[x264-devel] haswell benchmarking with x264

[Vittorio Giovara]

Hi all,

I got hold of a couple of Haswell servers and an Ivybridge one,
configured as such:

trand1
CPU: 56 x Intel(R) Xeon(R) CPU E5-2695 v3 @ 2.30GHz
Memory: 57.17 GB Free, 62.68 GB Total

trane1
CPU: 48 x Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz
Memory: 42.07 GB Free, 62.81 GB Total

tranb10
CPU: 48 x Intel(R) Xeon(R) CPU E5-2695 v2 @ 2.40GHz
Memory: 40.56 GB Free, 62.87 GB Total

I used x264 to benchmark the AVX optimizations and see how well
haswell machines ran compared to ivybridge. In order to do so I
encoded the Sintel trailer (from https://media.xiph.org/) at different
resolutions. I tried to load in memory the video before running any
benchmarking, so that the results would not be skewed by caching
problems

- The first set of results is a simple `time` execution and I noted
the "real" amount on an average of three rounds.

- The second set is a more detailed `perf stat` over 10 runs.

- Finally there is a functions breakdown with `perf record/report` (on
separate sheets).

https://docs.google.com/spreadsheets/d/18wDEchgkO6UMEaScdnuSUTSxLVF9enSwpsU0us79Jl4/edit?usp=sharing

>From what I can see there is a 30-50% improvement between generations,
and that the machine with less cores but faster cpu beats the other
one. You can also see how the AVX function occurrences do help and
speed up compression.

I am not sure how to interpret the results of the slower machine with
more cores, might I have hit a threading limit in x264? Does these
results match other people benchmarks?

I hope you find these numbers useful. I am more or less new to
benchmarking, so feel free to point criticism at them and suggest any
tip for improving them.

Cheers,
Vittorio