[Intel-gfx] gem cache size vs performance

Chris Wilson chris at chris-wilson.co.uk
Fri Jun 5 14:33:13 CEST 2009

On Fri, 2009-06-05 at 13:20 +0800, Li, Shaohua wrote:
> I recently did some benchmarks to verify the relationship about gem cache
> size and performance. The workload is x11perf -rgb10text. I use attached
> patch to gather statistics about gem cache.

Interesting. I'm currently experimenting with pruning the bo cache after
a miss, and seeing better results than setting arbitrary max_entries
whilst replaying firefox traces.

Paraphrasing from cairo-drm:
  while (dev->bo_cache_size > dev->bo_max_cache_size)
    /* trim caches by discarding the most recent buffer in each bucket */
where bo_cache_size is the number of bytes within buffer objects stored
in all buckets and bo_max_cache_size is currently set as gtt_max_size/4.

More information about the Intel-gfx mailing list