Re: performance libssh vs. openssh

[Aris Adamantiadis <aris@xxxxxxxxxxxx>]

Hi,

I think we need fair benchmarks to study the performance of libssh. What
Andreas has studied is synchronous sftp vs openssh's sftp (which is
asynchronous imho). Synchronous sftp means that only one packet of data
can be exchanged per RTT, which is desastrous on high latency * high
bandwidth links (i.e. your local gigabit network).

We'd need to implement similar test cases using the 3 different
technologies available (libssh, libssh2, openssh) and, beyond all, not
using the hard disk (even reading in /dev/urandom has kernel penalties).
Then only we'll be able to profile the different tools on one
configuration. We'd need to run them on different configurations
(localhost, host to VM, gigabit, ADSL, high latency link) to understand
where it doesn't work so well. And only then profile and find the
effective bottlenecks.

I don't have so much time to invest in such a project, but I find it
important, otherwise some people are going to say that X is 2.3 times
faster than libssh because they have one broken test which says so on
their computer.

Aris

Mikhail Kulinich a écrit :
>  Hi,
> 
> You can start on optimization of memory allocation functions.
> This can be and mostly the root cause of such libraries.
> 
> 2009/10/20 Vic Lee <llyzs@xxxxxxx <mailto:llyzs@xxxxxxx>>
> 
>     Hi,
> 
>     Do you already have some analysis on this situation already?
> 
>     I think the performance also affects the tunneling, which is the
>     essential feature I am using. I want to help on this but I don't know
>     where I should start yet.
> 
>     Thanks,
> 
>     Vic
> 
>     On Tue, 2009-09-08 at 18:17 +0200, Andreas Schneider wrote:
>     > Hey,
>     >
>     > I've tested sftp from openssh against libssh.
>     >
>     > openssh: ~8 MB/s
>     > libssh: ~2 MB/s
>     >
>     > We have a serious problem ...
>     >
>     >
>     >
>     >       -- andreas
> 
> 
> 
> 
> 
> 
> -- 
> Mikhail
>