On Fri, 18 Aug 2023 12:17:51 +0200
Martin Maechler wrote:
> I think it would be nice to provide the average R user with a
> (possibly super small) R package that allows to turn on (and off)
> such CNR reproducibility.
Would it be possible to effect this on/off via options()?
cheers,
> Bill Dunlap
> on Thu, 17 Aug 2023 07:31:12 -0700 writes:
> MKL's results can depend on the number of threads running and perhaps
other
> things. They blame it on the non-associativity of floating point
> arithmetic. This article gives a way to make results
Thanks a lot! I am not aware of this behavior! I will take this into
account in my work.
Regards,
Shu Fai
On Thu, Aug 17, 2023 at 10:31 PM Bill Dunlap wrote:
>
> MKL's results can depend on the number of threads running and perhaps other
> things. They blame it on the non-associativity of
Ah, super interesting! Thanks for the pointer, Bill.
I might not have seen the issue because I have
export MKL_NUM_THREADS=1
in my .profile, so MKL never goes beyond using one thread on my machine.
That's a potential source of non-reproducibility I have not come across. So the
'wisdom' that
Thanks, I was missing the point that this was *non-repeatability on
the same platform*.
On 2023-08-17 10:31 a.m., Bill Dunlap wrote:
MKL's results can depend on the number of threads running and perhaps other
things. They blame it on the non-associativity of floating point
arithmetic. This
Sorry for the confusion caused by that sentence. I did
not expect using the same seed generates the same
sequence of numbers unconditionally. What puzzled me
is not having the same sequence when running the
same code repeatedly in the same computer in the same
session of R. The following four sets
MKL's results can depend on the number of threads running and perhaps other
things. They blame it on the non-associativity of floating point
arithmetic. This article gives a way to make results repeatable:
But I have never seen a case where, on the same machine and using the same math
routines, using the same seed repeatedly generated different values. That is
quite bizarre (the issue I was discussing in the thread is quite simple in
comparison). I tried to reproduce the issue below, but was
> However, should the numbers
> generated identical if the same seed is used?
I don't see how using the same seed can overcome floating-point
differences across platforms (compilers etc.) stemming from differences
in an eigen() computation (based on arcane details like use of
registers,
Hi All,
When addressing an error in one of my packages in the CRAN check at CRAN,
I found something strange which, I believe, is unrelated to my package.
I am not sure whether it is a bug or a known issue. Therefore, I would like
to have advice from experts here.
The error at CRAN check occurred
10 matches
Mail list logo