Windows without parallel (-P 32) isn't better than linux with same config ?

[KevinDS-debug]

Hi,

I recently tested mutliple iperf3 (3.18) on my infrastructure but I notice strange througput results between Windows Server 2022 and Debian 12. Exactly, when I try the same command between Windows Server 2022 and Debian 12 I didn't have the same results on 10Gb Network Card on VMware ESXi Server.

My goal is to determine if I have misconfigurations on my Windows/Debian template, and if I need to optimize the network config beside OS.

Config:
Server => iperf3 -s --bind
Client => iperf3 -V --bind ip_interface -c ip_destination
Client MultiStream => iperf3 -V --bind ip_interface -c ip_destination -P 32

Windows Server without multistream: 1.12Gb
Debian without multistream: 9.1Gb

Windows Server with multistream: 8.96Gb
Debian with multistream: 9.28Gb

It's the big difference with and without multistream surprise me. Do you have any idea ?

[davidBar-On]

Hi, each stream in iperf3 is running in a separate thread. If the multi-streams threads are running on several CPUs/Cores it may explain the results.

What seems strange is why the throughput of single stream on the Windows Server is only 1.12Gb. Usually such limit is because of CPU limit, but 10Gb is usually an easy task for today's CPUs. It may be that somehow the TCP/System buffers (per CPU) are limited, somehow each CPU can use limited bandwidth of the interface, etc.

To better understand the Windows Server behavior you may try:

1. Try to run tests with only few streams, e.g. 2,3,4. If there is a limit per core/CPU than throughput will be about 1.12Gb times the number of streams.
2. Check whether each stream's is using a different CPU/Core and how much CPU each thread is taking.
3. Run Wireshark to check whether the 1.12Gb limit is caused by a lot of TCP retries.

[KevinDS-debug]

Thank you for your response.
My virtual machine have 4 vCPU and 16GB RAM configured, and it run on HP DL380 Gen10 server.

1. When you say I need to try with few stream 2,3,4 it's with the "-P" argument or it's with another method ?

For your second question, how can I check if iperf use different CPU/Core ? With Windows Tasks manager ?

I will try everything tomorrow

[davidBar-On]

When you say I need to try with few stream 2,3,4 it's with the "-P" argument or it's with another method ?

Yes, -P 2, etc. The idea to check whether for some reason the throughput from each CPU is limited, although if there are only 4 CPUs, the 1.12x4 is only about half of 8.96... Still, I believe it is worth checking this.

how can I check if iperf use different CPU/Core ? With Windows Tasks manager ?

On Linux I am using top -H. It seems that Task manager gives the %CPU usage per thread and the %CPU usage per core, but I am not sure it shows threads allocation per CPU. Searching the internet I found some possible alternatives, but I don't know if or how well they work: Process Explorer, Pworshell Script / NTop.