Implementing and Testing I/O in a Scientific Application on T3D
Elda Rossi, Stefano Evan · 1996
We will describe the implementation of an out-of-core algorithm on the CRAY T3D. The introduction of heavy I/O activity was necessary to handle larger problems, and required particular attention in order to maintain good performances. We were able to nearly reach disk peak performance; we obtained tranfer rate of 38 Mbyte/s using two DD-60 channels on I/O Phase I and 75 Mbyte/s with four DD-60 channels on I/O Phase III. As an application, we used the algorithm to compute the Full-CI energy of the Be 2 molecule with a (9s2p1d) basis set (all electrons), with a Full-CI space of more than one billion (10 9 ) of symmetry-adapted determinants. We needed 27 iterations to get convergence; a single iteration on the CRAY T3D at CINECA (64 processors) required about four hours of elapsed time, 30 minutes of which spent in I/O activity. Due to the scalability of our code, substantially larger calculations could be performed provided that more processors and a larger amount of disk space were available.