Hello,
I am working on trying to minimize drops of UDP/IPv4 data received at high data rates on a ConnectX-3 NIC. Currently, I see quite a few drops reported under vport_rx_dropped via ethtool -S <interface>. Based on the following post:
Counters Troubleshooting for Linux Driver | Mellanox Interconnect Community
it appears that this counter indicates that no software buffers were available to receive the data. However, I have been unable to find any way to modify the number of buffers available. Am I correct in assuming that this counter is incremented when netdev_alloc_skb_ip_align() returns NULL in the driver? I know that for non-NAPI devices, the kernel parameter netdev_max_backlog could be set to increase the number of available skbuffs. However, everything that I have seen indicates that the NAPI buffer size is set by the driver and presumably cannot be otherwise modified. Can someone point me to where in the driver or kernel the size/number of these buffers is set?
I believe that these packets are dropped during bursty transfers as opposed to being systemically overloaded, so I expect that increasing the number of buffers would help to alleviate the drops.
Regards,
Thomas