[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: random numbers on clusters

From: David Bateman
Subject: Re: random numbers on clusters
Date: Fri, 15 Apr 2005 13:40:59 +0200
User-agent: Mozilla Thunderbird 0.8 (X11/20040923)

Michael Creel wrote:

David Bateman wrote:
If your platform as /dev/urandom, then the octave-forge random generators will use it to seed the generators, and you will end up with independent number streams. Even if you don't have /dev/urandom, then these generators use the LSB of the clock to seed with (ie. usec) and so you'll probably find yourself with independent streams in any case.

Hi David and list,

Looking into this, I found the definition

"Least Significant Bit (LSB) The LSB is the right most bit in a binary number. For Example 00000001, 1 is the LSB."

"By default, the generator is initialized from /dev/urandom if it is\n\
available,otherwise from cpu time, wall clock time and the current\n\
fraction of a second.\n\"

I'm guessing that LSB here means decimal rather than binary numbers, but if that's true then when the clock is used to initialize, then there are only 10 possible starting points for the sequence of random numbers. (?) If this is the case then correlated streams on different nodes would be quite likely even if only a small number of nodes are used, and it would be a certaintly on any cluster with more than 10 nodes. I guess that this might be relevant on Windows machines.

Hoping I'm confused, Michael

The relevant piece of code is

/* Look for entropy in /dev/urandom */
FILE* urandom =fopen("/dev/urandom", "rb");
if (urandom) {
while (n<MT_N) {
unsigned char word[4];
if (fread(word, 4, 1, urandom) != 1) break;
entropy[n++] = word[0]+(word[1]<<8)+(word[2]<<16)+(word[3]<<24);

/* If there isn't enough entropy, gather some from various sources */
if (n < MT_N) entropy[n++] = time(NULL); /* Current time in seconds */
if (n < MT_N) entropy[n++] = clock(); /* CPU time used (usec) */
if (n < MT_N) {
struct timeval tv;
if (gettimeofday(&tv, NULL) != -1) {
entropy[n++] = tv.tv_usec; /* Fractional part of current time */

I'd thought the code only took the usec part of the time and cpu time for the entropy so that sequences that started near each other didn't have too many identical values in them. The fact is from the above, the wall time in seconds is used, pluse the cpu time in microseconds. If gettimeofday exists the microsecond part of the wall time is also added to the pool of entropy.... In general this will give you 96 bits of entropy that will be sufficient in most cases to have unique sequences, even in the absence of /dev/urandom...


David Bateman                                address@hidden
Motorola Labs - Paris +33 1 69 35 48 04 (Ph) Parc Les Algorithmes, Commune de St Aubin +33 1 69 35 77 01 (Fax) 91193 Gif-Sur-Yvette FRANCE

The information contained in this communication has been classified as: [x] General Business Information [ ] Motorola Internal Use Only [ ] Motorola Confidential Proprietary

Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:
How to fund new projects:
Subscription information:

reply via email to

[Prev in Thread] Current Thread [Next in Thread]