[darcs-devel] performance measurements of "darcs get", "hg clone", and "rsync -r"

[Alexander Staubo]

On 1/25/08, zooko <zooko at zooko.com> wrote:
> > Are you using Darcs' control master feature?
>
> I measured the default behavior of "darcs get --darcs-2", "darcs get
> --hashed" and "darcs get", with the matching formats of source
> (remote) repositories.

The control master feature was turned off after reports that it was
buggy on MacOS X and possibly Linux. As far as I know, the bug does
manifest itself on Linux (and Leopard apparently fixes what looks a
lot like a kernel bug in Tiger).

You really should run with the control master turned on -- it caches
SSH connections and thus eliminates all the connection setup/teardown
overhead, which is fairly expensive. To do this, add --ssh-cm to the
command line.

> > You should be able to see all the SSH commands
> > issued by Darcs by adding "-v -v -v".
>
> Actually, I did have "--verbose --verbose --verbose" on the command-
> line, but nothing about ssh command was emitted, and in fact only 3
> lines were emitted:

I see. I know Eric Kow added -v counting and SSH logging, but his
patches might not have been included on the branch from which David
builds darcs-2.

On Linux you should be able to trace SSH invocations with something like:

  strace -p PID 2>&1 | grep "exec.*(ssh|scp|sftp)"

Alexander.