Login
Computer Science @ Rochester
Friday, March 25, 2011
2:00 PM
Computer Studies Bldg. Room 703
Terence Kelly
Hewlett-Packard Laboratories
Senior Researcher
Abstract:

The massive hardware scale, labyrinthine software complexity, and tangled external interactions of tomorrow's enterprise data centers conspire to undermine reliability: Scale increases the frequency of failures---from hardware glitches and power interruptions to process crashes and kernel panics---while interconnectedness exacerbates their consequences by turning local mishaps into global disasters. Supporting enterprise-class reliability over failures is challenging: Recoverable consistency and outside-world interactivity together require frequent synchronous durable writes, which can degrade failure-free performance. Furthermore, availability demands that recovery be both fast and correct, which argues against manually written application-specific recovery code. Finally, cost- effectiveness and agility require that the average programmer endow enterprise applications with reliability, performance, and availability at reasonable cost in time and money.

This talk will describe current HP Labs research that addresses the challenges of next-generation enterprise computing. Specifically, we will consider a rollback-recovery protocol that is well suited to our requirements, and we will describe how emerging technologies including non-volatile memory make this protocol more attractive. Finally, we will outline ongoing joint work with academic researchers to develop enhanced infrastructure for our reliability techniques.

Bio:

Terence Kelly has been a researcher at Hewlett-Packard Laboratories in Palo Alto, California since earning his Ph.D. at the University of Michigan in 2002. Kelly's past work at HPL has involved performance debugging and performance modeling for enterprise computing, computational resource allocation, cluster scheduling, procurement auction decision support, and control-theoretic concurrency bug fixing. Recently Kelly has focused on fault tolerance techniques for distributed computing.\

Kelly is a named inventor on seven issued US patents, with several others pending, and has served on three dissertation committees and numerous program committees. Details available at http://ai.eecs.umich.edu/~tpkelly/papers/