Determining the GPU Topology

Configuration Scheme 1

1. Query GPU rendering nodes.

   1	ll /dev/dri/by-path/ | grep renderD

   Example command output:

   1 2	lrwxrwxrwx 1 root root 13 Oct 25 10:58 pci-0000:03:00.0-render -> ../renderD128 lrwxrwxrwx 1 root root 13 Oct 25 10:58 pci-0000:83:00.0-render -> ../renderD129

   This indicates that two AMD GPUs are inserted into the server, and the rendering nodes are renderD128 and renderD129.

2. Query the NUMA node to which a GPU rendering node belongs.

   1	cat /sys/bus/pci/devices/0000\:XX\:00.0/numa_node

   Replace XX in the command with the IP address of a node queried in step 1. Take renderD128 as an example. The query command is as follows:

   1	cat /sys/bus/pci/devices/0000\:03\:00.0/numa_node

   Command output:

   1

   0

   This indicates that renderD128 belongs to NUMA node 0.

Configuration Scheme 2

1

lspci -vvv -d :0200 | grep NUMA

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 0
NUMA node: 2
NUMA node: 2
NUMA node: 2
NUMA node: 2
NUMA node: 2
NUMA node: 2
NUMA node: 2
NUMA node: 2

The command output shows that, in the /dev/dri/ directory, rendering nodes renderD128 to renderD135 belong to NUMA0 and renderD136 to renderD143 belong to NUMA2.