|Jeffrey J. Evans|
||Home Teaching Research Publications|
|I have a broad range
of research interests, but generally they are
all tied to the area of distributed computing system dynamics and adaptation. Currently
my primary work is to understand the mechanisms that manifext into
performance degradation in distributed computing systems. My
recent research has included the impact of network performance
degradation on parallel application run time. I have developed a
framework that characterizes this sensitivity on computational clusters
made up of workstations (NOWs) without the need to instrument the
application itself. The framework can also be used by cluster system
administrators to characterize and evaluate their networks. If you are
interested in learning more about the framwork you can read below, then
explore the Adaptive Computing Systems Lab page. If you are interested in working in my lab please email me and we can arrange to discuss your interests.
High Performance Computing (Clusters and Grids)
The framework I have developed for studying communication variability includes a tool called PACE that performs Parallel Application Communication Emulation. PACE uses MPI to emulate real parallel application communication behavior. More than one application can be emulated simultaneously and each emulated application can use different message topologies and exchange semantics. An add-on to PACE has been developed called PARSE which performs Parallel Application Run time Sensitivity Evaluation using PACE. This extension allows applications to be evaluated without requiring the application to be instrumented to measure itself. This saves time and eliminates behaviors that would be caused by additional measurement code in the application. Applications we have begun to characterize include several of the NAS parallel benchmarks, the Parallel Spectral Transform Shallow Water Model used in climate simulations, and the GROMACS molecular dynamics application. There is considerable work to be done in this area that will investigate network adaptation strategies and triggers, job scheduling, and application migration.
To begin this work we have constructed an 80-node Linux NOW cluster in the Adaptive Computing Systems Lab. Our near term work includes growing the machine to 128 or more nodes and to install at least one high performance interconnect network (preferrably Myrinet) to compliment the Fast Ethernet interconnect already in place.
Sensor networks, especially those of the wireless variety has emerged as a popular research area that poses many challenges. We are collaborating with scientists from Agricultural and Biological Engineering (ABE) to develop efficient and reliable sensor networks specific to interesting combinations of existing and new sensor technologies for which common off-the-shelf (COTS) sensor network components are unsuitable. This requires knowledge and skills in a wide variety of disciplines including highly specialized electronic circuit development, electronics manufacturing, reliability, power management,data movement (routing), and systemic control.