[daip] GNU compilers and DDT/Y2K

[Eric Greisen]

Linux  1.3 GHz 512 Mbyte                   10/02/02
   Trying the new 3.2 version

   3.2      82 - 84        and  853 - 854      Cpu= 12.34 - 12.65 (min)
                               (881, 898 also occurred for no obvious
                               reason)

C-----------------------------------------------------------------------
Linux  1.3 GHz 512 Mbyte                   9/20,23/02
   Using a new compiler, there were no failures and there was
   acceptable accuracy, identical (nearly) in both versions  Using the
   2.95.3 compiled from 7.2 on primate rather than the one on bob:

   3.1.1    81 - 82        and  793 - 807      Cpu= 12.40 - 12.44
   2.95.3   70 - 71        and  678 - 692            9.98 - 10.12

   In fact there seems to be a modest improvement perhaps due to using
   the load modules on primate rather than on filehost.

C-----------------------------------------------------------------------
Linux  1.3 GHz 512 Mbyte                   2/05/02
   I have tested GNU 2.95.3, 3.0.2, and 3.0.3 compilers on the new
   primate.  All pass the DDT and Y2K tests with the following
   exception:

   3.0.3  Miscompiles COMB, messing up on an array (7,4) in COMMON.
          When it is accessed, other bits of memory get changed

          MATHS is reported to produce wrong answers.

          The host name becomes !!imate apparently right off when AIPS
          is started.

   With COMB compiled NOOPT, the DDTs and Y2Ks ran.

   Performance was interesting - times for DDT and Y2K

   2.95.3   70 - 75 sec    and  706 - 711 sec   Cpu= 10.32, 10.28, 10.217

   3.0.2    91 - 93        and  802             Cpu= 11.84

   3.0.3    92 - 93        and  814 - 829.      Cpu= 11.86, 11.77

C-----------------------------------------------------------------------

Linux 1.8 GHz dual 1 Gbyte memory (capella)   2/25/02

    2.95.3 only   49 - 50  sec   and  474 - 476 sec  (80 - 84 Amarks)

C-----------------------------------------------------------------------

Linux  1.3 GHz 512 Mbyte                      3/10/02

   3.0.4  Miscompiles COMB, messing up on an array (7,4) in COMMON.
          When it is accessed, other bits of memory get changed

          The host name becomes !!imate apparently right off when AIPS
          is started.

   With COMB compiled NOOPT, the DDTs and Y2Ks ran.

   2.95.3   70 - 75 sec    and  706 - 711 sec   Cpu= 10.32, 10.28, 10.217

   3.0.2    91 - 93        and  802             Cpu= 11.84

   3.0.3    92 - 93        and  814 - 829.      Cpu= 11.86, 11.77

   3.0.4    83 - 85, 91    and  741, 739, 736   Cpu= 11.68, 11.70, 11.77

C-----------------------------------------------------------------------

Linux  2.2 GHz Xeon dual-headed w 2 Gbytes ram                3/23/02

   DDT:   36   37 , 37   -> 108 - 111 AM
   Y2K   373, 379, 388   -> 103 - 107 AM

C-----------------------------------------------------------------------
One is a SunBlade 100, currently called tmpip, and the other is an
Ultra 60 known as tecs. The Ultra 60 belongs to the ALMA control
effort. The SunBlade sort of fell into our hands when an anticipated
data analyst position was frozen.

Test               tmpip            tecs

DDT real (s)   166 - 169       160 - 169
AM                 24              24
cpu (m)            2.5             2.0

Y2K real (s)  2168 - 2169     1795 - 1801
AM                18.4            22.3
cpu (m)           33.7            23.6

Thus tmpip (SunBlade) had a good cpu/real ratio but appears a bit
slower than the Ultra 60.  The tecs (Ultra 60) has a suprising
cpu/real ratio or it would have appeared better still relative to the
SunBlade.