mike@RICE.EDU (Mike Caplinger) (12/23/85)
I just got my 68020 upgrade, taking my Apollo up to a DN330. One of the first things I did was run the Dhrystone benchmark on it. I got a number somewhere around 1700 - about the speed of a VAX/11-780. Fine and good. Now I see some Sun-3 results that run in the 3000s! Are these for real? Would someone explain to me why the Sun apparently runs almost twice as fast as a DN330 with an identical processor? Mike Caplinger mike@bellcore.arpa ihnp4!bambi!mike
holtz%cascade.carleton.cdn%ubc.CSNET@CSNET-RELAY.ARPA (Neal Holtz) (12/24/85)
I sure can't say anything about those Sun-3 times, and I am
curious about it also. Could it be explained by simply a better
compiler?
For example, the following table is part of one
circulated in USENET a while back, with times I've added
for various Apollo's. Note that the Pascal versions run
much faster than the C versions (and likely do things
differently, so one shouldn't draw too many conclusions):
* MACHINE PROCESSOR OPERATING COMPILER DHRYSTONES/SEC.
* TYPE SYSTEM NO REG REGS
* -------------------------- ------------ ----------- ---------------
** Apollo DN300 68010-?Mhz Aegis SR8 cc r2.39 686 -
** Apollo DN320 68010-?Mhz Aegis SR9 cc r3.12 808 -
* VAX 11/750 - Unix 4.2bsd cc 862 877
* VAX 11/750 - VMS VAX-11 C 2.0 958 1091
* Sun2/120 68010-10Mhz Sun 4.2BSD cc 1136 1219
** Apollo DN300 68010-?Mhz Aegis SR8 pas r5.65 1332 -
** Apollo DN320 68010-?Mhz Aegis SR9 pas r6.59 1349 -
* VAX 11/780 - UNIX 5.2 cc 1515 1562
** Apollo DN660 Aegis SR9 cc r3.12 2030 -
** Apollo DN660 Aegis SR9 pas r6.59 2790 -
* IRIS 2400T 68020-16.67MHz Sys5 #14 cc 3105 3401
* SUN 3/75 68020-16.67Mhz SUN 4.2 V3 cc 3333 3571
*
** Apollo times added by N. Holtz, Nov. 29, 1985, from results
of March 84 version C/2 of Dhrystone (no register variables).
By the way, that figure reported of 1700 looks pretty low for a
DN330 -- I thought they were supposed to be slightly faster
than a DN660. Were all the 'fast' options turned on?
I suspect this is not the case (I haven't looked at the dhrystone
benchmark very carefully), but are there some funny optimizations
possible? For example, when I first tried the following naive
benchmark (out of BYTE + many other places, I believe), I got
incredible (orders of magnitude) speed ups over other machines.
The Apollo compiler quite rightly did all the arithmetic at compile
time, and replaced the 14 statements with 14 MOVE's (but why didn't
it remove the the first 13 moves, at least, plus the loop as well?).
So obviously, this is pretty useless as a benchmark.
#define CONST1 3.1415926
#define CONST2 1.7839032
#define COUNT 10000
main()
{
double a, b, c;
int i;
a = CONST1;
b = CONST2;
for (i = 0; i < COUNT; ++i) {
c = a * b;
c = c / a;
c = a * b;
c = c / a;
c = a * b;
c = c / a;
c = a * b;
c = c / a;
c = a * b;
c = c / a;
c = a * b;
c = c / a;
c = a * b;
c = c / a;
}
printf("Done\n");
}JW-Peterson@UTAH-20.ARPA (John W Peterson) (12/31/85)
One other question - when you got your DN330 upgrade, did you also get new compilers for them? The stock SR9 compilers have some problems on the 020's - I've found bugs in the floating point department, and could easily believe they might miss some optimisations for the DN330's. (The easy way to check is to try "/bin/cc kloo.c -M330" - if it doesn't work you're running old compilers). We got our hardware upgrades several weeks ago but had to nag the field office for the compiler upgrades. -------
apollo@ucbvax.UUCP (01/03/86)
Not all M68020's are alike. Apollo is using the 12.5 MHz version of the 68020. Sun is shipping the 16.67 MHz version. Sun put a lot of special wrinkles into the Sun 3. Your numbers are consistent with numbers they claim as well. There are many other games you can play too. It is not just the processor. I understand they played some games with their clock to avoid wait states. Right now, the Sun 3 is beating anything Apollo has. The ball is now in Apollo's court to see what they will come up with. Bob Clayton GE Calma R&D, San Diego