Hi Douglas, hi Spencer,

With all the measures, the total speed up should be around a factor of 20 (2 for openblas, 2 for auto tuning and 5 for the code optimization). 

Thank you for your support to solve the problems. 

Have a nice weekend,

Helmut


Von: Sherwin, Spencer J [s.sherwin@imperial.ac.uk]
Gesendet: Donnerstag, 19. Mai 2016 18:22
An: Kühnelt Helmut
Cc: Douglas Serson; nektar-users
Betreff: Re: [Nektar-users] EulerCFE - performance issue & MPI problem

Hi Helmut, Douglas

Thanks for Douglas’ analysis. He spotted that in some heavily used routines there were some unnecessary copies which made new arrays or shared pointers. This lead to a lock call and a counter initialisation which in your case where there is very little actually work in these methods were being dominated by these unnecessary calls. 

So I believe this orphaned code was the issue in your case. It also seemed related to your relatively slow mobile cpu since we were unable to reproduce this on our own laptops or desktops. Also it did not seem the boost version made much difference as you also mentioned.

Anyway good to have gotten to the bottom of this. We have this in a fix branch and will upload it into master and the web page distribution as soon as possible.

Cheers,
Spencer.

On 19 May 2016, at 15:48, Kühnelt Helmut <Helmut.Kuehnelt@ait.ac.at> wrote:

Hi Douglas, hi Spencer,

I run the test case with the feature/Euler1D branch 
with SumFac : 
Time-integration  : 18.7779s
Total Computation Time = 21s
This make a difference!

no tuning:
Time-integration  : 32.797s
Total Computation Time = 35s
Also no too bad!

I already notice the orphaned code that you removed. Was this the solution of my speed issue, or was it because I used the release version 4.3.1.?

Cheers,
Helmut


Von: Douglas Serson [d.serson@gmail.com]
Gesendet: Donnerstag, 19. Mai 2016 15:40
An: Kühnelt Helmut
Cc: Sherwin, Spencer J; nektar-users
Betreff: Re: [Nektar-users] EulerCFE - performance issue & MPI problem

Hi Helmut,

I pushed some changes to the feature/Euler1D branch which should make this simulation somewhat faster.
It would be great if you could test that to see if it helps with your problem.

Cheers,
Douglas

2016-05-18 13:29 GMT+01:00 Kühnelt Helmut <Helmut.Kuehnelt@ait.ac.at>:
Hi Spencer, hi Douglas,

Indeed, using Collections (auto => SumFac) definitively helps: Total Computation Time = 95s (180s before)

I noticed that the code runs single threaded, even though Boost_USE_MULTITHREADED is ON and OpenBlas is compiled as multi threaded. Any hints on that?

Cheers,
Helmut


Von: Sherwin, Spencer J [s.sherwin@imperial.ac.uk]
Gesendet: Mittwoch, 18. Mai 2016 10:25
An: Kühnelt Helmut
Cc: nektar-users
Betreff: Re: [Nektar-users] EulerCFE - performance issue & MPI problem

HI Helmut,

I am building with Release mode and nothing else in particular. I happen to have MPI turned on but I do not think that will make a difference.

One thing that might be worth checking is the use of openblas. I have had trouble with openblas in the past for small matrix sizes. The compilation I used will have been using the default installation in /usr/lib/libblas.a. I guess Douglas’ test will have used the Framework of Mac OS X.  However looking at the set up on the compute module I do seem to have openblas loaded 

(@Chris: Could you confirm what version of openblas is running on Victoria 

Other configurations are:
library
gcc             4.9.2
boost         1.58 - installed from a module

Do you have another machine/laptop we could turn this test on? 

Cheers,
Spencer.

PS I had a time step of 1e-6 and attach my .xml file below 

Spencer  Sherwin
McLaren Racing/Royal Academy of Engineering Research Chair, 
Professor of Computational Fluid Mechanics,
Department of Aeronautics,
Imperial College London
South Kensington Campus
London SW7 2AZ

+44 (0) 20 759 45052