While evaluating the Mellanox ConnectX-5 network cards, we’ve encountered some network bandwidth issues when using the inbox drivers.
To measure network performance, we equip two machines with ConnectX-5 cards and connect them using a QSFP28 100G copper cable. The client machine has an Intel Core i9 9900 CPU, the server has an Intel Xeon Silver 4208 CPU.
The performance obtained when running 3 instances of the iperf3 program are as follow:
-
~60Gbit/s with the version 5.13 of the Linux kernel that comes with Ubuntu 21.10
-
~80Gbit/s with the version 5.16.10 of the Linux kernel
-
~95-100Gbit/s with the drivers that come with the Mellanox OFED package
We followed the recommendations found in the performance tuning guide (https://community.mellanox.com/s/article/performance-tuning-for-mellanox-adapters), but it did not result in any improvement for both inbox drivers. The iperf program gives the same result.
We have multiple questions:
-
Is there some configuration we missed that could explain this gap of performance?
-
What exactly are the differences between the drivers found in the upstream Linux kernel and the ones in the Mellanox OFED packages?
-
What are the plans for the upstream driver ?