A question about openssh over MPTCP

A question about openssh over MPTCP

[Gang Yan]

Dear maintainers:

Hello everyone, there exists a method in 'https://www.mptcp.dev'
which can use openssh over MPTCP. But when I followed the guide
to use ssh's tools in two subflows, such as 'scp' and 'git clone <ssh link>', the
speed is the same as the TCP, which is weird. I have checked
the there is one 'msk' and two 'ssk' with 'ss -Mtie', and both subflows
is transferring the data, but the total speed is the same as TCP. 

Do you ever checked the 'scp' before? And maybe there exists a bug here?
By the way, can you provide some ideas to find out the root readson?

Thanks a lot!!!

Cheers!
Gang

Re: A question about openssh over MPTCP

[Matthieu Baerts]

Hi Gang,

On 31/03/2025 04:21, Gang Yan wrote:
> Dear maintainers:
> 
> Hello everyone, there exists a method in 'https://www.mptcp.dev'
> which can use openssh over MPTCP.

Out of curiosity, which methods did you use?
The FAQ suggests a few alternatives:

https://www.mptcp.dev/faq.html#how-to-enable-mptcp-support-with-openssh

> But when I followed the guide
> to use ssh's tools in two subflows, such as 'scp' and 'git clone <ssh link>', the
> speed is the same as the TCP, which is weird. I have checked
> the there is one 'msk' and two 'ssk' with 'ss -Mtie', and both subflows
> is transferring the data, but the total speed is the same as TCP.
> 
> Do you ever checked the 'scp' before? And maybe there exists a bug here?
> By the way, can you provide some ideas to find out the root readson?

Performances is a tricky topic: there can be many reasons not to "fill
the pipe". The bottleneck can be due to the CPU, the IO, the
application, the receiver, the sender, the network conditions, the TCP
CC algo, the MPTCP scheduler, a bug, a path marked as backup, etc. I
recommend you to monitor these resources. Analysing packet traces can
help as well to find on which side is the bottleneck (e.g. if the client
sends Zero windows, if there are a lot of retransmissions, etc.). Also
check the MIB counters (nstat) on both sides. This can be quickly
time-consuming, but probably worth it.

I didn't check the performances of 'scp' with MPTCP, but I remember in
the past to have a slow transfer because the compression was used,
and/or many small files had to be sent.

Cheers,
Matt
-- 
Sponsored by the NGI0 Core fund.