RE: Gabriel Benchmarks for 3675/3645/3650

[DUMOULIN@TITAN.KSC.NASA.GOV (Jim Dumoulin)]

  In reference to why 3650 Benchmarks were slow...
 
  When I ran benchmarks on our machines, I noticed a number of differences
between the results we got and those David Loewenstern presented in a
previous posting.  They are:

    1.  There are a number of tests where the times listed look like
        they were transcribed with the decimal point in the wrong
        place.  (For instance FRPOLY15R2 runs in 12.228 on our 3675
        and you list it as 1.397).

    2.  The times you post are in "CLOCK" times which is "CPU+PAGING".
        Most of the benchmarks published by others are in CPU time.  This
        makes your numbers appear high when compared to others.

 Some General Observations:

   A.  Your results showed very close timing between your 3645 and 3675.
       I don't find this suprising because the L-Bus Lisp processor
       board set is identical between these two machines.  The only
       thing that is usually different is the disk interface.  The
       original 3640's that we have were shipped with ST-506 disks
       and our 3675 has SMD disks.  Symbolics offers an upgrade
       package that upgrades ST-506's to SMD's or ESDI's so I'm not
       sure what disk configuration you have. 
 
           For comparison, at my site we have:

           On our 3675's:       Fujitsu Eagle 474Mbyte SMD Disk
                                14.872 Mbit/sec transfer rate
                                18msec average access time
                                6 platters / 20 Data heads
 
           On our 3640's        Maxtor XT1190 190 Mbyte ST-506 Disk
                                5 Mbit/sec transfer rate
                                30msec average access time
                                8 platters / 15 Data heads

           On our 3650's        Maxor XT-8760E ESDI Disks
                                15.06 Mbit/sec transfer rate (listed)
                                (but I think its closer to 10Mbit/sec)
                                19msec average access time
                                8 platters / 15 Data heads
 
                                Trident T-306 300Mbyte Removable SMD
                                9.68 Mbits/sec transfer rate
                                30msec average access time
                                10 platters / 19 heads

          On our MacIvories     Jasmine 300Mbyte SCSI Disks
                                1.5 Mbit/sec transfer rate
                                30 msec average access time.
 
          On our XL-400's       ESDI Disks (same as the 3650's)

       The type of disk (and other factors such as I/O) can make quite
       a difference in how your machine performs (regardless of what
       the Gabriel Benchmarks say).  For instance, the Gabriel suite
       runs faster on our MacIvory-2's than on our XL-400's but because
       of the slow disk and I/O on the Mac, our applications run MUCH
       slower on a MacIvory-2 than they do on an XL-400.

   B.  I think its a common misconception that people think the 3650's
       are faster than the 3675's.  Your benchmarks (even after they
       are adjusted for errors in decimal points) show that for most
       benchmarks (especially ones like PUZZLE and TRIANG that test
       searches and other IFU related operations) the IFU in a 3675
       is faster than the Instruction Pre-fetch hardware in a Gate-
       Array machine like 3650/53/30/20's.  I talked to my local
       Symbolics folks about this (Jeff Brooks and Mark Sorrells) and
       they were aware of this.  They suggested reading old SLUG
       archives at about the time the 3650's were first introduced.
       Mark seems to remember there was quite a discussion on SLUG
       about this.
 
                               Hope this helped,
                                       Jim Dumoulin
                                       NASA Payload Operations
                                       Kennedy Space Center FL