This blog only applies to M60 / M6 (Maxwell generation) GRID cards. It is not relevant to GRID 1.0 (K1/K2 cards).
Try this first!
If you are experiencing issues with a new M60 / M6 card, please run this command:
- lspci –n | grep 10de
You should get something that looks like:
[root@SAGRID-ESXi5:~] lspci –n | grep 10de
0000:04:00.0 Class 0300: 10de:11bf [vmgfx2]
0000:05:00.0 Class 0300: 10de:11bf [vmgfx3]
0000:08:00.0 Class 0300: 10de:11bf [vmgfx0]
0000:09:00.0 Class 0300: 10de:11bf [vmgfx1]
0000:86:00.0 Class 0300: 10de:11bf [vmgfx4]
0000:87:00.0 Class 0300: 10de:11bf [vmgfx5]
If your system reports 0302 instead of 0300, the GPU is in “compute” mode. If it reports 0300 as above then it’s in “graphics” mode and this is not the cause of your issue.
If you notice the GPUs are in compute mode – then use the gpumodeswitch utility and change it to graphics, make sure you reboot the server to apply these changes. The gpumodeswitch utility is supplied in the driver download and fully documented. I have blogged about it before.
NVIDIA’s Jason Southern has published a video to guide you through the utility too: https://www.youtube.com/watch?v=VAQhiNNFXxQ&feature=youtu.be
- Side note: lspci is a standard utility on many unix like OSs that probes pci devices (such as GPUs) – “10de” is an identifier allocated to the NVIDIA as a vendor, so this command is picking out just the information on NVIDIA pci devices from the massive list of pci devices you may have.
By default M60 / M6 cards are currently (April 2016) shipped to OEMs (server vendors) in “compute” mode and many so not switch them over when shipping a server designed for GRID remoting graphics.
What Symptoms / problems could you see if in “compute” mode?
1. VMs (Virtual Machines) fail to power on
- You see an error message such as: ” Could not initialize plugin “/usr/lib64/vmware/plugin/libnvidia-vgx.so” for ‘GPU ‘grid_m60-2q
2. An error is seen running the nvidia-smi command/utility (after loading the vib on ESX – see below on checking vib is loaded correctly)
- “NVRM: BAR1 is 8192M @ 0x387c00000000 (PCI:0000:83:00.0)NVRM: This is a 64-bit BAR mapped above 16 TB by the systemNVRM: BIOS or the VMware ESXi kernel. This PCI I/O region assigned
NVRM: to your NVIDIA device is not supported by the kernel.
NVRM: BAR1 is 8192M @ 0x387800000000 (PCI:0000:84:00.0)”
3. ECC Protection is enabled (this means you are in compute mode)
4. I/O Base BAR is disabled (this means you are in compute mode)
5. A Memory BAR of 8 gigabytes is being used (this means you are in compute mode)
On hypervisors with a full version of linux such as XenServer (VMware does not have a full version) you can use
- [root@localhost libvirt]# lspci -vvvv -d “10de:*” | grep Region
To check if GPU BARs correctly assigned on ESX you should use:
- # esxcfg-info -a | less
And scroll or search down to Tesla M60 label. ESX doesn’t have full version of lspci and the linux command “lspci -vvv” won’t work.
6. The memory on the GPU is missing 0.5 gigabytes of RAM
- i.e. in vCenter see RAM of 7.49GB rather than the expected 7.99GB, this could be a sign you are in compute mode
Side note: Other errors from nvidia-smi
- “NVIDIA-SMI has failed because it couldn’t communicate with Nvidia Driver”
This message usually means that driver is not loaded or incorrect version, so after running nvidia-smi, type:
And check all NVRM entries output.
But first check: for ESX/vSphere that the vib is loaded correctly
# esxcli software vib list | grep -i nvidia (to see if VIB is installed)
# esxcli system load -m nvidia (it should returns something like “Unable to load module /usr/lib/vmware/vmkmod/nvidia: Busy” that means driver is loaded)
Also if you catch any other message or error, please, use dmesg to check output info about the error.
Symptoms may have multiple root causes!
The symptoms above are signs you may need to switch from “graphics” to “compute” mode, but if you check and all is ok or change the mode and have issues after reboot then you probably have another issue. M60 and M6 cards come with full support and you should raise a support ticket with NVIDIA.
You can of course discuss issues on our support forums:
Citrix Customers should remember they have an automated support system to which you can upload logs and it will check for mis-configurations so it can be a good first check when encountering issues. It will also check for known issues and make recommendations for security patches etc. I’m not sure if they can check for GPU mode configuration yet, but I’ll be contacting them to see if possible:
- Service is here: https://cis.citrix.com/
- Very old user review of the service: https://adamgamble.org/2014/04/23/the-citrix-support-tool-half-of-you-dont-use-but-should/
- Citrix’s own (TaaS) CIS blog about the service: https://www.citrix.com/blogs/tag/citrix-insight-services/