MPRQ issues under high load with mlx5 PMD

[Wicht, Baptiste]

Hi



We have noticed a strange issue with MPRQ on the mlx5 PMD: When reaching rates close to line-rate (100G), we are seeing error conditions where a lot of the packets are dropped. When this happens, we also see errors in system logs (CQ error with syndrome 0x1). We have noticed that on both ConnectX-6 and ConnectX-7. When reaching this condition, we can go from dropping close to zero packets to dropping more than 50% of packets, likely because completion queues are stuck.



We have noticed this with multiple versions of DPDK (including 25.11) and with and without OFED drivers. We have also tested many configuration of MPRQ (log stride size and num).



Our suspicion is that this is caused when there is a lot of contention (16+ receivers) and when packets start dropping.



Does anybody have an idea on how to solve this issue? Currently, the only workaround we have found is to disable MPRQ.



Best regards

Baptiste Wicht

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mails.dpdk.org/archives/users/attachments/20260423/ab3543a9/attachment.htm>