Why care about indirect rendering ?

[Keith Whitwell]

Eric Anholt wrote:
> On Fri, 2005-09-02 at 10:26 +0200, Clemens Eisserer wrote:
> 
>>Hi there,
>>
>>
>>>X uses a network protocol.  While I love the network transparency and
>>>wouldn't lose it for anything it doesn't mean you want it when you're
>>>local.  In fact you want as little as you reasonably can between the
>>>application and the video card.  That's in part why I find this
>>>indirect-rendering centered discussion so strange.
>>
>>This is something I also never really understood. X forces you to go
>>through the network protocol even for drawing also if you're local.
>>Windows allows to plug-in a network-layer if you need it (however far
>>less professional than X).
>>Sure, Unix Domain Sockets are highly optimized - but at least parsing
>>out all the trafic should be quite expensive? Or am I completly wrong?
> 
> 
> Do you have any evidence that the travel over unix domain sockets is an
> actual bottleneck?
> 
> This would be the first step in any discussion about making a system to
> avoid the transport layer.  I have yet to see anyone bring evidence to
> the table.  On the other hand, everyone involved with X (myself
> included) that I've heard talk about it totally unconcerned with the
> overhead that might exist from network transport for core operations.  I
> know when I look at my profiles, network transport never shows up --
> it's always unaccelerated graphics operations.

Rik Faith put quite a bit of time into a Shared Memory Transport 
alternative to the network layer while at Precision Insight.  The 
results are committed to XFree86 on a branch (look for "smt" in the 
branch name).  As I understand it, results were not an improvement on 
domain sockets, and I think this has been the case for most or all 
attempts at this.

Keith