NUMA and why XFree86 should keep the wire protocol...

Just a quick note, as I read that there was on open discussion on the future
of X. Some people have been blaming the X wire protocol (and its network
retargetability) for poor performance. I think other people have made the
case that it imposes very little overhead quite well. What I am interested
in (as a hardware engineer) is that as the power of GPUs on graphics cards
increases, there will come an opportunity to incorporate the X-Server into
the graphics card directly. If this were to happen, then the separation of
server memory/client memory would map exactly onto graphics memory and
system memory, and shared memory would effectively be AGP mapped memory (on
the Intel-architecture). Then having an efficient wire protocol becomes
extremely important as this would be data that actually has to be sent to
the graphics card. This may not happen though as graphics card manufacturers
tend to target windows as the major platform, however, don't forget this
model can be extended to multi CPU machines - especially NUMA architectures.
In these kind of machines (AMDs hammer architecture is an example  of a NUMA
machine) one CPU is attached to the IO chain, the others are not. This means
that again the wire protocol directly maps to data packets sent between
CPUs... I don't even want to consider how inefficient a frame-buffer model
is for these up and coming architectures.

To summerise computer architecture is moving towards an internal network
structure (Intels - point-to-point bus replacement for PCI for example) and
NUMA architectures. X's separation of client and server with a wire protocol
in the middle suits these architectures very well (which is not surprising
as X was designed with networks in mind). When NUMA dominates (which it will
because it scales better than SMP) other windowing systems will be playing
catch up - don't throw away X's advantage too soon!

	Keean Schupke,
	Department of Electrical & Electronic Engineering,
	Imperial College, London.