Re: Post-doc position proposal

From: Cristiano (cristiano.pi@nsquipo.it)
Date: 03/07/03 

From: "Cristiano" <cristiano.pi@nsquipo.it>
Date: Fri, 07 Mar 2003 20:07:14 GMT

Seznec Andre wrote:
>
> HAVEGE combines entropy/uncertainty gathering from the internal
> hardware volatile states of the processors with pseudo-random
> number generation.

Although I think you generator has little entropy, it is very likely
that its complexity is extremely high (*perhaps* higher than a crypto
hash).

> HAVEGE presents an unprecedented throughput for an empirically
> strong random number generator: than 100 Mbits/s on recent PCs and
> workstations.

This is an other good reason to do some experiment, but you say:

"As we do not not know if the compiler optimization options vary from
implementation/installation of Visual C++ to another, we recommend to
always let HAVEGE2.0.obj unmodified."

and this is a big problem.

Perhaps you remember our old conversation about this issue.

You said that we must check the size of the inner loop in this way:

"I used the rather simple following way:
maintain a constant number total number iterations (that is the number
of iterations in the loop times the number of unrolled iterations).
Time the execution (with hardware clock counter), 10 times to avoid
accident associated with unfortunated other activity.
Vary the number of unrolled iterations: when there is a knee in the
curve, you are exceeding the I-cache size."

Perhaps someone, other than me, is interested in this topic.
Please, could you explain in a better way what I should vary and what I
should maintain constant?

Thank you
Cristiano