NYC-WTC_000187725

NYC 9/11 Public Portal Document

128

CHAPTER FIVE

SOFTWARE AND HARDWARE REQUIREMENTS AND SOLUTION OF THE MODELS ON COMPUTER NYREG Regional Equilibriiun Model

There are many aspects of a particular modeling problem that affect the cost of its solution. We have studied three of these in detail. The first is the choice of the computing environment, which is a subject of this chapter. It is clear that a supercomputer should be more efficient than a conventional IBM mainframe or a SUN workstation by an order of about two to eight times depending on whether the program is vectorized or not. Because, supercomputers are actually becoming more accessible than mainframes and workstations, and seem to be less affected by budget constraints, some of the NYREG computations were performed on the CRAY YMP-1 and on the CRAY-2. This hardware was accessible through the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Urbana-Champaign, via modem from Northwestern University. The code (software) which operationalizes the algorithm is written in FORTRAN which is one of the most widely used codes for mathematical computation. While the program has been solved on the CRAY-2 at the NCSA, the final version of the model has also been run on an IBM mainframe and on a SUN workstation. Hence, the code can easily be adapted to any of these computational hardware environments. Computational costs on an IBM mainframe and SUN workstation are similar. other factors that affect costs of modeling include the starting point from which model convergence begins. This is very important in determining the speed with which the solution reaches equilibrium. In formulating NYREG, we have found that starting points in which rents are set to low values give much faster results. In addition, the modeler should consider the aggregate vacancy rate of the housing market. It strongly affects the speed of convergence. ’’Tight” housing markets with a low vacancy rate such as 3% or less will take longer to converge than ’’loose” markets with 2 0% or higher vacancy rates. The combined affect of these factors on costs are discussed using several illustrative problems.

Initial Computational Experiments on the CRAY For comparative purposes, we present the computational results

■sad

CITY_21MC100_006662359

NYC-WTC_000187859