Linux ethtool Statistics - Linux ethtool Statistics - UG1739

port_rx_nodesc_drops

Number of packets dropped by the network adapter because of a lack of RX descriptors in the RX queue.

Packets can be dropped by the NIC when there are insufficient RX descriptors in the RX queue to allocate to the packet. This problem occurs if the receive rate is very high and the network adapter receive cycle process has insufficient time between processing to refill the queue with new descriptors.

A number of different steps can be tried to resolve this issue:

• Disable the irqbalance daemon in the OS
• Distribute the traffic load across the available CPU/cores by setting rss_cpus=cores. Refer to Receive Side Scaling section
• Increase receive queue size using ethtool.

• The matched filter indicates the packet must be dropped.
• The packet does not match any filter.
  Note: A copy of each received packet is always sent to the Enterprise datapath, to allow sniffing. If a packet is directed to the Express datapath, and the copy that is sent to the Enterprise datapath does not match any Enterprise datapath filter, this counter increments.

Number of packets dropped because there was not enough server packet memory available to receive them. Happens when packets cannot be delivered to the host as quickly as they arrive due to:

• Adapter is inserted into a low speed or low width PCI slot – so the PCIe bus cannot support the required bandwidth.
• Host memory bandwidth is being used by other devices resulting in poor performance for the adapter.
• All packets are very small (for example 64 bytes) and both interfaces run at 100G. (PCIe overheads are highest for small frames, and this combination can oversubscribe a PCIe Gen5 ×8 link.)

Number of bytes dropped because there was not enough server packet memory available to receive them. Happens when packets cannot be delivered to the host as quickly as they arrive due to:

• Adapter is inserted into a low speed or low width PCI slot – so the PCIe bus cannot support the required bandwidth.
• Host memory bandwidth is being used by other devices resulting in poor performance for the adapter.
• All packets are very small (for example 64 bytes) and both interfaces run at 100G. (PCIe overheads are highest for small frames, and this combination can oversubscribe a PCIe Gen5 ×8 link.)

Number of packets dropped by the network adapter because of a lack of superbuffers for an RX queue.

This problem occurs if the receive rate is very high and the host software does not refill the RX ring in time.

Number of bytes dropped by the network adapter because of a lack of superbuffers for an RX queue.

This problem occurs if the receive rate is very high and the host software does not refill the RX ring in time.

• Mismatch unicast address and unicast promiscuous mode is not enabled.
• Packet is a pause frame.
• Packet has length discrepancy.
• Due to internal FIFO overflow condition.
• Length < 60 bytes.

These PTP stats counters relate to the mechanism used by sfptpd to synchronize the system clock and adapter clock(s) in a server.

For each synchronization event, sfptpd selects a number of system clock times to be compared to the adapter clock time. If the times can be synchronized, the good_syncs counter is incremented, otherwise the bad_syncs counter is incremented. If sfptpd is unable to synchronize the clocks at this event, the sync_timeout counter is incremented.

sfptpd synchronizes clocks 16 times per second, so incrementing counters do not necessarily indicate bad synchronization between local server clocks and an external PTP master clock.