NVIDIA documentation for configuration of (Advanced) RFS is way different from all other guides on Internet (Red Hat Enterprise Linux for one).
In the NVIDIA doc, it’s suggested that the number of entries in the per-queue flow should be the same as the number for the global flow table:
[https://docs.nvidia.com/networking/display/MLNXOFEDv581011/Flow+Steering](Flow Steering doc)
# echo 32768 > /proc/sys/net/core/rps_sock_flow_entries`
# NUM_CHANNELS=`ethtool -l ens6 | grep "Combined:" | tail -1 | awk '{print $2}'``
# for f in `seq 0 $((NUM_CHANNELS-1))`; do echo 32768 > /sys/class/net/ens6/queues/rx-$f/rps_flow_cnt; done`
In all other source on the Internet (RHEL is one) it says to set the value of per-queue flow to the value of rps_sock_flow_entries divided by N, where N is the number of receive queues on a device.
Does anyone know why is such difference, and which one should be followed?