CS240A: Homework Assignment 0

Application Problem - Prediction of Storms

History

(Source: ARPS website)

Since 1993, the application was run on Cray C90(1993 rank 19) using 16 processors with a peak performance of 15.6 GFLOP/s. By 1995, using Cray T3D (1995 rank 18) -a vector super computer, 512 processors, with distributed Globally Addressable Memory, and peak performance of 76.8 GFLOP/s- at Pittsburgh Supercomputer Center (PSC) , CAPS achieved a milestone by successfully able to predict the location and structure of individual storms 6 hours in advance. This was further improved in 1996, when the location and timing of the storm was predicted seven hours in advance. The forecast was run in 100 minutes on 256 processors of the T3D.

With many of the improvements in the modelling and analysis, it has been mentioned in the article¹ that, the system could correctly predict around 80% of the storms, which was unprecedented. Furthermore, the system was able to correctly forecast a storm even when the data did not indicate any such conditions. Similarly, the system could successfully predict no storm on a day when the conditions suggested a high possibility of a storm. In recognition of its achievements, CAPS was awarded the Discover Magazine Award for Technological Innovations, and ComputerWorld-Smithsonian Award in 1997. Some of the pictures produced below clearly indicate that storm predictions do indicate high degree of agreement with the actual observation. However, the quality of such predictions is not always consistent. The ARPS system was further improved to include ensemble forecast to handle such uncertainty.

ARPS vs. Reality: May 24, 1996
The ARPS forecast (left) created at noon Oklahoma time for conditions at 5 p.m. compares well with an actual radar image at 5 p.m. Color indicates rainfall intensity, increasing from light blue to pink. The green square shows a three-kilometer fine-grid nested within the larger nine-kilometer grid of the forecast area. ARPS centers the small grid in the area where storms are likely to develop based on the large-grid forecast. Even though no storms were present in the forecast area at noon on May 24, ARPS successfully predicted the timing and location of the storm line that developed later that day (produced from article Weather Forecasting: Faster Than A Speeding Storm).

(source: CAPS website)

Current Application

It is clear from the information provided above that the application scaled well from a 16 processor (Cray C90) to 256 Beowulf cluster (running current version), although the current version also includes many improvements,and enhancements over the older versions of ARPS running on Cray C90. It also seems that the application was successfully ported from a Vector platform ( Cray C90, and T3D) to Linux Pentium4 Xeon (Beowulf) Cluster, also using SGI Origin 2000 machine in between.

References