Re: New function: secure-random-bytes

[Paul Eggert <eggert@cs.ucla.edu>]

On 06/24/11 02:57, Ted Zlatanov wrote:

> Do you think you could you provide three functions,
> `secure-random-bytes-{urandom,gnutls,openssl}' instead of three versions
> with the same name?

I dunno, I kind of like the simpler interface.
It's better to have the notion of a random-source,
independent of the notion of generating random bytes from it.
Whether that source should be represenated by an atom or
by something else is a different matter.

> I'm not sure what the auto priority should be.  Are the GnuTLS/OpenSSL
> functions better than /dev/urandom?  They are certainly slower.

If they're slower than /dev/urandom, they must be pretty slow.

I suggest using ISAAC-64, which is what coreutils' random-bytes
generators use by default.  (Coreutils originally defaulted to
/dev/urandom, but users complained because that was too slow.)
On my list of things to do is to gnulib-ize the coreutils
ISAAC-64 and random-bytes generators, and I can do that if
there's interest.

Here's one benchmark of the performance difference, run on
Fedora 14 x86-64.  The benchmark creates 1 GB of random data
and outputs it to /dev/null.  With ISAAC-64, it takes 0.48
seconds real-time.  With /dev/urandom, it takes 146 seconds
real-time.  So ISAAC-64 is about 300x faster.

$ time shred -v -n1 -s 1GB /dev/null
shred: /dev/null: pass 1/1 (random)...

real	0m0.476s
user	0m0.397s
sys	0m0.072s
$ time shred -v -n1 -s 1GB --random-source=/dev/urandom /dev/null
shred: /dev/null: pass 1/1 (random)...
shred: /dev/null: pass 1/1 (random)...32MiB/954MiB 3%
shred: /dev/null: pass 1/1 (random)...63MiB/954MiB 6%
...
shred: /dev/null: pass 1/1 (random)...954MiB/954MiB 100%

real	2m25.969s
user	0m0.013s
sys	2m24.189s