Document instructions for how to do performance simulations with P4-programmable simulated network devices

[jafingerhut]

Using only BMv2 simple_switch_grpc and Mininet, e.g. as provided in the https://github.com/p4lang/tutorials repository, there are significant issues where when simulated links are congested, the queueing behavior in terms of latency and throughput is significantly different than what occurs with physical switches and constant bit rate links between them. See [1] [2] for examples of the kinds of issues encountered.

In research paper [3], the authors say they have implemented software that enables one to write P4 programs and simulate them on a system where performance-level details similar to real physical switches are modeled correctly. If true, it would be good to have explicit instructions on how to run these, perhaps documented in the tutorials repository [3] for one or more exercises that rely on queue depths for their behavior, e.g. especially the ecn [4] and link_monitor exercises [5].

[1] https://forum.p4.org/t/unexpected-queue-depth-spikes-in-bmv2-v1model-during-congestion-control/1391
[2] https://forum.p4.org/t/when-i-conducted-an-attack-experiment-using-bmv2-the-tcp-congestion-control-seemed-to-fail/1373
[3] https://github.com/p4lang/project-ideas/issues
[4] https://github.com/p4lang/tutorials/tree/master/exercises/ecn
[5] https://github.com/p4lang/tutorials/tree/master/exercises/link_monitor

[Vineet1101]

@jafingerhut where are the author instructions that gave result similar to real physical switches???

[jafingerhut]

Mingyu Ma gave a P4 Developer Days talk on p4sim, some software he developed that is supposed to integrate BMv2 with the ns3 network performance simulation system. I have not used it before myself.

Here is a link to the recording of his presentation: https://p4.org/event/p4-developer-days-p4sim-protocol-independent-packet-processors-in-ns-3/

I would recommend contacting Mingyu directly if you have further questions.

[Vineet1101]

This is easy as this was my GSOC project. So I have already setup instructions and running the program using P4Sim.
https://github.com/Vineet1101/P4Simulator/blob/main/doc/vm-env.md here are the instructions to install P4Sim and
here are some examples to run p4 programs in ns-3 https://github.com/Vineet1101/P4Simulator/tree/main/examples

[fruffy]

@Vineet1101 Great!

@jafingerhut So these should be added to tutorials?

[Vineet1101]

@jafingerhut can you tell where to add these so that I can start working on it

[Dscano]

Can we close this Issue @jafingerhut ?

• added docs for installing p4sim tutorials#706
• doc: rewrite and restructure doc/P4sim/README.md tutorials#722

[jafingerhut]

Can we close this Issue @jafingerhut ?

• added docs for installing p4sim tutorials#706
• doc: rewrite and restructure doc/P4sim/README.md tutorials#722

Those PRs do explain how to install P4sim and to run one of the examples, but it is not clear to me whether there is an existing example that demonstrates something that is more performance-simulation-oriented than bmv2. Examples of things like that, that would convince me we've completed this issue, would be:

• A simple P4 example showing hardware-like performance results with a simulated network where there is a link configured with a known small speed (e.g. 100 Mbits/sec) between two simulated network devices, and the queueing and packet scheduling behavior on both ends of that link are similar to hardware behavior, and NOT like current bmv2 behavior.

[Mingyumaz]

I have modified some P4sim code and also tried to clear the queue/performance/behavior issue in that PR-726. Maybe you could have a look at that?

[jafingerhut]

I have modified some P4sim code and also tried to clear the queue/performance/behavior issue in that PR-726. Maybe you could have a look at that?

Definitely, and thanks for the additional PR. It might be a few days before I get to it, though.