[Wien] Problem with Goto library on Itanium2 ...

Vladimir Timochevski Vladimir.Timochevski at umontreal.ca
Thu Jun 17 17:57:01 CEST 2004


Dear Wien users,

Unexpectedly I faced a strange behavior of Goto library on Altix 
Itanium2 platform. After a certain matrix size (5000) Goto library goes 
  crazy and gives completely wrong eigenvalues. The output for the 
benchmark example looks like (I increased RK to 8.0):

       MATRIX SIZE 5039  WEIGHT= 1.00  PGR:
      EIGENVALUES ARE:
      -6.9796877   -6.9097097   -6.7796200   -6.7435624   -6.6421138
      -6.3867387   -6.1792930   -6.1251950   -5.8655448   -5.8171472
      -5.6519827   -5.6213795   -5.3374324   -5.2785844   -5.0657352
      -4.8613504   -4.8510333   -4.7831951   -4.4799485   -4.4201590
      -4.2822613   -4.1265467   -3.9850248   -3.8917386   -3.7772234
      -3.6793028   -3.6122531   -3.4510670   -3.2727999   -2.9833373
      -2.8502134   -2.5283070   -2.3909689   -2.2270614   -1.9973846
      -1.7946345   -1.6771091   -1.4776752   -1.3641985   -1.2552173

      -1.2241327   -1.0943078   -0.8017713   -0.7922804   -0.7320057
      -0.5912242   -0.5091300   -0.4767198   -0.3515971   -0.1713052
      -0.1273214    0.0331076    0.1060666    0.2308597    0.2709128
       0.3947352    0.5600952    0.6981032    0.7814915    0.8320300
       0.8601380    1.1090431    1.1544852    1.2144447    1.2388167
       1.3874186    1.4676762
           189 EIGENVALUES BELOW THE ENERGY   -7.00000

which is completely crazy: the smallest eigenvalue should be -0.769 ...

However there are no problems with SCSL and MKL libraries (except for 
they are almost twice slower ...).

Details:
Platform: Altix Itanium2
Compiler: Intel 8.0
Comp. options:  -FR -w -mp -tpp2 -O2
library: goto_it2-r0.94

I will appreciate any comments and suggestions.

Thanks.
Vladimir.

---------------------------------------------
Vladimir Timoshevskii
Departement de Physique, Universite de Montreal
C.P.6128, succ. Centre-Ville,
Montréal (Québec) Canada, H3C 3J7
Tel.: +1 (514) 343 611 - 4226
email: Vladimir.Timochevski at umontreal.ca
---------------------------------------------





More information about the Wien mailing list