Computer Science Seminar Series
Remote Visualization and Computational Steering for Large-Scale Scientific Applications
April 2, 3:00pm
Weir Hall, Room 235
Qishi Wu
Assistant Professor
Department of Computer Science
University of Memphis
Collaborative Research Associate
Computer Science and Mathematics Division
Oak Ridge National Laboratory
Abstract:
The progress in supercomputing technology has fundamentally changed the way of conducting basic and applied sciences, ranging from astrophysics, computational biology, climate research, computational materials, fusion simulation, neutron sciences, to nanoscience. The large volumes of scientific data generated by the simulation based on large-scale computation have presented unprecedented challenges on the traditional computing sciences. We propose Distributed Remote Intelligent Visualization Environment (DRIVE) that incorporates various research practices and emerging computing technologies to support large-scale scientific applications. This talk mainly focuses on two aspects of DRIVE: transport control for goodput maximization and system optimization for dynamic visualization pipeline configuration.
To address the bandwidth needs for large data transfer, we design a new class of UDP-based transport protocols that utilize a rate control scheme founded on the stochastic approximation method to achieve high throughputs at the application level. These protocols operate around a local maximum of the throughput regression curve by dynamically adjusting the source rate in response to acknowledgements and losses based on the statistical behavior of the network connection. We analytically show that this protocol generates a TCP-friendly flow, and also stochastically converges to the maximum throughput under a monotone loss rate condition. Our implementation achieved very robust performance over diverse Internet connections with different characteristics: it tracked the peak throughput in presence of time-varying cross traffic and consistently achieved 2-5 times the throughput of default TCP without significantly affecting the concurrent regular traffic.
To address the scientific visualization and computational steering needs in a distributed environment, we develop an approach that dynamically decomposes and maps a visualization pipeline onto wide-area network nodes to achieve fast interactions between users and applications. This scheme enables the selection and aggregation of computing nodes with disparate capabilities as well as communication connections with different properties. We estimate the transport and processing times of various subtasks, and present a polynomial-time algorithm to compute a decomposition and mapping solution that achieves the minimum end-to-end delay. We present experimental results based on a deployment at several geographically distributed nodes to illustrate its efficiency in terms of resource utilization.