EJFAT UDP General Performance Considerations

NIC queues on multi-cpu nodes

Scaling in the Linux Networking Stack

Contemporary NICs support multiple receive and transmit descriptor queues. On reception a NIC distributes packets by applying a filter to each that assigns it to one of a number of logical flows. Packets for each flow are steered to a separate receive queue, which in turn can be processed by a separate CPU. The goal of this is to increase performance.

The filter used is typically a hash function over the network and/or transport layer headers. Typically and for ejfat nodes this is a 4-tuple hash over IP addresses and ports of a packet. The most common implementation uses an indirection table (256 entries for ejfat nodes) where each entry stores a queue number. The receive queue for a packet is determined by masking out the low order seven bits of the computed hash for the packet (usually a Toeplitz hash), taking this number as a key into the indirection table and reading the corresponding value.

Some advanced NICs allow steering packets to queues based on programmable filters.

To find out which filter is being used:

// See how packets are distributed to Qs
ethtool -x enp193s0f1np1

// See the details of the hash algorithm
ethtool -n enp193s0f1np1 rx-flow-hash udp4

Contemporary NICs support multiple receive and transmit descriptor queues. On reception, a NIC can send different packets to different queues to distribute processing among CPUs. Find out how many NIC queues there are on your node by looking at the combined property:

// See how many queues there are 
sudo ethtool -l enp193s0f1np1

Effect of NIC queues on UDP transmission

In the case of ejfat nodes, there are a max of 63 queues even though there are 128 cores. It seems odd to me that there isn't 1 queue per cpu, and it does not appear to be changeable so most likely it's built into the kernel when first created.

Jumbo Frames