Learning Forum

[pbrunk]

We're running ANSYS 19.2, and its Electronics Desktop, on a bunch of CentOS 7.5 nodes.  All are headless and most lack discrete GPUs.  A handful have some model of Tesla GPU installed.

We're having a problem with libGL and libswrast.  When running "ansysedt" for the Electronics Desktop, for example, if I click "Project", then "Insert HFSS Design", I get

 libGL error: failed to load driver: swrast.so

Without mesa-drivers-dri installed, then I also get this error

 libGL error: unable to load driver: swrast_dri.so

From strace I can see that nodes without mesa-dri-drivers installed lack libswrast altogether, and those with it installed can't use the libswrast that the RPM provides.

Some other instances of Ansys use on our cluster might see these same symptoms--I don't know.

If I set Ansys to use only software rendering, then those libGL errors go away, but the program still freezes, just as reported in the thread

 /forum/forums/topic/libgl-error/?order=all#comment-a9ac6bb7-6081-46ff-b60f-a9ac01561f7b

All nodes show one of "ast" or "mgag200" (depending on motherboard) as the output of "lspci -nnk | grep -i vga -A3 | grep 'in use'".

I think all required RPMs, including i686, are installed on these nodes.

For GUIs, our users SSH (from whatever platform) into our cluster, with X11 forwarded.

"cat /etc/*release" on all nodes looks as below.  I know Ansys 19.2 doesn't officially support CentOS 7.5, but suspect that the GL swrast problem is independent of that.

/etc/centos-release:
CentOS Linux release 7.5.1804 (Core)

/etc/os-release:
NAME="CentOS Linux"
VERSION="7 (Core)"
ID="centos"
ID_LIKE="rhel fedora"
VERSION_ID="7"
PRETTY_NAME="CentOS Linux 7 (Core)"
ANSI_COLOR="0;31"
CPE_NAME="cpeo:centos:centos:7"
HOME_URL="https://www.centos.org/"
BUG_REPORT_URL="https://bugs.centos.org/"
CENTOS_MANTISBT_PROJECT="CentOS-7"
CENTOS_MANTISBT_PROJECT_VERSION="7"
REDHAT_SUPPORT_PRODUCT="centos"
REDHAT_SUPPORT_PRODUCT_VERSION="7"

/etc/redhat-release:
CentOS Linux release 7.5.1804 (Core)

/etc/system-release:
CentOS Linux release 7.5.1804 (Core)

[pbrunk]

Here's 'glxinfo' info:

 'glxinfo | grep GL' shows this on the GPU-less nodes:
libGL error: No matching fbConfigs or visuals found
libGL error: failed to load driver: swrast
Error: couldn't find RGB GLX visual or fbconfig
276 GLX Visuals
359 GLXFBConfigs:

'glxinfo' outputs this on a GPU-bearing node:
X Error of failed request:  BadValue (integer parameter out of range for operation)
  Major opcode of failed request:  154 (GLX)
  Minor opcode of failed request:  24 (X_GLXCreateNewContext)
  Value in failed request:  0x0
  Serial number of failed request:  37
  Current serial number in output stream:  38

[tsiriaks]

Hi pbrunk ,

Thank you for the info and for following the information in this thread

/forum/forums/topic/libgl-error/?order=all#comment-a9ac6bb7-6081-46ff-b60f-a9ac01561f7b

It seems you do your homework well.

Unfortunately, as also stated in that thread, your best bet for the systems that you described would be to ask users to install AnsysEM on their local machines, which must have dedicated graphics card, then do all their pre/post processing works on the local machines and only submit the jobs to solve on this Linux cluster. You just need to setup AnsoftRSM (if Windows to these Linux cluster nodes) or simply setup to point to the submit node of the cluster since AnsysEM has tight integration with some schedulers. Ref:

https://ansyshelp.ansys.com/Views/Secured/Electronics/v192/Subsystems/HFSS/HFSS.htm#HPC/WindowstoLinuxJobSubmission.htm

Thanks,

Win

 

[tsiriaks]

Sorry, I forgot another possible workflow, users can also do all their pre/post processing works on the local machines and transfer the files to the cluster , then use batch submission commands to submit the jobs to solve on the cluster , then transfer result files back to their local machines.

Thanks,

Win

[pbrunk]

 hanks.  Are the NVIDIA Tesla GPUs, in the few nodes that have them, of use here?

The Nvidia driver and libraries are installed on our nodes via RPMs.  I did a yum reinstall" of them on a GPU node.  'ansysedt' still has the same libGL swrast problem.  But now, "glxinfo |grep OpenGL" shows this:

[n584 ~]$ glxinfo |grep OpenGL
OpenGL vendor string: Brian Paul
OpenGL renderer string: Mesa X11
OpenGL version string: 2.1 Mesa 7.7.1
OpenGL shading language version string: 1.20
OpenGL extensions:


Can we tell whether this GPU node any closer to having HFSS function inside Ansys EDT, compared to the GPU-less nodes?

[tsiriaks]

Depends on Tesla model, type of analysis, and how the job is run, you 'might' be able to take advantage of Tesla for (job) computing (solving) related. I don't think this helps you run the GUI interactively.

Unfortunately, it still has to be one of the workflows that I mentioned earlier. 

Or, you can try setting up a test Windows machine (even without the great/dedicated graphics card) and see if that works. Our software seem to be more 'flexible' on Windows than Linux when there is no tested/supported graphics card installed on the machine.

Thanks,

Win