gomp barriers on aarch64

[cbz]

[forget CC'ed to the list.]

On Mon, Feb 27, 2017 at 12:23:39PM +0100, Balázs Oroszi wrote:
> 2017-02-24 13:03 GMT+01:00 cbz <cbz@baozis.org>:
> > Hi,
> >
> > I have recently been running several programs (caffe, mxnet, openblas, blis...) on aarch64.  And I found performance regression when libgomp is used and OMP_NUM_THREADS is set to be >=2. Almost half of the execution time is consumed either in gomp_barrier_wait_end() or gomp_team_barrier_wait_end(). The version of libgomp I used is 5.3.1-14, which is shipped with Ubuntu 16.04. I'm wondering whether it is a known issue on aarch64. Or it might relate to some other factors of the system/hardware?
> 
> I also encountered this behavior some time ago, and I implemented my
> own simple threading to avoid this issue, so I second that.
> 

I have run the EPCC OpenMP micro-benchmark suite[1] to test the synchronization
performance for gcc-5.3.1 on aarch64. As what is expected, the results are
pretty bad, the overheads increase linearly as the number of threads grows.
(800ms overhead when OMP_NUM_THREADS is set to 2.) Then I have updated gcc
to the latest git version (7.0.1 20170226), which looks much more normal.
Though it is still not as good as x86, the overhead when OMP_NUM_THREADS is
set to 2 lows down to ~1ms, which is an acceptable result I think.

Cheers,

Baozi.

[1]
https://www.epcc.ed.ac.uk/research/computing/performance-characterisation-and-benchmarking/epcc-openmp-micro-benchmark-suite)