Forecasts from the Network Weather Service

PARTICIPANTS
Neil T. Spring, Univ. of Washington
Jim Hayes, UC San Diego
Martin Swany, University of Tennessee

o choose the fastest line at the bank or supermarket, you have to make a prediction based on the length of the line
and the needs of the other people in line. Metasystems, which combine geographically dispersed computing
and data resources into a single virtual computer, have to make similar choices between seemingly equivalent resources on networks. University of Tennessee computer scientist Rich Wolski has created the Network Weather Service (NWS) to dynamically forecast the usability of system resources. NWS makes this information available to automated schedulers and to humans who want to gauge the network's quality of service.

"The Network Weather Service both monitors and dynamically forecasts the performance that various network and computational resources can deliver to an application," Wolski said. "Just as in weather forecasting, predictions are based on how the system has behaved in the past under similar conditions."

EVOLVING PREDICTIONS

REFERENCE
Rich Wolski, Neil Spring, and Jim Hayes. "The Network Weather Service: A Distributed Resource Performance Forecasting Service for Metacomputing." Journal of Future Generation Computing Systems, 1999 (in press).

Figure 1. Network Weathervane

Top: Four days of bandwidth measurements between UC Santa Barbara and Kansas State University. Throughput peaks early in the morning, then drops later in the day as users load the system. Bottom: SWEB project data show the corresponding NWS forecasts. The large variance in the raw data would foul up a forecast based solely on current measurements.

EVOLVING PREDICTIONS

NWS gathers performance data from sensors--network monitors, CPU monitors, and so on--distributed throughout the system. It uses several different numerical models to generate internal forecasts of conditions in the near future. As time goes by, it tracks the accuracy of the models' predictions, and uses the model with the lowest cumulative error at any given moment to generate a forecast for an application. This method not only gives excellent results, but also allows new predictive models to be added and compared.

Wolski's team is putting together an NWS Java sensor for the NPACI HotPage. "You will be able to watch the network forecast from your laptop or desktop computer--wherever you are--to SDSC and back simply by accessing the HotPage," he said. "No other center in the country will be able to offer this kind of 'live' connectivity analysis to users."

Other NPACI projects are benefiting from the maturation of NWS. The AppLeS scheduler now makes extensive use of NWS facilities, and Wolski's group has developed prototype implementations for Legion and Globus. Each prototype forecasts process-to-process network latency and bandwidth and the available processing percentage for each machine that it monitors.

"We are also building a small but brave user community," Wolski said. Jack Dongarra's group at the University of Tennessee is integrating NWS with NetSolve. And Dan Andresen's SWEB project at Kansas State University is using it to do Web server selection in conjunction with NPACI's Alexandria Digital Library project at UC Santa Barbara (Figure 1).

News of NWS is spreading beyond NPACI. The Ninf project (Ninf: A Network-based Information Library for Global World-Wide Computing Infrastructure) under Satoshi Sekiguchi of the MITI Electrotechnical Laboratory in Tsukuba, Japan, is using NWS to control its computer system and in conjunction with their simulation efforts. Hidemoto Nakada is visiting UC San Diego to help with the software integration effort.

NWS has been under development for several years, and both the system and the project have evolved. Wolski has moved to the University of Tennessee and gradually is transferring the main development effort there, where graduate student Martin Swany soon will take over as the main NWS programmer from UC San Diego researcher Jim Hayes, who has recently redesigned the NWS Application Program Interface.

"Jim Hayes has done a wonderful job of hardening release 1.1.1.1," Wolski said. "The NWS has run on NPACI machines for quite a while without any substantial down-time due to software failure--the system itself has failed only once in the last 18 months. The NWS is one of the first computational grid services to achieve this kind of robustness record." --MG