[Wien] WIEN and Mathematical libraries under Linux

Stefaan Cottenier Stefaan.Cottenier at fys.kuleuven.ac.be
Fri Jul 4 12:17:08 CEST 2003


> In the last weeks there were several emails regarding WIEN and
> mathematical libraries under LINUX, in special Intel's MKL.
>
> I have two comments that I think can be interesting for some of you.
>
(...)
> So, users can continue to use the library without any problem.

That's good news. The small catch is that you have to register with your
email address for the first download of mkl, and therefore can not use that
mail address any more when you would like to get a newer trial version say 1
year later.
>
> 2) I have also compared the performance of WIEN with MKL and ATLAS, for
> different matrix sizes (I have tried only the real version of lapw1).
> My computers are Athlon XP 2000+ under Redhat 8.0 and with ifc compiler.
>
> The behavior (time of running for lapw1 in seconds) was:
>
> # Matrix Size      MKL     ATLAS     MKL/ATLAS
>    521 (10kpts)   16.52    17.80      0.928
>   1083 ( 9kpts)  108.02   111.22      0.971
>   1913 ( 9kpts)  466.42   478.72      0.974
>   2490 (10kpts) 1081.68  1077.16      1.004
>
> As you can see ATLAS behaves more or less similar to MKL at least for
> my system.

I can supplement these data with similar tests I have done (P-IV 2.4 GHz and
Athlon 2200+, IFC 7.1, times are for lapw1, matrix sizes are 1580 for case 1
and 1751 for case 2) :

      CPU
     compiler
     library
     Case 1
     Case 2

      Pentium IV
     IFC 7.1
     (precompiled)
     1:03:57
     1:06:01

      Pentium IV
     IFC 7.1
     ATLAS 3.4.1
     0:29:17
     0:31:31

      Pentium IV
     IFC 7.1
     mkl 6.0
     0:22:41
     0:24:05

      Athlon
     IFC 7.1
     (precompiled)
     1:59:24
     2:07:43

      Athlon
     IFC 7.1
     ATLAS 3.4.1
     0:44:01
     0:45:01

      Athlon
     IFC 7.1
     mkl 6.0
     0:43:53
     0:46:43


 On Athlon, there is no real time difference between mkl and ATLAS, which
agrees with your conclusion. On P-IV however, the ratio mkl/ATLAS is around
0.75 (which makes sense, as mkl is designed for Intel cpu's). Note also the
huge speed difference with the precompiled executable, and the fact that a
P-IV with mkl is twice as fast as Athlon (with these clock speeds).

Stefaan




More information about the Wien mailing list