[dpdk-users] Packet drop observed due to increased processing time of rte_eth_tx_burst( ) after dpdk upgrade

[Kaustubh Mendki <kaustubh.mendki@gmail.com>]

Hi,

Our app is currently using dpdk 17.11.10 (CentOS 7.9 Mellanox
OFED 5.2-2.2.0.0-rhel7.9-x86_64). It is quite similar to test-pipeline
example with io_rx, workers and io_tx threads running on separate lcores.

After upgrading to either dpdk 19.11.2 or dpdk 20.11.1, we observed few
packets being dropped between workers and tx threads because tx thread is
not able to keep up with the rate at which worker threads are enqueuing
packets in the ring, and eventually ring becomes full. it occurs only for a
moment initially at the start of traffic generation (@500 Mbps), then
packet drops are not seen thereafter.

The bottleneck seems to be rte_eth_tx_burst() call in tx thread which seems
to be much slower in dpdk 19.11 and 20.11 consuming ~150 milliseconds
initially for few packets and later improving with processing time in
nanoseconds. With dpdk 17.11.10, where no drop is seen, the same call
executes in ~50 microseconds for initial few packets and later in
nanoseconds. The burst size used in transmission is 1.

Is there any change in the implementation or configuration (offloads?)
required for  rte_eth_tx_burst() in dpdk 19.11.2 which could impact the
initial performance? With only dpdk upgrade and all other platform specific
things (OS, drivers) being the same, I am not sure what could be the reason
behind the higher execution time for the burst API in dpdk 19/20?

Thanks,

Kaustubh