[LAD] CUDA

Jens M Andreasen jens.andreasen at comhem.se
Thu Sep 4 15:52:04 UTC 2008


Some time ago I mentioned having this performance problem:

On Mon, 2008-08-25 at 20:48 +0200, Jens M Andreasen wrote:
> For 1024 bytes the transfer will take about 0.2 ms. That is 20% of the
> available time if we use a time granularity of a single millisecond.
> OTOH this appears to be mostly an intial constant, (much) more data can
> be transferred in the exact same amount of time if that is what is
> needed.

I have since then updated to kernel 2.6.24.7-rt.17 (and compiled for
this systems E1200/Core2)

This by itself doubled the throughput, so the upload would be done in
approx 100us (or better.)

By chance I then tried to run the test back from runlevel 3? This got me
down to 25us! ... which also happens to be the target I had been aiming
for.

A minimal X-setup, as in: 

$ startx /usr/bin/xterm -- :1 -depth 16

.. yields the same fine results, and so does a lightweight environment
like IceWM.

But running the (default) Gnome Desktop Environment is consistently
quadrupling latencies from main to graphics memory? I have no idea what
is going on here, but kicking out gnome seems to solve the problem.


/jens




More information about the Linux-audio-dev mailing list