If the assumption is made that random number

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: external e ects and are thus, for all practical purposes, impossible to keep aligned in time in a predictable way. If the assumption is made that random number generation from a single generator will occur, across processors, in a certain predictable order, then that assumption will quite likely be wrong. A number of techniques have been developed that guarantee reproducibility in multiprocessor settings and with various types of Monte Carlo problems. We will consider only simple extensions to our previous discussion of LCGs, but acknowledge that there are many approaches to parallel random 26 number generation in the literature. The rst situation we address involves using LCGs in a xed number of MIMD processes, where that number is known at the beginning of a run. Suppose we know in advance that we will have N independent processes and that we will need N independent streams of random numbers. Then the best strategy for using an LCG is to split its period into nonoverlapping segments, each of which will be accessed by a di erent process. This amounts to nding N seeds which are known to be far apart on the cycle produced by the LCG. To nd such seeds, rst consider (for c = 0), the LCG rule successively applied: Xn+1 = aXn (mod m) Xn+2 = aXn+1 = a2Xn (mod m) Xn+3 = aXn+2 = a3Xn (mod m) ::: Xn+k = aXn+k = ak Xn (mod m) Thus we can \leap ahead" k places of the period by multiplying the current seed value by ak mod m. For our purposes, we would like N starting seeds, spaced at roughly k = P=N steps apart. Since k is likely to be quite large, it is not practical to compute ak one step at a time. Instead we compute an (L + 1)-long array, d, the power-of-two powers of a: d0 = a d1 = d02 d2 = d12 ::: dL = dL;1 2 where dL is the largest power-of-two power of a that is still smaller than k. I.e. L is the integer part of the log (base 2) of a. For example, assume that k = 91 = 10110112 (very small, but big enough to show how it works). Then d = (a a2 a4 a8 a16 a32 a64) and since 91 = 64 + 16 + 8 + 2 + 1, then a91 = a64 a16 a8 a2 a1 = d6 d4 d3 d1 d0 Thus for any k, ak = di for all i for which bit i in the base-two representation of k is a one. Therefore we can leap ahead by k cycle steps with no more than log2k multiplies. Once ak is computed, the N seeds can be determined by the procedure: Choose seed1 Remarks 27 seed2 = ak seed1 (mod m) seed3 = ak seed2 (mod m) ::: seedN = akseedN ;1 (mod m) With these seeds, each of the N processes will generate random numbers from nearly equallyspaced points on the cycle. As long as no process needs more than k random numbers, a condition easily met for some applications, then no overlap will occur. Everything just said for MIMD processes applies equally well to SIMD programs, where the number of random number streams needed is (usually) known at run time. The development of the leap ahead technique just described assumed that c = 0 in the LCG rule. For c 6= 0, Leap ahead can still be accomplished in a similar way, if one constructs the log2k-long array of partial sums of the form: sj = Pj=0 ai where, as before, j is a power i of two. The details are left as an exercise for the reader. The second and more di cult case to consider is when we do not know at the beginning of program execution how many processes (generators) we will need. The splitting of processes in such programs are data driven and in most cases occur as the result of prior Monte Carlo trials taken many steps earlier. The problem is to spawn new LCG seeds in a way that is both reproducible and which yields independent new streams. Here we only mention a generalized approach that works within limits. Further details can be found in Frederickson, et al., 1984] Consider an LGC with the property that each X has two successors, a \left" successor, XL, and a \right" successor, XR . These are generated as follows: L(X ) = XL = aLX + cL (mod m) and R(X ) = XR = aRX + cR (mod m) Figure 7 shows the action of these operations with respect to a starting seed X0 . Taken separately, the XL and XR sequences are simple LCGs that traverse the set f0 1 2 : : : m ; 1g in di erent order. Alternatively (and the method in which these generators are typically used), the XL rule produces a pseudo-random leap-ahead for the XR sequence, thus deterministically producing a seed for a new, spawned, subsequence of the \right" cycle. With such a mechanism that uses only local information from a process, reproducibility can be established. Frederickson gives a formula for the selection of the constants in the succession rules that satis es a particular independence criterion, given some constraints. The interested reader is referred to Frederickson, et al., 1984 for further enlightenment. Exercise 8 - Vectorization of Cray's LCG Complete Exercise 6 above to determine the parameters a and m (probably m = 248) of Cray's ranf(). Develop a vectorized version of ranf() by creating a vector of successive multiples of the coe cient a. For...
View Full Document

Ask a homework question - tutors are online