Configuring GPU Passthrough in OpenStack

Configuring GPU Passthrough in OpenStack

- 12 mins

As part of my effort to “cut the cord”, I recently install Plex Media Server onto a virtual machine instance running in my home OpenStack environment. The compute node is a few generations old, and while perfectly capable of running many different workloads, transcoding Ultra HD (4K) content is not one of them.

To combat this, I recently installed an NVIDIA Quadro P2000 video card with the goal of passing it through to the virtual machine instance to improve the viewing experience.

My OpenStack cluster consists of the following:

The virtual machine instance running Plex Media Server has the following specs:

All media is stored on a NAS and accessed over NFS.

Preparation

In a recent post, I documented the process of procuring the proper riser and installing the GPU into this server. This process will vary from machine to machine, of course.

To pass a GPU through to virtual machines, you will want to enable VT-d extentions in the BIOS. This, too, will vary from platform to platform.

Next, update the grub configuration at /etc/default/grub to ensure IOMMU support is enabled. For Intel, boot the machine, and append intel_iommu=on to the end of the GRUB_CMDLINE_LINUX line in the grub configuration file:

$ sudo nano /etc/default/grub
...
GRUB_CMDLINE_LINUX="splash=quiet console=tty0 ... intel_iommu=on
...

For AMD, boot the machine, and append amd_iommu=on to the end of the GRUB_CMDLINE_LINUX line in the grub configuration file:

$ sudo nano /etc/default/grub
...
GRUB_CMDLINE_LINUX="splash=quiet console=tty0 ... amd_iommu=on
...

Refresh the grub configuration and reboot the host:

$ sudo update-grub
$ reboot

Driver Shenanigans

Once the host has been rebooted, the next step in preparing the GPU for passthrough is to ensure the proper drivers are configured. A few useful pieces of information that will be needed as we progress through the process include the following:

These can all be determined with the lspci command shown here:

$ sudo lspci -nn | grep NVIDIA
24:00.0 VGA compatible controller [0300]: NVIDIA Corporation GP106GL [Quadro P2000] [10de:1c30] (rev a1)
24:00.1 Audio device [0403]: NVIDIA Corporation GP106 High Definition Audio Controller [10de:10f1] (rev a1)

In this example, the PCI Bus ID is 24:00.0, the Vendor ID is 10de and the Product ID is 1c30.

There is a good chance that the kernel has detected the GPU and loaded a graphics driver. In the case of NVIDIA cards, this likely means the open source nouveau driver, as shown here:

$ sudo lspci -s 24:00.0 -k
24:00.0 VGA compatible controller: NVIDIA Corporation GP106GL [Quadro P2000] (rev a1)
	Subsystem: Dell GP106GL [Quadro P2000]
	Kernel driver in use: nouveau
	Kernel modules: nvidiafb, nouveau

While the video card is bound to a graphics driver, GPU passthrough will not be possible.

Blacklist

In order to ensure the graphics drivers are not loaded, we need to blacklist them. Create and edit a dedicated blacklist file at /etc/modprobe.d/blacklist-nvidia.conf and add the following contents:

blacklist nouveau
blacklist nvidiafb

Whitelist

Next, we need to ensure that the vfio-pci driver gets loaded and binds the GPU. Create and edit a dedicated file at /etc/modprobe.d/vfio.conf and add the following contents:

options vfio-pci ids=10de:1c30,10de:10f1

Then, add the vfio-pci driver to the /etc/modules-load.d/modules.conf file to ensure it’s loaded:

8021q
br_netfilter
dm_multipath
dm_snapshot
ebtables
ip6table_filter
ip6_tables
ip_tables
...
vfio-pci

Reboot the host.

$ sudo reboot

Once the host has returned, you can verify the proper driver has been bound using the following lspci command:

$ sudo lspci -s 24:00.0 -nnk
24:00.0 VGA compatible controller [0300]: NVIDIA Corporation GP106GL [Quadro P2000] [10de:1c30] (rev a1)
	Subsystem: Dell GP106GL [Quadro P2000] [1028:11b3]
	Kernel driver in use: vfio-pci
	Kernel modules: nvidiafb, nouveau

The Kernel driver in use: vfio-pci line indicates the proper driver is now in place.

Nova changes

The OpenStack Compute service needs to be configured in two places in order to recognize and utilize the GPU.

First, configure a PCI passthrough whitelist on the compute node where the GPU resides. Update the [PCI] section of the /etc/nova/nova.conf file with the following:

[PCI]
passthrough_whitelist: { "vendor_id": "10de", "product_id": "1c30" }

Ensure that the vendor_id and product_id match that retrieved earlier in the process. The whitelist can be limited to referencing a particular PCI slot if needed, but since I only have a single card installed, that isn’t really necessary. Restart the nova-compute service:

$ sudo systemctl restart nova-compute

Next, configure a PCI alias on the node hosting the Nova API service. Update the [PCI] section of the /etc/nova/nova.conf file with the following:

[PCI]
alias: { "vendor_id":"10de", "product_id":"1c30", "device_type":"type-PCI", "name":"quadro-p2000" }

Restart the Nova API service:

$ sudo systemctl restart nova-api

Lastly, ensure that the Nova scheduler has been configured with a PCI Passthrough filter. You can verify PciPassthroughFilter is listed in enabled_filters and/or available_filters in the /etc/nova/nova.conf file, as shown here, and add it if necessary:

[filter_scheduler]
enabled_filters = RetryFilter, AvailabilityZoneFilter, ComputeFilter, ComputeCapabilitiesFilter, ImagePropertiesFilter, ServerGroupAntiAffinityFilter, ServerGroupAffinityFilter, PciPassthroughFilter
available_filters = nova.scheduler.filters.all_filters

Restart the Nova scheduler service if necessary:

$ sudo systemctl restart nova-scheduler

Create a flavor

Nova uses flavor metadata and properties to determine what resources should be associated with an instance. In the case of GPU passthrough, one must configure a flavor with the pci_passthrough:alias property. In the following example, I’ve created a new flavor with 6 vCPUs, 8 GB of RAM, a 40 GB disk, and a pci_passthrough:alias property referencing the quadro-p2000 alias we configured earlier:

openstack flavor create \
--vcpus 6 \
--ram 8192 \
--disk 40 \
--property "pci_passthrough:alias"="quadro-p2000:1" \
6-8-40-gpu

In quadro-p2000:1, the quadro-p2000 directly references the alias name, while the 1 instructs Nova that a single GPU should be assigned.

Boot an instance

To test the new flavor and new GPU passthrough capabilities, spin up a virtual machine instance using the new flavor, as shown here:

openstack server create \
--flavor 6-8-40-gpu \
--image ubuntu-bionic \
--network LAN \
--key-name imac-rsa \
--security-group plex \
plex-pgu-server

After a brief moment, the server should be ACTIVE:

root@lab-infra01-utility-container-cee33e10:~# openstack server list
+--------------------------------------+-------------------+---------+-------------------+---------------+---------------+
| ID                                   | Name              | Status  | Networks          | Image         | Flavor        |
+--------------------------------------+-------------------+---------+-------------------+---------------+---------------+
| 854ba22b-52c0-44a1-ab66-b78eba142bc5 | plex-pgu-server   | ACTIVE  | LAN=192.168.2.111 | ubuntu-bionic | 6-8-40-gpu    |
+--------------------------------------+-------------------+---------+-------------------+---------------+---------------+

Login to the machine and verify the GPU is recognized:

ubuntu@plex-pgu-server:~$ lspci | grep NVIDIA
00:06.0 VGA compatible controller: NVIDIA Corporation GP106GL [Quadro P2000] (rev a1)

For best performance, updated drivers can be installed from the upstream Ubuntu repository or from NVIDIA’s website. In this case, I’ve chosen to install drivers from the upstream repo using the ubuntu-drivers utility:

ubuntu@plex-pgu-server:~$ sudo apt update
ubuntu@plex-pgu-server:~$ sudo apt install ubuntu-drivers-common

ubuntu@plex-pgu-server:~$ ubuntu-drivers devices
== /sys/devices/pci0000:00/0000:00:06.0 ==
modalias : pci:v000010DEd00001C30sv00001028sd000011B3bc03sc00i00
vendor   : NVIDIA Corporation
model    : GP106GL [Quadro P2000]
driver   : nvidia-driver-390 - distro non-free recommended
driver   : xserver-xorg-video-nouveau - distro free builtin

The utility recommends the nvidia-driver-390 driver, and even though there may be newer drivers available, this works for my purposes.

ubuntu@plex-pgu-server:~$ sudo apt install nvidia-driver-390

Once the drivers have been install, reboot the instance:

ubuntu@plex-pgu-server:~$ sudo reboot

After rebooting the instance, verify the new drivers have been installed using lspci and nvidia-smi:

ubuntu@plex-pgu-server:~$ sudo lspci -nnk -s 00:06.0
00:06.0 VGA compatible controller [0300]: NVIDIA Corporation GP106GL [Quadro P2000] [10de:1c30] (rev a1)
	Subsystem: Dell GP106GL [Quadro P2000] [1028:11b3]
	Kernel driver in use: nvidia <---
	Kernel modules: nvidiafb, nvidia_drm, nvidia

ubuntu@plex-pgu-server:~$ nvidia-smi
Wed May  8 18:08:14 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.116                Driver Version: 390.116                   |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro P2000        Off  | 00000000:00:06.0 Off |                  N/A |
|  0%   38C    P0    19W /  75W |      0MiB /  5059MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Voila! Now it’s off to the races!

Benchmarking

The best way to get a BEFORE and AFTER is to kick off the transcoder. For this, I fired up the AppleTV 4 (HD) to force Plex to transcode an 4K movie to 1080p. To do this, you’d need a Plex Pass to enable Hardware Acceleration, but there may be alternative methods of verifying.

Here’s a look at the CPU statistics before playing the movie:

transcode cpu noop

Once the movie started playing, CPU usage shot up over 85%:

transcode cpu software

A look at the Plex dashboard reflected software-based transcoding (note the lack of ‘hw’):

dashboard software

We can see the Quadro P2000 just sitting there not utilized:

Wed May  8 19:04:40 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.116                Driver Version: 390.116                   |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro P2000        Off  | 00000000:00:06.0 Off |                  N/A |
|  0%   38C    P0    19W /  75W |      0MiB /  5059MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Enabling Hardware Acceleration in the Plex server settings and restarting the movie reflected a significant drop in CPU utilization:

transcode cpu hardware

A drop from 80% to ~50%? Not shabby, but not great. A look at the Plex dashboard reflected hardware-based transcoding was in effect:

dashboard hardware

We can see the Quadro P2000 barely trying:

Wed May  8 19:11:03 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.116                Driver Version: 390.116                   |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro P2000        Off  | 00000000:00:06.0 Off |                  N/A |
| 49%   42C    P0    19W /  75W |    161MiB /  5059MiB |      1%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0     26847      C   /usr/lib/plexmediaserver/Plex Transcoder     151MiB |
+-----------------------------------------------------------------------------+

Turns out, fully-offloaded transcoding isn’t yet supported in Plex Media Server for Linux. Further research led me to some FFmpeg flags that would really make this card shine. To implement these, however, one must a) wait for a future update to Plex Media Server or b) create a wrapper script to implement the flags. I chose the latter.

The script

To offload both encoding and decoding, a wrapper script should be implemented that adds an addition flag to enable the NVDEC engine.

First, rename the Plex Transcoder to Plex Transcoder2:

ubuntu@plex-pgu-server:~$ cd /usr/lib/plexmediaserver
ubuntu@plex-pgu-server:~$ sudo mv Plex\ Transcoder Plex\ Transcoder2

Then, create a file named Plex Transcoder with the following contents:

#!/bin/bash
marap=$(cut -c 10-14 <<<"$@")
if [ $marap == "mpeg4" ]; then
     exec /usr/lib/plexmediaserver/Plex\ Transcoder2 "$@"
else
     exec /usr/lib/plexmediaserver/Plex\ Transcoder2 -hwaccel nvdec "$@"
fi

Next, set the file to be executable:

ubuntu@plex-pgu-server:~$ sudo chmod 755 /usr/lib/plexmediaserver/Plex\ Transcoder

The changes should be noticed immediately. Restarting the stream, we can see CPU utilization drop to below 20%:

transcode cpu wrapper

Now we’re cooking with gas! A quick look at nvidia-smi shows the GPU barely breaking a sweat, but performing admirably:

Wed May  8 19:25:14 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.116                Driver Version: 390.116                   |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro P2000        Off  | 00000000:00:06.0 Off |                  N/A |
| 49%   41C    P0    21W /  75W |   1034MiB /  5059MiB |     11%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0     27118      C   ...ib/plexmediaserver/Plex Transcoder Orig  1024MiB |
+-----------------------------------------------------------------------------+

Summary

GPU Passthrough and hardware offloading? Yes, please!

If you have 4K content, chances are your viewers would benefit from Direct Play functionality only possible with true 4K-capable devices. Transcoding 4K to something like 1080p will likely result in some color manipulation that astute viewers would dislike, but for kids, it’s good enough. Letting the Xeons try to transcode the media results in a lousy viewing experience, so the GPU is worth it to me. I also consider reducing the amount of vCPUs reserved for Plex in my lab as a win.

If you have some thoughts or comments on this process, I’d love to hear ‘em. Feel free to reach out on Twitter at @jimmdenton.

James Denton

James Denton

openstack, networking, and random musings

rss facebook twitter github gitlab youtube mail spotify lastfm instagram linkedin google google-plus pinterest medium vimeo stackoverflow reddit quora quora