[Intel-gfx] [PATCH] [RFC] drm/i915: Restore LRU evict order, with a twist!

[Chris Wilson]

On Fri, 14 May 2010 18:37:46 +0100, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> When we need to clear some space in the GTT in order to pin a new
> buffer, scan through the inactive list amalgamating objects in LRU order
> until we find a large enough contiguous space to fit the new buffer.
> 
> ...
> This is benchmarking much better than the current scan through for a
> single buffer large enough to fit.
> ...

Ok my recollection of the current performance of the various backends on
that box was slightly off so the impact isn't as great as I hoped (nor
worse than I feared).

old: unfair
new: lru
Speedups
========
  xcb       firefox-planet-gnome    45640.36 (45789.01 1.80%) -> 41498.25 (41842.95 0.67%):  1.10x speedup
Slowdowns
=========
  drm                  evolution    4146.72 (4229.41 4.16%) -> 4408.15 (4509.70 1.69%):  1.06x slowdown
   gl                    poppler    29186.84 (29233.64 0.40%) -> 31165.07 (31173.13 0.05%):  1.07x slowdown

The most interesting impact is the firefox-planet-gnome trace which does
cause mass eviction, and where xcb consistently lags the xlib backend but
with the fair eviction xcb improves and is on a par with xlib.

This only leaves the question of whether it provides sufficient protection
against the page-fault-of-doom.
-ickle

-- 
Chris Wilson, Intel Open Source Technology Centre