Open osevan opened 3 years ago
Another quesrion:
How we can manage two nics with openonload,when we want working together? ,maybe 4 NICs too
We do not publish benchmarks, sorry. The best performance results will be obtained with a Xilinx/Solarflare NIC, since the AF_XDP compatibility layer used for other NICs is quite costly.
Support for multiple NICs in Onload is transparent and automatic - it will just use the current Linux configuration, including routes, bonding, VLANs, etc.
@osevan please see below:
I can answer this question:
For memcached on Intel 82599 with Onload-on-AF_XDP vs memcached on VMA on Mellanox Connect X 4 Lx, the numbers are:
for 64 byte payload memcached on VMA is 795K TPS for 64 byte payload memcached on Onload-on-AF_XDP on Intel 82599 is 744K TPS
So performance difference is just 5%. XDP is that good!
Another quesrion:
How we can manage two nics with openonload,when we want working together? ,maybe 4 NICs too
You cant do NiC bonding, NiC Teaming with Onload-on-AF_XDP.
You cant do NiC bonding, NiC Teaming with Onload-on-AF_XDP.
It is not true. It works, and it is tested.
You add the NICs (not the team/bond interface, but the underlying NICs!) to /sys/module/sfc_resource/afxdp/register
. You configure team/bond. (These 2 steps can go in any order, I believe.) And Onload-on-AF_XDP accelerated traffic over team/bond.
Or you can use 2 NICs separately, without team/bond. There is a limitation for the NIC number. IIRC you can't use more than 8 NICs without modifying some constants in the source code. 4 NICs is definitely supported by default.
oh ok, this is news to me @ol-alexandra . I will retest it and let you know. I think I tried adding the two NiCs independently and hoped the bond will pick up.
From my experience of Onload-on-ef_vi you dont have to do anything and Onload automagically picks it up.
Someone got great results on intel NICs with openonload?
For example vpp needs only 2 CPU cores to reach the network cards limit.
Any Benchmarks available for openonload how efficient they works
Thanks and
Best regards