[FFmpeg-user] ffmpeg multithreaded performance bad for specific physical machine

[P. Remek]

Hello,

I am facing rather interesting problem and am slowly running out of
ideas. It seems that ffmpeg running on one of our physical machines
has performance problems when running with multiple threads.

We have two physical boxes:

First: 2-socket Intel Xeon E5-2630 v2
Second 4-socket Intel Xeon E7-8867 v3

First seems ok, when I run encoding with 1 thread I get ~20 fps, with
2 threads ~40 fps. So it scales up as expected.

The second however, with 1 thread I get ~25 fps, and with 2 threads it
stays roughly the same on ~25 fps - so, no scale up.

What is also interesting that the CPU consumption (viewed using htop)
looks comparable on both machines when running 2 threads - basically 3
cores are spinning somewhere around 80% but with different end result
(20 vs 40 fps).

I already excluded hyperthreading (disabled on both), numa issues (I
am starting the command with taskset to bind the execution to cores
running on same socket) and specific instructions which are different
on our different CPU families - like for example AVX2 which is only on
our E7 cpu (I disabled everything using -cpuflags 0, -x264opts asm=0)

I also tried to compile the binaries directly on the tested machines
to exclude some CPU-specific compile time optimizations which might
take effect.

Does anybody have an idea where else could be the problem?

Here is the command which I use to run the test:

root at lab-node3:~/ffsource/ffmpeg-3.1.3# taskset -c 1,2,3 ./ffmpeg
-cpuflags 0 -f lavfi -i testsrc=duration=60:size=1024x576:rate=25
-threads 2 -preset veryfast -c libx264 -x264opts asm=0 -f mpegts -
>/dev/null

Regards,
Prema