Computer Sciences Dept.

Mary K. Vernon

Professor of Computer Science
                  and Industrial Engineering
Picture of Mary K. Vernon
Research Interests: Techniques and applications of computer systems performance analysis, high performance network protocols, networked system security, parallel/distributed applications and architectures, Grid and cluster job scheduling.

Research Summary

The theme of my research to date has been the development and application of analytic modeling techniques that enable the design of software, hardware, and communication networks with near-optimal performance. The contributions have included (1) customized equations that directly specify an optimal system design, (2) analytic bounds that quantify the opportunity for improvement as well as the target system performance, (3) customized analytic models that reflect the mechanics of the system and that accurately estimate measured system performance for various design options, and (4) application of the models to derive or invent new near-optimal systems that significantly outperform previous systems.

Analytic models that are accurate, yet as abstract as possible, readily expose system features that optimize or inhibit system performance, including any performance bottlenecks that can be eliminated. Thus, the research in analytic design of commercially important systems has led to significant insights into system design bottlenecks, leading to important new system designs in addition to the new system design techniques.

The modeling techniques I've developed previously together with graduate students, an undergraduate student, and faculty colleagues include:

  • the Generalized Timed Petri Net (GTPN) (with Mark Holliday),
  • Customized Approximate Mean Value Analysis (CMVA) (with Derek Eager, Ed Lazowska, Haonan Tan, and John Zahorjan),
  • deterministic task graph analysis (with Vikram Adve),
  • interpolation approximations for evaluating parallel processor scheduling policies (with Rajesh Mansharamani),
  • LoPC (with Matthew Frank and Anant Agarwal), and
  • models for determining the proxy server content that minimizes delivery cost for multicast streaming media (with Jussara Almeida, Derek Eager, and Michael Ferris).
Our recent innovations in CMVA include analysis of high variability in service times and the resulting highly bursty arrivals to downstream servers, as well as models that estimate client loss probabilities as low as one in ten thousand.

We have validated these techniques and used them to obtain important new design insights for bus arbitration, cache coherence protocols, mesh interconnection networks with wormhole routing, the Sequent Symmetry bus, parallel shared memory system architectures (with complex modern processors), the Cray UNICOS operating system semaphores, complex parallel/distributed applications, parallel processor scheduling policies, global memory management in NOWs, scalable on-demand continuous media delivery protocols, and content distribution networks for popular media objects.

My current research includes the design of key parallel/distributed applications, high performance network transport protocols, efficient network bandwidth estimation techniques, methods for estimating the arrival rate of bursty arrival processes, storage systems, and optimized server and client software co-designed with optimized CMP hardware support. Further information about these current research activities is available in the web pages for my current research projects.

Computer Sciences | UW Home