What are the performance limits of UPG-VPP?

[Rorsachach]

We compiled UPG in the physical server and did not use Docker. In the process of trying to measure the speed, we used a simulator to simulate 10Gbps uplink traffic and sent it through a UE. What we see is that the final reception rate at the receiver is only 2Gbps. Through show run, we found that the vector was much larger than 100, and a large number of RX misses appeared. It is worth mentioning that using non-GTPU packaged UDP packets can easily run up to 10Gbps.

I would like some help with performance tuning. I also have a question about why some nodes choose to handle a single vector. Could these parts of the code be modified to make it more efficient? For example, the upf_input function and the upf_forward function.

[RoadRunnr]

We compiled UPG in the physical server and did not use Docker. In the process of trying to measure the speed, we used a simulator to simulate 10Gbps uplink traffic and sent it through a UE. What we see is that the final reception rate at the receiver is only 2Gbps. Through show run, we found that the vector was much larger than 100, and a large number of RX misses appeared. It is worth mentioning that using only non-GTPU packaged UDP packets can easily run up to 10Gbps.

Performance heavily depend on the network drivers that you use. The af_packet and af_xdp drivers suck in terms of performance. DPDK drivers might need careful tuning with their queue settings.

Also, the layout of your PDR has a big impact on performance. PDRs and ACLs are both scanned linearly. There is no optimization on that.

Which ones did you use? What PDRs/ACLs did you use?

I would like some help with performance tuning. I also have a question about how some nodes choose to use a single vector. Could this part of the code be modified to make it more efficient? For example, the upf_input function and the upf_forward function.

Sure, patches (and sensible suggestions) are welcome.