[comp.sys.acorn] ARM-3 benchmarks

as@prg.ox.ac.uk (Andrew Stevens) (05/08/91)

Dave Redman obtains the following results for Dhrystone (2.1?)
benchmarks.

>       Machine         Dhyrstones per sec
>        ARM2            5289
>        ARM3 @ 25Mhz    14883
>        SUN3/60         2888
>        SUN/Spark1      8000
>        SUN/Spark2      16000

The ARM results are exactly in line with what I've observed but
I am *amazed* at the low figures for the SUN machines.   I am
sure the SS1 does much more than 8000 dhrystones.  I think 18000
is much nearer the mark - but I don't recall the last benchmark
postings. Certainly I don't believe 8000 - the thing clocks at
20Mhz and has a fair sized cache.  Furthermore, I have never known
it to run *anything* slower than my 30Mhz ARM3 A410.  In fact things
typically seem to run around 50 per cent faster (unless you do
lots of 32x32 integer multiplies :-) ). I guess this is
probably due to faster main memory.   The SS2 is approximately
twice as fast again, an ARM-3 A4x0 simply isn't in the same ballpark.
The SUN 3/60 (25Mhz 68020) when I observe it chugs along somewhere between
ARM2 and ARM3 A4x0's. 2900 dhrystones is more SUN3/50 or Macintosh-II
territory.

Can anyone shed any light on Dave's mysterious results?




		
        Andrew Stevens                  
      Programmming Research Group       JANET: Andrew.Stevens@uk.ac.oxford.prg         
 Oxford University Computing Laboratory INTERNET: Andrew.Stevens@prg.ox.ac.uk
     11 Keble Road, Oxford, England     UUCP:  ...!uunet!mcvax!ukc!ox-prg!as
     OX1 3QD

nbvs@cl.cam.ac.uk (Nicko van Someren) (05/10/91)

In article <1694@culhua.prg.ox.ac.uk> as@prg.ox.ac.uk (Andrew Stevens) writes:
>
>Dave Redman obtains the following results for Dhrystone (2.1?)
>benchmarks.
>
>>       Machine         Dhyrstones per sec
>>        ARM2            5289
>>        ARM3 @ 25Mhz    14883
>>        SUN3/60         2888
>>        SUN/Spark1      8000
>>        SUN/Spark2      16000
>
>The ARM results are exactly in line with what I've observed but
>I am *amazed* at the low figures for the SUN machines.   I am
>sure the SS1 does much more than 8000 dhrystones.  I think 18000
>is much nearer the mark - but I don't recall the last benchmark
>postings. Certainly I don't believe 8000 - the thing clocks at
>20Mhz and has a fair sized cache.  Furthermore, I have never known
>it to run *anything* slower than my 30Mhz ARM3 A410.  In fact things
>typically seem to run around 50 per cent faster (unless you do
>lots of 32x32 integer multiplies :-) ). I guess this is
>probably due to faster main memory.   The SS2 is approximately
>twice as fast again, an ARM-3 A4x0 simply isn't in the same ballpark.
>The SUN 3/60 (25Mhz 68020) when I observe it chugs along somewhere between
>ARM2 and ARM3 A4x0's. 2900 dhrystones is more SUN3/50 or Macintosh-II
>territory.
>
>Can anyone shed any light on Dave's mysterious results?

I went and tried it.  The above figures seem about right if you use the SUN
standard compiler but if you use gcc the SUN goes lots faster (50%).  Roll
on gcc for the Archi.!


+-----------------------------------------------------------------------------+
| Nicko van Someren, nbvs@cl.cam.ac.uk, (44) 223 358707 or (44) 860 498903    |
+-----------------------------------------------------------------------------+

ecwu61@castle.ed.ac.uk (R Renwick) (05/10/91)

In article <1694@culhua.prg.ox.ac.uk> as@prg.ox.ac.uk (Andrew Stevens) writes:
>
>Dave Redman obtains the following results for Dhrystone (2.1?)
>benchmarks.
>
>>       Machine         Dhyrstones per sec
>>        ARM2            5289
>>        ARM3 @ 25Mhz    14883
>>        SUN3/60         2888
>>        SUN/Spark1      8000
>>        SUN/Spark2      16000
>
>The ARM results are exactly in line with what I've observed but
>I am *amazed* at the low figures for the SUN machines.   I am
>sure the SS1 does much more than 8000 dhrystones.  I think 18000
>is much nearer the mark - but I don't recall the last benchmark
>postings. Certainly I don't believe 8000 - the thing clocks at
>20Mhz and has a fair sized cache.  ......

	Well here are the sun figures I have found using the same
benchmarks as Dave.

	Machine		Dhrystones/sec	  
	 Sparc SLC	  10000
	 Sparc 2	  25000
	 Sun 3/60	   3333

	Sorry, can't seem to find a sparc 1 to try :-(  I'm not sure
whether SLC's are sparc 1 or 2 but I could find out if pushed.  The
sparc station 2 figures were obtained on a Sparc 2 with 64Mb of ram and
100Mb of local swap space and just a tad more expensive than a 540 :-) 
The sparc 2 also had 8 users on it.  So all in all not a bad performance.
	I'd say that Dave's 8000 for a sparc 1 looks about right to me
judging by the SLC figure I gave.

Rik

news@prg.ox.ac.uk (news) (05/13/91)

My mystification at the SUN vs Arch dhrystone ratings (what
would Freud say about this ? :-) ) abates.

Dave Redman got
>       Machine         Dhyrstones per sec
>       ARM2            5289
>       ARM3 @ 25Mhz    14883
>       SUN3/60         2888
>       SUN/Spark1      8000
>       SUN/Spark2      16000

whilst Rik Renwick obtained

>	Machine		Dhrystones/sec	  
>	 Sparc SLC	  10000
>	 Sparc 2	  25000
>	 Sun 3/60	   3333

It turns out these figures are right (mod deviation due to clock
granularity etc) ... for the benchmark run.  However, as run, the
benchmark is more than a little unfair on the SUNs.  This is because the
benchmark (as given to me by Rik) uses the *default* optimisation level
of the compilers.  On the SUNs this is no optimisation at all, on the
Arch this is full optimisation!  If you use comparable optimisations
(CSE, auto register, loop tweaking = default on the Arch, -O2 on SUN) the
figures are (roughly):

SPARCstation-1/SLC = 14500  (a -2 runs around twice as fast)
SUN-3/60	   = 4000   (gcc squeezes out around 4500)

which is in line with published figures.  The fact my 30Mhz ARM3 A410
runs *my* programs a lot slower than SPARCstation-1's is explained by
the style of program: big, heavy register usage and pointer bashing.
SUN's hardware - SPARC register windows and big (but dumb) cache - does
well on the former two points.  SUN's compiler does well on the latter
because it can optimise pointers (and other stuff) in ways that Acorn's
cannot.  The fact that the machine my memories of run-time related to
was a 1+ (25% faster clock) helped a bit too!


        Andrew Stevens                  
      Programmming Research Group       JANET: Andrew.Stevens@uk.ac.oxford.prg         
 Oxford University Computing Laboratory INTERNET: Andrew.Stevens@prg.ox.ac.uk
     11 Keble Road, Oxford, England     UUCP:  ...!uunet!mcvax!ukc!ox-prg!as
     OX1 3QD

jpff@maths.bath.ac.uk (John ffitch) (05/13/91)

If you use the Norcroft C compiler on the SPARC you get even better
Dhrystone figures that gcc with optimisation.  The SUN compilers are
really not very good.  Also with Norcroft you get real ANSI.

==John

PS I am biased

gcwillia@daisy.waterloo.edu (Graeme Williams) (05/14/91)

In article <1991May13.145154.21732@maths.bath.ac.uk> jpff@maths.bath.ac.uk (John ffitch) writes:
>
>If you use the Norcroft C compiler on the SPARC you get even better
>Dhrystone figures that gcc with optimisation.  The SUN compilers are
>really not very good.  Also with Norcroft you get real ANSI.

OK I have to ask this - how do prices for Suns and ARM3 machines compare??

Graeme Williams
gcwillia@daisy.waterloo.edu

nbvs@cl.cam.ac.uk (Nicko van Someren) (05/14/91)

In article <1991May13.192221.22184@watdragon.waterloo.edu> gcwillia@daisy.waterloo.edu (Graeme Williams) writes:
>OK I have to ask this - how do prices for Suns and ARM3 machines compare??

I think that the cheepest Sparc IPC, which has about the same hardware spec
as an R260 but has a 17" mono monitor and NO SunOS (pay more for it) is about
5650 quid, or was last time I looked.  To have SunOS you need to pay an extra
250 for the CD ROM, 450 for the manuals and 650 for the CD ROM drive.  Then
you need to spend 300 on a thin-ether drop box if you don't want to pay more
to have thick-ether wiring.  That looks like 7300 quid to me, compared to 4000
for an R260 with an extra 300 for the on site service that Acorn make you get.
Note that the R260 is alot faster than an R140 with an ARM3 as the RAM and disk
speed make a great deal of differenc.



+-----------------------------------------------------------------------------+
| Nicko van Someren, nbvs@cl.cam.ac.uk, (44) 223 358707 or (44) 860 498903    |
+-----------------------------------------------------------------------------+