[Wien] MKL 10 linking

[Andrzej Wronka]

Hello Andrey,

I've observed that strange compilation dependent phenomena, too. Number cycles in both compilation cases are different but  a final effect at self-consistency level seems to be the same and that is more important.  I tested a few manual example  properties  as for DOS, etc. and they all look sensible.  It can suggest some different calculation precision on different "layered" architecture. I will try to run more benchmarks, but it takes some time. 

I think so your benchmark results are very OK for Core 2 Duo desktop. Did you compare it to some MKL9 linking runs as for timing?

 

Greetings,

Andrzej


>Hello Andrzej,

>Sorry for delayed answer as I've been away from the mailing list a while. I've run some benchmarks on my computer (Intel Core2Duo T6400 @ 2.13 GHz, 1 GB RAM with Fedora 7) using my default and your pure layered compiler options. All programs compiled, linked and executed successfully with -O3 -xT, including lapw2, so your problem should lie in some system (or 3rd party) libraries.

>The results of the benchmark (fccni as shown in the UG) are 24 secs with my default model vs. 21 secs with your pure layered model. What bothers me most is the difference in results - it needs 4 cycles to converge with my parameters and 3 cycles w/yours. I'd like to ask whether you observed such a behaviour in your benchmarks. Now I'll try other systems to see if the results obtained are significantly different.

>Cheers,
>Andrey.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20071216/3ea6d1ca/attachment.html