Can anybody tell me the maximum number of Flows we can offload when configured through TC? It installs(shows) 10K flows easily but traffic does not match all of the flows. We also noticed that the latency increases many folds and throughput decreases when the installed flows reaches just above 200. And after around 450 flows, unknown/unmatched traffic stops coming to the ARM cores.
Could you shed a light on how you configure the system (host + SmartNIC side)? Are you using SMFS or DMFS? Are you using IPSec?
I am using the default settings currently. SMFS is configured.
Following Mellanox OFED and SmartNIC user manuals you should be able to configure more flows. In general, limitations for amount of the flows come from OVS. You might check ovs-vswitchd man page for the details.
In the case if everything is configured properly and you still seeing performance degradation you may open a support ticket with Nvidia in the case if your organization has a valid support contract.
Changing the mode to dkms only increased the throughput from 6Gbps to ~13Gbps for 200+ TCP flows. We have tried to configure the DPU for performance tuning as mentioned in the user manuals but it did not have any affect on the number of flows. And, we are not using OVS. We are using TC utility to configure the flows.