Re: [Flightgear-devel] Multi core discussion

[Anders G. <and...@gi...>]

On Tue, 6 Jul 2010, Alex Perry wrote:

> Most of the _non-minimal rendering options use considerable CPU power
> on any GPU that is routinely available in a mobile form factor.  If we
> want to be able to support non-desktop users (as well as simplify
> demonstrations at trade shows), we need to make sure that the main
> thread (and in fact that whole core) are entirely dedicated to
> rendering.

Or hope that the OSG and driver developers find good ways to parallelize 
the CPU part of the rendering process.

> The reason why FG itself doesn't use any CPU power is that, whenever
> any of the subsystem developers tries to implement an underlying
> simulation improvement that requires non-trivial amounts of CPU, there
> is massive complaint from the eye candy side of the community.  This
> appears as a tradeoff precisely because FG doesn't support
> multithreading. If we had threading working safely, any simulation
> subsystem could choose to run in its own thread and eat an entire core
> as needed.
>
> The single threaded limitation currently impacts the implementation of
> real time weather, microcell airmass, AI operations, non-steam
> instruments, as well as radio navigation realism.  That I'm aware of
> ... there are presumably others too.

My suggestion would be to rather look into parallelization inside the 
subsystems that need a lot of computation (think worker threads) - not 
least since things like microcell airmass simulation may well need more 
than one thread/core/CPU - and the problem itself could be of a type that 
is easy to parallelize. This would keep the top level "main loop" 
integration simple too.

Cheers,

Anders
-- 
---------------------------------------------------------------------------
Anders Gidenstam
WWW: http://www.gidenstam.org/FlightGear/