Ping performance improves during iPerf3 network congestion test

akastack

3/29/24, 3:23 AM

I am beginning to learn about networking and testing network performance (specifically on containers). I am trying to learn how to perform good network tests to see networking capabilities of docker containers and whether they can support deterministic networking.

So, I am learning iperf3, and using Ping to track RTT and Jitter.

My current setup is two docker containers running on the same host using a bridge network to connect the two containers. I wrote a simple program to read Ping output and graph the RTT (or latency) and then calculate the Jitter between the two.

I want to simulate latency or impact of latency during a congested network, so I thought the best way was to use iperf3 to do so.

On one docker container, I ran iperf3 -s. The other docker container I ran iperf3 -c server_ip -t 500 and ping -i 0.01 -c 1000 172.19.0.4 > output.txt I ran three Ping tests, and arrived at the following result.

Ping Test

To me, this is strange, why does PING RTT improve on a congested network? If my understanding is wrong, how do I properly simulate network congestion to understand its impact on network latency and jitter?

Thanks

1 + 0

jitter

ping

docker

perf

Score:1

Server

Ginnungagap

3/29/24, 5:03 AM

Local containers aren't a good means of simulating physical network phenomenon. You're load testing your CPU at best here.

KVM has options to simulate network parameters like packet loss, rate and maybe congestion (although I'm really not sure about that last one) but I don't think that Docker/Podman do.

In any case, if you really want to try out real network phenomenon, plug into a real (test) network.

+ 2

akastack

3/29/24, 5:28 AM

What is a "real network". If I extend my docker container network, for example, to be on two different nodes (different hosts), I think the method of developing correct tests for latency and jitter are the same.

Håkan Lindqvist

3/29/24, 8:32 AM

@akastack I think the way to read it is basically "real, as opposed to virtual". And I think the main point is that you are testing a whole different thing than you seem to derive your own expectations from. You say essentially that your results from a fully virtual environment, with no actual network hardware or network links involved, only code running on the CPU, does not match your expectations of how a regular physical network should behave. Depending on what you want to test, your "container network across two nodes" may be relevant, but it also may involve more pieces than necessary.

Elon Musk

I sit in a Tesla and translated this thread with Ai:

EN: Ping performance improves during iPerf3 network congestion test

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.