[Intel-gfx] An apparent large performance regression of 3D display over the network

[Alan W. Irwin]

Three years ago on a Debian testing system that was to become Debian
lenny, I made a test of running low-end 3D games (tuxracer and
foobillard) on a remote box on my 100Mbit LAN while displaying them on
my local Debian testing box X server with g33 chipset.  The games were
quite playable (in fact indistinguishable from playing the games
locally if I recall correctly).

Fast forward to today, when I once again tried the foobillard part of
the test (I didn't bother with tuxracer) with the same local (g33)
hardware but this time with Debian testing (squeeze) installed on that
locally displaying box.  foobillard has become completely unplayable
over the network with the former smooth movement reduced to what looks
like a series of snapshots with large gaps in between.  The LAN
network I have now is 1 gigabit as opposed to the older 100Megabit LAN
network I had when the remote 3D games worked well.  I can play
foobillard and tuxracer just fine locally on that machine so it
appears local 3D is in reasonable shape.

foobillard and tuxracer are just subjective tests of whether 3D
rendering works reasonably efficiently over the network, but my
impression is the regression in that regard is at least one or two
orders of magnitude in speed in order to reduce smooth effects to a
series of snapshots.  Thus, objective tests of remote 3D efficiency of
the old Intel stack from three years ago versus the current one should
pick up this performance regression easily.

I recently bought another computer (ASUS Eee Box with 945GME chipset)
that shows foobillard is unplayable over the network in the same way
while local running of that game is fine on that box.  I have now
configured that box to be an X-terminal (a configuration I far prefer
because it reduces sysadmin issues a lot).  The 2D KDE desktop
displays well for that configuration, but I have extreme doubts
(haven't tried them yet) about whether remote 3D desktop effects will
work at all considering this huge slowdown I get with remote running
of foobillard over the local 1 Gigabit LAN with that X-terminal.

One possibility is there may be something extra I have to do now to
make remote 3D display efficient over an ssh connection.  Advice in
that regard would be helpful.  (Currently, I just set ForwardX11 yes
and ForwardAgent yes for the host in question in .ssh/config for
the local computer.)

But if ssh configuration is not the issue, then it appears there has
been an efficiency regression for remote 3D at least for the 945GME
(GMA 950) and g33 (GMA 3100) chipsets. Has anyone here found
foobillard or similar low-end 3D games to be playable or 3D desktop
effects to work reasonably efficiently over fast LAN networks with
today's Intel graphics driver?

Of course, if this really turns out to be a general regression in
remote 3D display efficiency, then that regression obviously
corresponds with the X stack reorganization by Intel that has occurred
over the last 3 years. I expect making local 3D display efficient for
that newly organized stack is still one of the top priorities for
Intel developers, but I hope dealing with this efficiency regression
for remote 3D (if that is what it is) is at least on the agenda. After
all, with 3D desktop effects becoming more and more important and with
low-end 3D games as a "would be nice", reasonably efficient X network
transparency for 3D display is an important issue for those using X
terminals.

Let me know if there are more quantitative tests of efficiency you
would like me to run between local and remote display of 3D on either
the 945GME or g33 boxes.

Alan
__________________________
Alan W. Irwin

Astronomical research affiliation with Department of Physics and Astronomy,
University of Victoria (astrowww.phys.uvic.ca).

Programming affiliations with the FreeEOS equation-of-state implementation
for stellar interiors (freeeos.sf.net); PLplot scientific plotting software
package (plplot.org); the libLASi project (unifont.org/lasi); the Loads of
Linux Links project (loll.sf.net); and the Linux Brochure Project
(lbproject.sf.net).
__________________________

Linux-powered Science
__________________________