Bug#684344: openmp problematic

[Sébastien Villemot]

Le mardi 04 février 2014 à 23:17 +0100, Julian Taylor a écrit :
> There are well known issues with the gnu openmp variant of openblas in
> respect to forks of applications.
> This causes numerous issues especially in the python world where forking
> is very common.
> see e.g. https://github.com/xianyi/OpenBLAS/issues/294
> http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=737675
> 
> As openblas is integrated into Debians blas alternative system it needs
> to be rock solid compared to ATLAS and netlib blas.
> This is known not to be the case for openmp openblas but somewhat
> working with pthreads.
> As my viewpoint is very narrow coming from python I want to ask if there
> are known and ideally reproducable cases where pthread openblas breaks
> applications but openmp openblas does not?

Please find attached an example of such a breakage, using
parallelisation in R. The example is run simply with:

  R --vanilla < foo.R

On sid, with openblas 0.2.8-3 (OpenMP), the program runs and reports 8
condition numbers.

On sid but with openblas 0.2.8-2 (pthreads), the program hangs.

-- 
 .''`.    Sébastien Villemot
: :' :    Debian Developer
`. `'     http://www.dynare.org/sebastien
  `-      GPG Key: 4096R/381A7594

-------------- next part --------------
library(parallel)

num.cores <- 4

f <- function(x)
    {
        n <- 40
        A <- replicate(n, rnorm(n))
        B <- replicate(n, rnorm(n))
        rcond(A %*% B)
    }

z <- 1:8

mclapply(z, f, mc.cores = num.cores)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part
URL: <http://lists.alioth.debian.org/pipermail/debian-science-maintainers/attachments/20140212/5b1b2aa3/attachment-0001.sig>