[Linux-aus] strange ATI video issue

[James Cameron]

On Fri, Feb 18, 2022 at 09:00:22PM +1100, Russell Coker wrote:
> On Friday, 18 February 2022 18:12:55 AEDT James Cameron via linux-aus wrote:
> > At a telescope we had a performance problem with a four headed
> > workstation, which uses XDMCP to host an X11 display from an
> > instrument computer, then uses x11vnc to provide that to multiple
> > remote observers.  It was limited CPU vs GPU bandwidth.  Once we had
> > enough display updates happening, the TCP queues for the X server
> > would begin to grow, and interactive latency for the VNC users would
> > climb from 250ms to about 10 seconds.  I reproduced it with many
> > instances of `xclock -update 1`.
> 
> So you are saying that the GPU isn't fast enough to keep up with
> requests and has some sort of buffer overflow and congestion control
> problem?

It wasn't possible to tell if the GPU was slow, the kernel was slow
to pass buffers, or the buffer management was faulty.  But slow GPU
was the most likely.  We knew the hardware was capable of more, before
we upgraded the operating system.  So it looked like a regression.

> So the solution would be to have the system do less graphics stuff
> or get a faster GPU?

We suspect we got a faster GPU without changing it, by using a driver
that engaged a proprietary thermal management algorithm.  Presumably
it ran various clocks at a higher rate and knew how to shift those
clocks as needed.  That's often a special sauce that is slow to be
reverse engineered.

We could have verified that with our thermal cameras, by measuring
temperature distribution across the card using the same processing
load but with different drivers.

> I'm not going to use a proprietary driver.

Indeed.  I feel the same way.  Yet I use proprietary silicon, so I'm
not sure where to draw the line.