Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Weaver]

iperf ?

[Alex Atkin UK]

Iperf seems kinda unpredictable for me, it often comes up short compared to browser speed tests.

[XGS_Is_On]

Add more threads and ensure you've enough CPU to handle it. Far more reliable than browser tests.

[Weaver]

Agrees with XGS.

[Alex Atkin UK]

Add more threads and ensure you've enough CPU to handle it. Far more reliable than browser tests.

I have some of the most powerful hardware on the market and increasing threads doesn't help when the problem occurs.

Its not that it universally doesn't work, its inconsistent and doesn't seem to be related to threads or CPU power.

[XGS_Is_On]

Strange. iPerf3 is what we use in work to test enterprise networks from edge to edge. Controlling both the receiver and sender gives way more visibility than an Internet speed test someone else owns.

It's also what I've used troubleshooting connections between users and ISPs. It's very simple code. It uses the same APIs everything else does to open sockets and push data down them. The two sides exchange a bit of telemetry so there is a clearer picture of link quality.

Far less to go wrong than Ookla. Ram enough connections through something and it'll certainly look consistent as any issues will be compensated by other connections ramping up to cover up deficiencies but how realistic a guide that is is debatable.

At home I run these tests from clients not the edge device to keep the test in the same datapath and IP chain as my traffic. Work software can manipulate which path it takes depending on source interface specified. I would like to think that our business isn't using inconsistent metrics with our customers.

[Alex Atkin UK]

Yeah its weird, most of the time it works great but there are just some outlier scenarios where it gives odd results that are inconsistent with the real-world performance.

Its still my primary tool for testing how the internal network is performing though, but I don't 100% rely on it - not that you should ANY synthetic test as it doesn't tell you how different protocols might react.

I actually stumbled onto SSHFS while migrating my VPS over to one that supports newer kernels with the security mitigations.  Was really surprised to see it comfortably maxed out my upstream while moving the Virtualmin backup files over.

[Weaver]

Does UDP iperf3 work for you? Without destroying civilisation, that is?

[Alex Atkin UK]

I'll have to try if/when I encounter the problem again.

[XGS_Is_On]

Good plan. I've just been doing some work and happily saturated 3 links consisting of 2 x 10G and 1 x 25G with it reliably, so not sure why it's problematic for you.

[Alex Atkin UK]

Don't get me wrong, 99% of the time it works as expected, but the fact its not 100% is concerning as in the past I've seen it claim it was doing faster than the link it was running over.  Of course that might have been a bug long since squashed, but it highlights how you should never rely on just a single tool for testing.

[Weaver]

I had a big problem with it, and with my own ignorance, when I tried UDP mode across the internet. Nearly destroyed the civilised world, as I said. So I need to find out what I should be doing before I go near that again. And I can’t see any point in using it in TCP mode, given my objections to the use of TCP in speedtesters, where imho they only measure TCP, not the actual path.

[Alex Atkin UK]

Most of the time all you want to test is TCP, as its where the majority of the traffic is and where you can identify single-thread speed issues.

I've just been trying to test between my desktop and server over 10Gbit using UDP on the LAN and it seems to become CPU bottlenecked on an i5 8600K at 4.2Gbit.  I don't understand this when TCP can do 9.4Gbit absolutely fine and surely TCP should use more CPU than UDP?

[XGS_Is_On]

TCP offload engine would potentially account for some of the difference. No UDP offload in network cards or drivers so all CPU-based.

[Alex Atkin UK]

Okay this is weird, sending from a Ryzen 9 5950X (Win11) to an Intel i5 12700F (Linux) I can push 9.57Gbit with a 1M window size.  Receiving it only does 3.4Gbit.  Increasing the windows size doesn't help, although reducing it tanks the receive speed to 4.4Gbit.

However if I run the iperf3 server on Windows, performance tanks both ways.

I think I'm going to give up on this and just say its useless for testing in UDP mode, which strengthens what I said before - we test TCP for a good reason.

This does highlight something though, wont this be a problem with QUIC given the drivers will only see it as UDP traffic?  Or is this mostly down to a bad implementation in iperf3?