoding methods, we study the theoretical limit of encoding efficiency that can be attained. These ideas are embodied in a programming model and its realization in a small programming language, targeted at networks of failure-prone, resource-constrained computational devices. The language enables programs to internally encode the amount of deviation of their values that are acceptable, the specification of latency and loss constraints on communications, and the change of control-flow when these constraints are violated.

This talk will overview the language model and discuss the analytic and empirical constructions necessary for bounding per-variable value deviations in programs. To provide context on the relevance to the construction of real systems, the design and fabrication of a prototype hardware platform intended for implementing our target systems is briefly outlined.


Phillip is a postdoctoral researcher at TU/e (Eindhoven, The Netherlands). He received the undergraduate degree in materials science from Rutgers University (NJ, USA), having sufficiently damaged his health by engaging in undergraduate research on lead-zirconate-titanate (PZT) composites for piezoelectric applications. He joined the graduate program in computer engineering at Rutgers, escaping after 2 years, with a "TA of the year" award and an M.Sc. in computer engineering, in 2001. He then joined CMU (Pittsburgh, USA). After a fair amount of wandering in the wilderness, he completed his Ph.D. in 2006, supported in part by a Lamme/Westinghouse graduate fellowship award (2004-2005).

Prior to, and during graduate school, he held various intern and full-time positions at AT&T Microelectronics, Lucent Microelectronics, Philips Consumer Communications, Lucent's Data Networking Group, and NEC research labs. He is the author of a programming language textbook published by John Wiley & Sons (Chichester, UK) in 2003. Phillip is a member of the ACM, IEEE, Sigma Xi, and USENIX, and from 2003-2004 served as the copy editor for the ACM Mobile Computing and Communications Review (MC2R) journal. His current research focus is on bits that misbehave, bits that cause the movement of atoms, and atoms that control the movement of bits, sometimes collectively referred to as failure-prone battery-powered computing systems.