Crashing XenServer 7 with Nvidia M60 (and Dell R730)

During the last weeks we had to – or better are – facing a strange problem with Dell R730 / Dell R7910, NVIDIA M60 and Citrix XenServer 7.0. It’s not solved until know. I will update this Blog-Post when there are new findings (or solutions). Before I come to the problem I would like to give you some background information’s about the whole project:
Our Company gets a new CAD program. For the training we decided to use Virtual Desktops with Graphics acceleration instead of classical Workstations. Round about 40 users are trained at the same time. Thus we discussed the topic with Dell and at the end we bought (on recommendations from Dell) the following Hardware (twice):

Dell 7910 Rack Workstations
512GB RAM
2X Intel Xeon E5-2687W v4 (3.00GHz)
1X NVIDIA M60

We did some initial testing’s and everything worked without any problems. With ~10 Key Users we “simulated” a training to check how many users we get on one machine and if we have any problems. Luckily a small NVIDIA vGPU profile (M60-0B – yes B) was fast enough. So we can handle up to 64 users with the two Workstations – more than enough for our trainings.

Due to some problems in the project the trainings had to be delayed – fine for us because we thought everything was working and we had nothing to do in this area. In the meantime we checked the current CAD Hardware. This check showed that around 30 clients must be replaced before the new CAD software goes live – because it will not work on these clients. Another 30 need to be replaced during the year. A lot of discussions and calculations started if we now start to provide productive CAD users also a Virtual Desktop instead of a classical physical one. Again we had discussions with Dell and ended with the following configuration:

Dell R730
576GB RAM (512 for VMs – 64 for Hypervisor)
2X Intel Xeon E5-2667 v4 (3.20GHz)
1X NVIDIA M60

The system was ordered three times (for redundancy). At this point it was not clear how many users we can get on one Server – we planned to use these systems to do more tests and find the best NVIDIA vGPU Profile which fulfills the user requirements.

After the hardware arrive it was directly installed. At the same time the trainings started – our problems as well. After a few training days suddenly one of the R7910 froze – no machines or the Hypervisor (XenServer 7) itself reacted any longer. We had to hard reboot the whole system. These continued to happen on both systems in the next days. There were no Crash Dumps on the XenServer or Event-Logs in the iDRAC Server Log. Thus we decided to activate one of the R730 to get a more stable training environment and investigate more relaxed the problem. However – they also had a problem. Instead of freezing there was a hard Hardware-Reset after some time. In the iDRAC Server Logs the following errors appeared:

Fri Dec 02 2016 10:43:16    CPU 2 machine check error detected.   
Fri Dec 02 2016 10:43:10    An OEM diagnostic event occurred.   
Fri Dec 02 2016 10:43:10    A fatal error was detected on a component at bus 0 device 2 function 0.   
Fri Dec 02 2016 10:43:10    A bus fatal error was detected on a component at slot 6.   

And shortly after that:

Fri Dec 02 2016 10:45:24    CPU 1 machine check error detected.   
Fri Dec 02 2016 10:45:24    Multi-bit memory errors detected on a memory device at location(s) DIMM_A3.   
Fri Dec 02 2016 10:45:24    Multi-bit memory errors detected on a memory device at location(s) DIMM_B1.   
Fri Dec 02 2016 10:45:24    A problem was detected related to the previous server boot.

Furthermore the boot screen showed the following message:

Ok – Time for a Ticket at Dell. The support person told me that there are two known problems when the M60 is used:

1. The M60 is connected with a wrong cable and does not get enough power
2. The power cable of the M60 is in the Airflow of the card

We checked both. The results were quite promising:
R7910:
A wrong cable was used – so the M60 can’t get enough power. The wrong one looks like this:

The correct one has two plugs instead of one. It’s an CPU-8-pin auxiliary power cable (see Nvidia M60 Product Brief – page 11): 

The side with one plug is connected to the M60.

Of course this cannot be directly connected to the power-plug on the riser card. Thus you also need the following cable: 

As you can see this has also two plugs on one side – connect these two plugs with the two from the correct cable and the white with the rise card.

Another thing you need to make sure is that the power cable is not in the Airflow of the M60 cooling. It looks like it often happens that the cable is not under the M60 (like shown in the picture) and instead behind it.

The rest of the cable fits next to the card.

Furthermore Dell advised us to change the Power Supply Options. You can find them on the iDRAC: Overview => Hardware => Power Supplies. On the right side there is now an Option Power Configuration.

They suggested changing the Redundancy Policy to Not Redundant and Hot Spare to Disable.

After correcting this it first looked like it fixed the problem – unfortunately it did not. After a few days in production on of the 7910 froze again. Later I discussed this topic again with another Dell Engineering. After that we changed the setting back to the following:
Redundancy Policy:

You can find a detailed description of the settings in the iDRAC User Guide on page 161.

At this point Dell thought they have another customer with the same configuration (Dell R730 + Nvidia M60 + XenServer 7). Therefore they started to check for other hardware differences between both systems. They found the following differences and replaced our ones with the one of the other customer:

Power Supply:
We had one from Lite-On Technology.

In addition the other had one from Delta Electronics INC.

Furthermore their motherboard had a different revision number. It was changed to one with the same revision. (sorry didn’t make a picture of both). To make sure even one M60 was replaced.

No replacements made any difference. Later we figured out that the other customer was using XenServer 6.5 and not 7. During all the tests I found that rebooting 32VMs several times (often just one reboot was enough) led to the problem – thus it was reproducible

In the meantime the case was escalated. Dell did many Hardware replacements (really uncomplicated) but nothing changed. Interestingly even now it didn’t look like Dell Engineering was involved. We created a case at NVIDA and contacted System Engineers from NVIDA and Citrix. In parallel I created a post on the NVIDIA Grid Forums – especially BJones gave some helpfull feedback.

After discussing the problem we did a few more tests:

Testing different vGPU Drivers:

The Problem was always the same

There had been more tests – but to be honest I don’t remember every detail.

One of the things we also changed was to add the iommu=dom0-passthrough parameter to the xen boot-line. Although we didn’t have the problem that the driver inside the VM did not start – the behavior changed a little bit. It first looked like the problem was fixed – but at the end only more reboots had been necessary to crash the system. After that a crash dump appeared on the XenServer. This had never happened before. Unfortunately the crash dump didn’t contain any useful information.

At the moment we are working with Citrix Engineering and Nvidia (although it’s mainly focused on Citrix because we don’t see the Nvidia driver included in the problem). One of the other thoughts was that a Microcode Patch from Intel could help to solve the problem. This patch is not included in the current Dell Bios. The current version is 0X1F – currently the Dell Bios has the 0X1E integrated. Fujitsu has updated his Bios at the end of December with 0X1F. We installed the update manually – but it didn’t change anything. (If I have some time later I will publish a separate blog post how to do that…).

Our current workaround is to use v3 CPUs. The performance is lower – but it was the option where we didn’t have to change anything in our environment (no reinstallation or reconfiguration) except the CPU.

Since yesterday we are running another test with v4 CPUs. We disabled a new feature from the CPU that was adopted in Xen 4.6 (=> XenServer 7). Until now we have no crashes *fingers crossed*. User tests will follow tomorrow.

One last (personal) thing at the end:
I really appreciate it that Dell did many Hardware-Replacements really fast and uncomplicated. Nevertheless it never felt like the Dell engineering was involved in the problem solution. Citrix and Nvidia Engineers asked us a lot of questions – but we didn’t receive any from Dell engineering.
The point I am more worried about is how Dell “continues” to support an environment with XenServer. During the discussions we often got the information like “We don’t support that because it’s not on the Citrix HCL” – interestingly on the XenServer HCL you can find the Dell R720 (a quite old system) in conjunction with the Nvidia M60. On the other hand the R730 is listed with the M10. Furthermore Dell send a R730 to Citrix for testing it and adding it to the HCL. The “normal” way I know is that the Hardware-Vendor does all the tests for the HCL and just sends Citrix the results. We have asked our Dell representatives for an official statement – hopefully we will receive that soon and it’s positive about XenServer Support from Dell (and they don’t try to replace  XenServer with VMWare…).