[Libdlo] [PATCH] udlfb: high-throughput urb pool

[Andrew Kephart]

Bernie,

   
On Sat, 2010-12-18 at 14:31 -0800, Bernie Thompson wrote:
> HI Andrew,
> 
[some deleted]
> I've always been nervous about the on-demand model because we can
> easily suck up all the memory we're given (so we need a limit), and we
> churn through a unusually large number of alloc/dma map/free
> operations even for relatively light use.
> 
> We can't avoid needing a system to work within a limit we set on
> memory consumption.
> 
> Best wishes,
> Bernie

All true.  The top-end limit is an obvious requirement, but I think that
we can implement something with minimal alloc churn.  If we work with
the assumption that device usage is fairly consistent (i.e. full screen
video at time T[n] implies full-screen video at time T[n+m]), we could
start with simple on-demand pool expansion with no pool contraction,
which would avoid the churn entirely. Context management might be
tricky, but I'd be willing to give it a shot.

A smoothed contraction mechanism wouldn't be too bad, either, if we find
that the expansion is too hungry for general use.

What do you think?

-andrew