19 November 2008—Of the many things that have gone wrong on Wall Street this past year, the use and misuse of computational algorithms meant to give financiers a clear picture of the risk of big losses was one of them. One important calculation, called Value-at-Risk (VaR), is a way of assessing the probability that an investment portfolio will lose a specified value over a certain period of time. Though VaR’s reputation is much maligned, experts say firms have little choice but to continue, if not accelerate, their use of computational algorithms as the need to calculate risk and value has become more acute. The surviving financial firms might get a little help from code produced for Cell, the processor behind Sony’s PlayStation 3 as well as a number of high-end IBM servers and supercomputers.

By taking advantage of Cell’s unique architecture, computer scientists at Georgia Tech, in Atlanta, say they have found a way to accelerate the generation of random numbers up to 33-fold compared with what’s possible using commercially available Intel or AMD processors. Random-number generation is key to many so-called Monte Carlo simulations, but as a proof of concept, the researchers built their random-number generator into a program that efficiently runs a VaR algorithm. David A. Bader, executive director of high-performance computing at Georgia Tech, says his group has been working with several financial firms, whose names he would not disclose due to legal agreements, on their use of the VaR software. Bader says the source code behind the random-number generators for Cell will be made available during the SC08 supercomputing conference this week in Austin, Texas.

”What we’re able to do is generate not just a single random number fast, but streams of random numbers that can be fed into Monte Carlo simulations,” says Bader, who is also a professor of computational science and engineering.

The basis of many scientific and financial computational algorithms, Monte Carlo methods all start with some random numbers, then perform an operation on them, and repeat the process to arrive at a result. (Imagine playing the game Battleship, for example.) Often the most time-consuming part of a Monte Carlo computation is the generation of a good set of random numbers (or pseudorandom numbers, as computers don’t produce true random numbers). Bader and his graduate students Aparna Chandramowlishwaran and Virat Agarwal adapted two well-known random-number algorithms to Cell: the 64-bit linear congruential generator (LCG) and the 32-bit Mersenne twister. With the LCG, their speed increased 33-fold and with the Mersenne twister, between 14-fold and 29-fold.