Skip to content

METASYSTEMS| Contents | Next

A Ripening Crop of AppLeS

Fran Berman, UC San Diego
Rich Wolski. University of Tennessee

Henri Casanova, Walfredo Cirne, Marcio Faerman, Jaime Frey, Jim Hayes, Graziano Obertelli, Gary Shao, Shava Smallen, Alan Su, Dmitrii Zagorodnov, UC San Diego

Bruce Johnson, Todd Bryan, University of Tennessee

Legion, Globus
Network Weather Service
Data-Intensive Computing
Digital Libraries and Data Repositories
Molecular Science
Biomolecular Structure and Energetics
Refining and Linking Brain Data

M etasystems let users tap into many computers on multiple networks to run demanding applications. But shared resources need careful scheduling to get good application performance. Researchers at UC San Diego and the University of Tennessee have created Application Level Schedulers (AppLeS) to help applications achieve their best possible performance on a metasystem. Each AppLeS scheduler, called an agent, integrates with a single application to forecast the capacity of resources, schedule computation, and coordinate data transfer. AppLeS schedulers have enhanced the performance of more than a dozen applications, both within and outside of NPACI, and demonstrated the value of adaptive scheduling.

"In other words, the AppLeS approach is bearing fruit," said project co-leader Fran Berman. "AppLeS complements Legion and Globus by providing an adaptive scheduling mechanism." AppLeS applications also leverage Rich Wolski's Network Weather Service (NWS), another NPACI Metasystems project, to monitor and predict the varying performance of potentially usable resources.





UC San Diego graduate students Marcio Faerman and Alan Su are implementing AppLeS agents for the Data-Intensive Computing thrust area. In particular, many data-intensive applications make remote file transfers over dynamically changing networks. Faerman has developed an Adaptive Regression Modeling (ARM) prediction method to forecast the performance of remote file transfers based on past behavior and live dynamic information from the NWS. The ARM method automatically adapts its forecasting model using statistical regression.

Faerman has accurately predicted performance for experiments with the SDSC Storage Resource Broker and the JPL Synthetic Aperture Radar Atlas (SARA). Su has been working on an AppLeS agent that can select the best source from which to obtain remote SARA images. Su is also studying scheduling techniques for digital libraries and Web publishing in heterogeneous environments.

Top| Contents | Next


AppLeS group member Graziano Obertelli is working with the University of Virginia's Legion team on the Department of Defense HPC Modernization project as part of the Naval Oceanographic Office Programming Environment and Training program. He and other group members are developing an AppLeS agent for a magnetohydrodynamics code package called PMHD3D that is being targeted to Legion.

Experiments have shown that PMHD3D performance with adaptive scheduling is up to 2.5 times faster than with static scheduling because AppLeS assigns more work previously underloaded processors.

Shava Smallen, researcher Jaime Frey of the National Center for Microscopy and Imaging Research, and Walfredo Cirne, an assistant professor at Brazil's Universidade Federal da Paraiba on leave with Berman's group, are scheduling a parallel tomography code across both batch and interactive workstation clusters. (The tomography application, which reconstructs a 3-D image from 2-D slides, is being developed with Mark Ellisman from the Neuroscience thrust, Carl Kesselman and Mei-Hui Su from the Globus project, and the AppLeS group.) The AppLeS approach can improve "turnaround time"--batch queue waiting time plus execution time, the most realistic measure of performance for users

"Applications typically target either batch or interactive environments," Berman said. "The tomography code demonstrates that some applications can achieve even better performance by targeting both and using all of the resources available at execution time."

Top| Contents | Next


AppLeS agents have been developed for many applications, but each has been a customized effort. The group is now developing templates that combine scheduler, performance model, and other software for particular classes of structurally similar applications. Classes amenable to this approach include parameter sweeps (applications that run the same algorithm with various input parameters), master/worker applications, and distributed data applications.

A parameter sweep template is the goal of graduate student Dmitrii Zagorodnov and AppLeS project scientist Henri Casanova. Testbed applications for this template include INS2D, a fluid dynamics code from NASA Ames Research Center, and MCell, a neuron physiology simulator from NPACI partners Tom Bartol and Terry Sejnowski at the Salk Institute.

Graduate student Gary Shao and programmer Jim Hayes are developing another template that will ultimately target master/worker and scatter/gather applications. Shao is developing user-level scheduling techniques to reduce the impact of performance bottlenecks. Hayes has developed a library that handles the scheduling details, allowing the programmer to concentrate on the application problem. Adapting an application to use this template will require only a few new functions linked with the communication library. The template is being tested with several programs, including the DOT molecular docking code from SDSC and The Scripps Research Institute.

"Our earliest AppLeS had to be integrated 'by hand,'" Berman said. "Templates allow classes of applications to achieve good adaptive performance on the metasystem with less effort." --MGend note

Top| Contents | Next
Top| Contents | Next