A new class of random number generators required for advanced computer architectures

This is the fiial report of a three-year, Laboratory-Directed Research and Development (LDRD) project at the Los Alamos National Laboratory (LANL). The advent of ever more powerful computers allows one to run Monte Carlo computations of unprecedented length. Currently used random number generators (RNGs) do not have the cycle length necessary for these computations. It is possible to cycle completely through most RNGs used on workstations in a few minutes computations. Even having a long period may not qualify a RNG as suitable. We are developing tests that will allow us to develop high quality RNGs for use in long computations. 1 . Background and Research Objectives Reliability of Monte Carlo methods ultimately rests upon the quality of "random" numbers used in a computation. These numbers are not truly random, but are generated by some deterministic mathematical process. Such random number generators (RNGs) can only produce a finite sequence of numbers before repeating themselves. Advanced computer architectures and fast scientific workstations are capable of exhausting the entire sequence of many presently used generators in a single computation. A new class of RNGs is required or else these new computing capabilities will be of limited value for complex Monte Carlo simulations. Many of the currently used RNGs are inadequate for the future. For example, many computations on workstations are done with an RNG whose cycle is only 232 (about 4,000,000,000). A 25 Mhz workstation that takes 10 cycles to generate a new random number would exhaust the RNG in about 30 minutes. In addition, the cycle need not be exhausted to exhibit problems. For example, if one were to run a problem on a lattice of size 2563 using a *Principal investigator, e-mail: ttw@lanl.gov