Vadim Baranovsky wrote in a message to Mike Bilow:

 MB> Hey, that's not what I said when I gave you the ELF hash code!  I told
 MB> you it would produce a uniformly distributed seed, not a random one.
 MB> There's a big difference there.

 VB> I use for random seed the combination of Os/2 milliseconds
 VB> timer and Unix time. Then I hash its by MD5 and use as
 VB> initial seed for free running random generator(implemented
 VB> as thread) and use its(random generator) production for
 VB> DES/Russian_GOST encription - it produce very good results.

This has the problem that MD5 will produce a 128-bit value from a 64-bit
input, because the milliseconds clock and Unix time are each 32 bits.  This
has some complicated issues in terms of the distribution of the results. 
No one really understands the math behind distribution analysis of
cryptographic hashes -- which is why they are useful -- but all such
attempts to analyze them start out with the assumption that the message to
be digested is large, at least as compared to the digest.

By contrast, a simpler non-cryptographic hash such as ELF is known to be
well distributed even for short inputs.  This will usually promote better
patterns if the goal is to seed a pseudo-random number generator.

When constructing a seed for certain kinds of pseudo-random number
generators, especially the linear congruential kind which is most commonly
used, the partial dsitribution of the sequence may not be very good.  There
are well known techniques for working around this, either by using only the
least significant half of the generated bits or by running two linear
congruential generators and taking the combined result.

We are well beyond the rational level of concern for a graphics generator.

 VB> P.S. Sorry for my poor English :-(

 VB>    Best wishes, Vadim!
 VB>    Santa-Fox Team
 VB>    E-Mail:  Vadim{at}tso.emissia.spb.su

Your English is perfectly understandable.

I can't believe I'm discussing cryptography with someone in Russia...
 
-- Mike


---