Testing 10 GbE throughput on Windows - iperf3 is outdated

iperf3 only showing 4.5 gbps in Windows

Recently I upgraded my AMD-based PC on a livestream, and I installed an Innodisk EGPL-T101 10 Gbps M.2 NIC (link to Innodisk product page).

Under Linux, I could get through 9.4 Gbps using iperf3 between the PC and my Mac Studio. But under Windows, I could only get up to about 4.5 Gbps (tested around 1h 27m into the stream)!

I had downloaded the latest version of iperf3 for Windows from the iperf.fr website, which was listed as version 3.1.3 from June of 2016! I thought that was pretty old, and indeed, looking at all the downloads on that page... they are all very old. What gives?

Well, even though that site seems to be at the top of all search results for iperf3 downloads, it doesn't have the latest version of iperf3. On Linux and macOS, I just apt|brew|dnf install iperf3 and I have a fairly up-to-date version.

On Windows, it looks like choco (Chocolatey) pulls iperf3 binaries for Windows from files.budman.pw, and there is also a GitHub project maintaining releases of iperf3 for windows, ar51an/iperf3-win-builds.

I have downloaded and run iperf3 successfully from both sources (getting 9.40 Gbps on that PC), and it seems like both builds come out of the kind efforts of Neowin users:

So I'm not sure if one is recommended over the other, really—especially for my needs, where I'm doing local bandwidth testing between two machines and don't need SSL support or multithreading. The files.budman.pw site is linked from the obtaining.rst doc in the esnet/iperf project, FWIW.


So what throughput did you get with an updated version?

Try Cisco TREX, it's a purpose-built appliance (on ancient centos, yay), but it's used with dpdk nics like mellanox x6's to hit 200gb bidir per nic. Their equivalent of a Sprient or Ixia bandwidth tested you used to hear about opened to the world luckily knowing Cisco...

No way I’ve been wondering about the old builds on the iPerf3 website too but never looked into it. Thanks for sharing this! I wonder if it‘ll show a difference in speed on my 2.5Gbe Network

I’ve also noticed quite a few public iperf servers are defunct as of late. Still useful for testing consistency of connection rather than momentary SpeedTest ™.

It might be good to try iperf 2 as well. Of particular interest may be the latency or one way delay (OWD) measurements. I think you've some Rapsberrry Pi's with clock sync. Iperf 2 will take advantage of that. Use --trip-times to enable these metrics on a test.