[self-interest] Re: Sparc 2 with Solaris 2.6

[David Ungar]

Jecel,

Thanks for the numbers!
The only ones I would not trust would be the bytecode-per-second
ones.
In Smalltalk and Self, costly primitives, such as BitBLT, or an 
enumeration through
memory, show up as a single bytecode. You cannot average those
with pushes and pops.

- Dave





At 3:04 PM -0200 12/30/99, Jecel Assumpcao Jr. wrote:
>Hi, John!
>
>>  Thanks a bunch, Jecel.  Sorry for the extra email I sent you about this
>>  :-(
>
>Strangely enough, I only got that one today.
>
>>  On your recommendation I decided to go ahead and buy it (plus, the
>>  price wan't bad at only $200).
>
>I guess you didn't get a monitor with that? It is a good bargain in any case,
>but getting one of those old workstation monitors is really a treat.
>
>I was looking through some old emails to find some performance 
>numbers. Here is
>the number of bytecodes per second that Squeak reported in April, 1997 (in
>millions):
>
>    Sparc 10    1.7
>    Sparc 20    3.6
>    UltraSparc  5.2
>
>While it is not possible to do a direct comparison, my Ultra 5 does 10 million
>bytecodes per second in Squeak 2.2. I can't find any numbers for the Sparc 2,
>but I seem to remember it was around half that of the Sparc 10.
>
>In May 1997, I wrote:
>Here is the raw data for the two popular Squeak benchmarks
>on two different Sparc machines (and the same for the
>programs translated into Self):
>
>Machine: UltraSparc 1
>
>   Squeak:
>      Time millisecondsToRun: [ 10 benchmark ] 948 963 966 984 1045
>      Time millisecondsToRun: [ 26 benchFib ] 1264 1252 1251 1274 1238
>
>   Self 4:
>      [ 10 benchmark ] realTime 1619 329 510 239 501 252
>      [ 26 benchFib ] realTime 556 40 42 45 38 39
>
>Machine: Meiko Multi Sparc 20
>
>    Squeak:
>       Time millisecondsToRun: [ 10 benchmark ] 1508 1492 1491 1524 1513
>       Time millisecondsToRun: [ 26 benchFib ] 3895 3941 3925 3923 3920
>
>    Self 4:
>       [ 10 benchmark ] realTime 1478 268 253 266 180 286
>       [ 26 benchFib ] realTime  1797 156 146 415 162 154
>
>Well, these numbers are all very obsolete now. But they are the best 
>I have for
>testing the relative speeds of various processors. Of course, all benchmarks
>are suspect. For example, the "bogoMIPS" number on my new 350MHz Pentium II
>machine is almost exactly the same as on my old 166MHz Pentium computer. The
>ultimate test is to run the application you want and see how it performs.
>
>64MB of memory should be enough as long as you do normal things (I used to get
>by with half of that). If you start to enumerate all objects (like I 
>do all the
>time) then you will be constantly swapping to disk and things will be
>fantastically slow no matter what CPU you have.
>
>Note that the first number beside each Self expression above is always much
>larger than the ones that follow it. This is the effect of adaptive
>compilation. This will also make the menu take much longer to pop up the first
>time you press the mouse button and so on. So a quick test of the Self UI can
>sometimes leave people with a very negative impression of its performance. It
>is much more reasonable in real use, however.
>
>-- Jecel
>
>------------------------------------------------------------------------
>Tired of the same old games?
>Play the most popular gameshows online by the co-creator of Jeopardy
>At BoxerJam.com—Join now and enter to win $1000 cash
>http://click.egroups.com/1/485/5/_/8257/_/946578404
>
>-- 20 megs of disk space in your group's Document Vault
>-- http://www.egroups.com/docvault/self-interest/?m=1



     David Ungar
     Sun Microsystems Laboratories
     (650) 336-2618