Contents
This roll enables GPU pass-through to a guest VM. The roll installs rocks commands for GPU management a frontend and vm-container nodes.
This roll assumes that GPU cards are NVIDIA cards that have Video/Audio function for each card which are designated by a function 0/1 respectively on the PCI bus. For example, to list nvidia devices
[root@gpu-1-6]# lspci -D -d 10de: 0000:02:00.0 3D controller: NVIDIA Corporation GF100GL [Tesla T20 Processor] (rev a3) 0000:02:00.1 Audio device: NVIDIA Corporation GF100 High Definition Audio Controller (rev a1)
This roll assumes that GPU cards for each physical host are given logical names with the prefix
gpupciThe names are unique on a single host and the sequence starts withgpupci1. The name is not unique across different hosts.A single GPU card will be assigned to a VM host. The logical name for a GPU card on any VM is always
gpupciwithout indexes. See example below for usage.Cuda roll is installed on vm-containers and on the guest VMs.
To build the roll, execute :
# make roll
A successful build will create gpupt-*.x86_64*.iso file.
To add this roll to existing cluster, execute these instructions on a Rocks frontend:
# rocks add roll gpupt-*.x86_64.disk1.iso # rocks enable roll gpupt # (cd /export/rocks/install; rocks create distro) # rocks run roll gpupt > add-roll.sh # bash add-roll.sh
On the vm-container nodes (only GPU-enabled):
# yum clean all # yum install rocks-command-gpupt
The following commands are enabled with the gpupt roll:
rocks add host gpu ... rocks dump host gpu ... rocks list host gpu ... rocks remove host gpu ... rocks report host gpu ... rocks set host gpu ...
A plugin
plugin_device.pyto manage guest VM GPU pass-through PCI addressing. Used by rocks commandrocks report host vm config.A command
gpupcito manage GPU cards PCI addressing (list, detach, attach). This command is executed on the GPU-enabled hosts to get information (list) or to make GPU card available/non-available on the physical host PCI bus. For more info usegpupci -h
The Intel VT-d extensions provide hardware support for assigning a physical device to a guest VM. There are 2 parts in enabling extensions (assuming that the hardware provides a support for it). The changes are done on the physical host that has GPU cards and will be hosting VMs.
Enable VT-D extensions in BIOS Verify if your processor supports VT-d extensions. The extensions differ among manufactureres. Consult the BIOS settings.
Activate Vt-d in the kernel Append the following flags to the end of the
kernelline in boot.grub:intel_iommu=on iommu=pt pci=realloc rdblacklist=nvidia
The last flag is to disable loading of nvidia driver.
Uninstall nvidia driver. This step is important otherwise when booting VMs later the following errors may be present in /var/log/libvirt/qemu/VMNAME.log and VM will not boot:
Failed to assign device "hostdev0" : Device or resource busy 2017-08-31T22:17:28.117713Z qemu-kvm: -device pci-assign,host=02:00.0,id=hostdev0, ... Device 'pci-assign' could not be initialized
To uninstall the driver
/opt/cuda/driver/uninstall-driver more /var/log/nvidia-uninstall.log
Reboot.
When the host is rebooted, check if the changes are enabled:
# cat /proc/cmdline ro root=UUID=575b0aac-0b20-4024-8a2d-26f8d3cc460b rd_NO_LUKS rd_NO_LVM LANG=en_US.UTF-8 rd_NO_MD SYSFONT=latarcyrheb-sun16 KEYBOARDTYPE=pc KEYTABLE=us rd_NO_DM rhgb quiet intel_iommu=on iommu=pt pci=realloc rdblacklist=nvidia
the output should contain added flags
The following two commands shoudl show PCI-DMA and IOMMU
# dmesg | grep -i PCI-DMA PCI-DMA: Intel(R) Virtualization Technology for Directed I/O # grep -i IOMMU /var/log/messages Aug 28 15:06:23 gpu-1-6 kernel: Command line: ro root=UUID=575b0aac-0b20-4024-8a2d-26f8d3cc460b rd_NO_LUKS rd_NO_LVM LANG=en_US.UTF-8 rd_NO_MD SYSFONT=latarcyrheb-sun16 KEYBOARDTYPE=pc KEYTABLE=us rd_NO_DM rhgb quiet intel_iommu=on iommu=pt pci=realloc rdblacklist=nvidia Aug 28 15:06:23 gpu-1-6 kernel: Kernel command line: ro root=UUID=575b0aac-0b20-4024-8a2d-26f8d3cc460b rd_NO_LUKS rd_NO_LVM LANG=en_US.UTF-8 rd_NO_MD SYSFONT=latarcyrheb-sun16 KEYBOARDTYPE=pc KEYTABLE=us rd_NO_DM rhgb quiet intel_iommu=on iommu=pt pci=realloc rdblacklist=nvidia Aug 28 15:06:23 gpu-1-6 kernel: Intel-IOMMU: enabled Aug 28 15:06:23 gpu-1-6 kernel: dmar: IOMMU 0: reg_base_addr fbffe000 ver 1:0 cap c90780106f0462 ecap f020fe Aug 28 15:06:23 gpu-1-6 kernel: IOMMU 0xfbffe000: using Queued invalidation Aug 28 15:06:23 gpu-1-6 kernel: IOMMU: hardware identity mapping for device 0000:00:00.0 ... Aug 31 10:57:53 gpu-1-6 kernel: IOMMU: hardware identity mapping for device 0000:04:00.1 Aug 31 10:57:53 gpu-1-6 kernel: IOMMU: Setting RMRR: Aug 31 10:57:53 gpu-1-6 kernel: IOMMU: Prepare 0-16MiB unity mapping for LPC
Check that nvidia driver is not loaded
lsmod | grep nvidia
should return nothing
The commands to detach GPU cards from physical hosts are run once for each GPU card on each host. The list below includes some informational commands.
Run
gpupci -lcommand on all GPU-enabled vm-containers to get information about the GPU cards. For example, on vm-container-0-15 the output is# gpupci -l gpupci1 pci_0000_02_00_0 gpupci2 pci_0000_03_00_0
The output means there are 2 GPU cards and for each there is a logincal GPU name and its PCI bus info.
Run commands to add this information in the rocks database:
# rocks add host gpu vm-container-0-15 gpupci1 pci_0000_02_00_0 # rocks add host gpu vm-container-0-15 gpupci2 pci_0000_03_00_0
Verify that GPU info now is in the database:
# rocks list host gpu HOST GPU PCI_BUS vm-container-0-15: gpupci1 pci_0000_02_00_0 vm-container-0-15: gpupci2 pci_0000_03_00_0
Detach the GPU cards from the physical host. This is an actual command that detaches the GPU from the physical host PCI bus. This needs to be done once for each GPU card before any VM can use the GPU PCI in pass-through mode. This can be done as a single command for all cards
# rocks run host vm-container-0-15 "gpupci -d all"
or using a specific logical name for a single GPU card on a given host
# rocks run host vm-container-0-2 "gpupci -d gpupci1"
Once the GPU card is detached from a physical host it is ready for use by the guest VM. We assume that a single GPU card is assigned to a VM and that a VM is run on a GPU-enabled vm-container. For example, if there is a VM rocks-33 that is created and running on a vm-container-0-15 and we want to assign a GPU to it:
rocks stop host VM rocks-33 rocks add host gpu rocks-33 gpupci pci_0000_02_00_0 rocks report host vm config rocks-33
The first command stops VM, the add command adds a GPU attribute to the VM in the rocks database.
The report command verifies that the xml file that describes the VM configuration has device information
for the GPU card. For this example, the output would contain:
...
<hostdev mode='subsystem' type='pci' managed='yes'>
<source>
<address domain='0x0000' bus='0x02' slot='0x00' function='0x0'/>
</source>
</hostdev>
</devices>
At the next start of the VM the GPU card will be available to the VM.
PCI bus address
On the VM the GPU PCI bus address will be different from the GPU PCI address of the physical host. For eample, a GPU card on a physical host
[root@gpu-1-6]# lspci -D -d 10de: 0000:02:00.0 3D controller: NVIDIA Corporation GF100GL [Tesla T20 Processor] (rev a3)
shows on a VM as
root@rocce-vm3 ~]# lspci -d 10de: 00:06.0 3D controller: NVIDIA Corporation GF100GL [Tesla T20 Processor] (rev a3)
check nvidia driver is loaded
# lsmod | grep nvidia nvidia_uvm 63294 0 nvidia 8368623 1 nvidia_uvm i2c_core 29964 2 nvidia,i2c_piix4
check if the GPU card is present
# nvidia-smi Thu Aug 31 17:37:32 2017 +------------------------------------------------------+ | NVIDIA-SMI 346.59 Driver Version: 346.59 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M2050 On | 0000:00:06.0 Off | 0 | | N/A N/A P1 N/A / N/A | 6MiB / 2687MiB | 0% E. Process | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+
run a few commands form nvidia toolkit to get more info about the GPU card
nvidia-smi -q /opt/cuda/bin/deviceQuery /opt/cuda/bin/deviceQueryDrv
The first set of commands can be run on physical and virtual hsots, the rest are run on aphysical host.
listing of pci devices
lspci -D -n lspci -D -n -d 10de: lspci -D -nn -d 10de: lspci -vvv -s 0000:03:00.0
For example, the output below shows info for 2 GPU cards, for video and audio components
# lspci -D -n -d 10de: 0000:02:00.0 0302: 10de:06de (rev a3) 0000:02:00.1 0403: 10de:0be5 (rev a1) 0000:03:00.0 0302: 10de:06de (rev a3) 0000:03:00.1 0403: 10de:0be5 (rev a1)
The video card component ends on
0abd audio card component ends on1.virsh info for the devices as a tree
virsh nodedev-list --tree
Note, that 4 devices from the above lspci command in the output of this command become
+- pci_0000_00_03_0 (comment: parent pci device) | | | +- pci_0000_02_00_0 | +- pci_0000_02_00_1 | +- pci_0000_00_07_0 (comment: parent pci device) | | | +- pci_0000_03_00_0 | +- pci_0000_03_00_1
This syntax for pci bus is used in all
virshcommands below.virsh detach and reattach devices
virsh nodedev-detach pci_0000_02_00_0 virsh nodedev-detach pci_0000_02_00_1 virsh nodedev-reattach pci_0000_02_00_1
GPU cards info
virsh nodedev-dumpxml pci_0000_02_00_0 > pci-gpu1 virsh nodedev-dumpxml pci_0000_03_00_0 > pci-gpu2
check device symbolic links
readlink /sys/bus/pci/devices/0000\:02\:00.0/driver
check xml definition of the VM
virsh dumpxml rocce-vm3 > vm3.out
For a GPU-enabled VM,
hostdevsection described in the sections above should be in the output.
Useful links for enabling PCI passthrough devices
- Enabling PCI passthrough with KVM
- Determine if your processor supports Intel Virtualization Technology
- Red HAt Guest VM device configuration