[lustre-discuss] Odd behavior with Matlab.
Laifer, Roland (SCC)
roland.laifer at kit.edu
Thu Jun 2 02:31:56 PDT 2016
we had the same problem. The reason is a Lustre bug and a odd way how
Matlab creates directories. This is all clearly described in
New Matlab versions just try to create all directories in the path and
they continue if the return code is EEXIST. Lustre *sometimes* provides
the wrong return code (EPERM instead of EEXIST). This is also described
in the update from 15/Feb/16 9:31 PM in LU-4185. Other applications
possibly have the same problem, too.
The good news is that future versions of Lustre will allow to change the
default behaviour (if I understood the last update correctly). It seems
that it is not planned to fix the bug as default since this would have a
performance impact. The bad news is that it will probably take some time
before the patch is available in stable Lustre versions. (I opened a
case and asked our Lustre support to provide the patch in their Lustre
version.) Also bad news is that Mathworks is not willing to provide a
workaround (we asked them, too).
Am 02.06.2016 um 11:04 schrieb Phill Harvey-Smith:
> Hi all,
> I have recently compiled the lustre clients on Ubuntu 16.04, with help
> from this list I managed to get that working with source checked out on
> May 31.
> In our environment we have the user's home directories on a lustre volume.
> However when a user starts Matlab they have problems :
> R2016a just dumps them straight back to a command prompt with the error:
> Warning: failed to create preference directory /home/stsxab/.matlab/R2016a.
> Starts up, displays the splash screen and then a big page of java
> errors, the first of which is :
> Internal error: No write permission on
> Works as expected.
> When I login with a user who's home directory is on the local
> filesystem, all 3 versions work correctly, which suggests a problem with
> the interaction between Lustre & Matlab.
> Does anyone know how to fix this?
> lustre-discuss mailing list
> lustre-discuss at lists.lustre.org
More information about the lustre-discuss