The performance of the GTF and LTF algorithms will depend strongly on the type of parallel machine used. A machine using NFS with fast local disks and communication is likely to perform better with the LTF algorithm. However, a system with fast (parallel) filesystem and fast communication can actually perform better with the GTF algorithm. In the following we will consider test cases run on very different machines:
We have run 2 test cases to analyze the behavior of the different algorithms on different machines. The small test case was designed to execute on the PPro system. It uses a small line database (about 550MB) with about 35 million lines of which about 7.5 million lines are selected. The second test uses an about 16 times larger line database (about 9GB) and also selects 16 times more lines. This large test could not be run on the PPro system due to file size limitations and limited available disk space. The line opacity calculations were performed for about 21,000 wavelength points that are representative for typical calculations. The tests were designed for maximum I/O usage and are thus extreme cases. In practical applications the observed scaling is comparable to or better than which found for these tests and appears to follow the results shown here rather well. In the following we will discuss the results for the tests on the different computing systems.