Re: [libdc] Stress test

[David M. <dcm@MIT.EDU>]

On Thu, 2007-12-20 at 21:33 +0900, Damien Douxchamps wrote:
> - if there's a link between two interfaces, then the cameras will appear
> twice (since they can be seen by two cards). There used to be a check of
> the camera ID before we append it to the camera list in order to avoid
> this, but this was removed in David's "big patch". Shall we reinstate
> the check? (in enumerate.c, around line 230) Or maybe there was problem
> with it?

Yes, we should probably reinstate the check.

We should also have a mechanism for selecting which host controller to
use in the case that a camera is available on more than one.  I'll look
into how Mac OS X and Juju handle this case and propose a new function
or two for choosing a host per camera.  I think it can be done in a
backwards-compatible way.

> - the time required to boot coriander with 16 cameras is 7 seconds, or
> about 0.5 seconds per camera. This seems a lot. Does anyone observe a
> similar delay? On other platforms? I haven't tested multi-camera setups
> in a long time so this may be normal.

In my experience it depends mostly on the camera.  I have an old Sony
camera that takes about 7 seconds just for that one camera.

The usual offender is the dc1394_feature_get_all() function, which reads
a huge number of registers.  There are some easy optimizations that can
be done in there.  At the moment, it loops through every possible
feature and reads 3 different presence registers for each.  It could be
optimized by reading the presence registers only once up front, and then
looping through the features.  Doing so would require some re-factoring
of code though.

-David