Clearly my "native" implementation somehow fails to capture some goodness from
TBB. I'll see if I can figure out the missing ingredient. In the mean time,
we'll keep TBB in place until we can match its performance.
-- lg
On Apr 18, 2012, at 7:43 AM, Ramon Montoya Vozmediano wrote:
> Here is the comparison between the old code with TBB and the patch with TBB
> off. It does look like performance is hit as threads increase.
>
> Done on a 24/12 cores machine running linux, rendering production data with
> arnold. I ran two comparisons, one with AA_samples 1 and another with
> AA_samples 6,
>
> Columns are:
> number of threads
> time
> cpu time
> system time
> time speedup
> threading efficiency ( adjusted to disregard hyperthreading effects)
> I applied the patch to OpenImageIO ver 0.10.7,
>
> TBB no TBB + patch
> ------- textures ON ------- textures ON
> t1 29.9 29.8 0.1 x1.0 100% t1 29.6 29.5 0.1
> x1.0 100%
> t6 6.5 31.4 0.2 x4.6 94% t6 6.6 31.9 0.2
> x4.5 92%
> t12 4.0 32.7 0.5 x7.5 90% t12 4.0 33.6 0.5
> x7.3 87%
> t18 3.6 41.9 1.1 x8.2 70% t18 3.9 45.4 2.3
> x7.6 62%
> t24 3.6 51.6 3.3 x8.3 55% t24 4.7 67.5
> 10.5 x6.3 38%
> ------- textures OFF ------- textures
> OFF
> t1 23.8 23.8 0.1 x1.0 100% t1 23.4 23.3 0.1
> x1.0 100%
> t6 5.4 25.3 0.1 x4.4 94% t6 5.3 24.6 0.1
> x4.4 95%
> t12 3.2 25.3 0.1 x7.5 94% t12 3.2 25.1 0.1
> x7.4 92%
> t18 3.0 32.1 0.1 x7.9 74% t18 3.0 31.4 0.1
> x7.8 74%
> t24 2.8 36.6 0.1 x8.6 65% t24 2.7 35.8 0.1
> x8.5 65%
>
> And the same scene with AA_samples set to 6
>
> TBB no TBB + patch
> ------- textures ON ------- textures ON
> t1 785.8 785.6 0.1 x1.0 100% t1 797.7 797.5 0.1
> x1.0 100%
> t6 140.7 833.5 3.0 x5.6 94% t6 143.3 847.8 4.0
> x5.6 94%
> t12 75.9 878.9 12.0 x10.4 88% t12 82.3 944.6
> 23.5 x9.7 82%
> t18 80.2 1348.0 52.6 x9.8 56% t18 95.7 1509.1
> 149.5 x8.3 48%
> t24 85.6 1729.8 219.7 x9.2 40% t24 133.9 2460.0
> 545.2 x6.0 27%
> ------- textures OFF ------- textures
> OFF
> t1 617.2 617.1 0.1 x1.0 100% t1 620.4 620.2 0.1
> x1.0 100%
> t6 107.0 634.7 0.1 x5.8 97% t6 107.4 637.2 0.1
> x5.8 97%
> t12 55.2 646.1 0.2 x11.2 96% t12 55.0 643.3 0.2
> x11.3 96%
> t18 55.1 963.1 0.2 x11.2 64% t18 50.2 878.2 0.3
> x12.4 71%
> t24 47.1 1088.8 0.3 x13.1 57% t24 47.5 1101.5 0.4
> x13.1 56%
>
> r
>
> On Thu, Apr 12, 2012 at 7:42 AM, Ramon Montoya Vozmediano
> <[email protected]> wrote:
> Thanks for the clarification!
>
> Will do.
>
> r
>
>
> On Wed, Apr 11, 2012 at 6:44 AM, Larry Gritz <[email protected]> wrote:
> Fantastic, thanks. But please don't just test TBB vs no TBB with old code...
> you need to test my patch from this pull request, which eliminates much of
> the performance gap of the "no TBB" case.
>
> -- lg
>
>
> On Apr 10, 2012, at 7:56 AM, Ramon Montoya Vozmediano wrote:
>
> > Hi,
> >
> > At Solid Angle we have some customers' scenes that we use to test threading
> > performance, we will report back on how TBB vs no TBB performs on linux and
> > windows,
> >
> > Ramon
>
> --
> Larry Gritz
> [email protected]
>
>
> _______________________________________________
> Oiio-dev mailing list
> [email protected]
> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
>
>
>
> _______________________________________________
> Oiio-dev mailing list
> [email protected]
> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
--
Larry Gritz
[email protected]
_______________________________________________
Oiio-dev mailing list
[email protected]
http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
