Performance of BMv2 using stress_test_ipv4.py

[CaffeineXavier]

Hi everyone,

I follow the instrcution in performance.md, and then i run the BMv2 correctly.

I found something interesting: The Throughput using iperf comes to 5Gbps.

I would like to know why.

And after setting the TCLink to 1000 in stress_test_ipv4.py, the avg throughput comes to 955 Mbps.

Different people may have different settings, but i still wonder that if it is reasonable, because i want to use BMv2 for testing link load balancing.

Since the topic "performance of bmv2" has been a subject of ongoing debate, I'm sharing some initial details for reference.

"Another question: is there a way to quantify how the number of lines or specific operations in P4 code affect forwarding performance? Which features have the most significant impact? If anyone has experience with this, I’d love to hear your insights.

I followed the benchmarking steps from earlier issues by turning off logs and recompiling BMv2 with -O3. For reference, here are my hardware specs, which likely influenced the results: A physical machine equipped with AMD Ryzen 9950X, 32G DDR5 6000, Mellanox ConnectX-5 100GbE NIC, and a Realtek 2.5GbE NIC.

[jafingerhut]

If you want to do accurate performance modeling that reflects how real physical switches behave, then I suspect that BMv2 by itself will not suit your needs, e.g. if you need queues to build up and drain at rates that match specific link rates, with buffering and delay behavior that matches real hardware, BMv2 plus veth virtual links between multiple instances of those tend to have quite different behavior than physical links with constant bit rates.

If that is your goal, then I'd recommend investigating something like P4sim to see whether it might better meet your needs. I have not tested it myself, so cannot comment from personal experience that it would do what you want. https://github.com/p4lang/tutorials/pull/706/files

You say that you compiled BMv2 with particular options and got 5 Gbps of throughput, and you would like to know why. I am guessing you wish it were higher. The short answer is that BMv2 was not designed to be a speed demon on packet processing performance. It was designed for debugging and developing P4 programs on inexpensive widely available computers.