🚀 go-pugleaf

(This message concentrates on performance, the system architecture and config details
will have to wait for another time...)

Performance numbers for the SPARCcenter 2000 (a.k.a SunDragon) SPARCclassic
and SPARCstation 10 model 41 running Solaris 2.X and SPARCcompilers 3.0alpha.

The numbers were measured using Solaris 2.1 on all these platforms, for
the SPARCclassic and SPARCstation 10 model 41 this is a beta release that
ships in December 92, for the SPARCcenter 2000 this is a development release
and the product will ship next spring with Solaris 2.2. SPARCcompilers 3.0
will be released next year. The SC2000 will initially be supported in
configurations with up to 8 CPU's. With further time for testing the
huge configuration matrix and performance tuning 20 way systems will be
supported towards the end of 1993. 2 CPU systems start under $100K and
an 8 CPU system starts at about $200K.

SPARCcompilers 3.0 include an improved code generator, with similar
performance to the Apogee compilers used for the original SS10 SPEC figures
but combined with a KAP *Parallelising* Fortran front end and a microtasking
parallel fortran runtime harness to give the first quoted SPEC figures
that are measured with a single Unix process running on multiple CPU's.

To make it totally clear, these numbers are *NOT* produced by running two
copies and adding the results, some of the SPEC benchmarks paralellise a bit
and some of them parallelise pretty well. The C benchmarks are run straight
and give single CPU performance levels.

Here are the SPECfp92 results

Name		SPARCclassic	SS10/41		SC2000 with 8 M41 CPU's
013.spice2g6	 17.0		 43.1		 42.0
015.doduc	 17.0		 49.3		 48.7
034.mdljdp2	 24.6		 67.7		 65.8
039.wave5	 13.7		 49.1		 65.5
047.tomcatv	 20.8		 59.3		155.9
048.ora		 31.2		118.5		570.8
052.alvinn	 34.0		147.3		144.0
056.ear		 28.9		 78.3		 77.3
077.mdljsp2	 14.5		 32.6		 32.3
078.swm356	 13.3		 39.0		339.6
089.su2cor	 25.9		 85.7		259.0
090.hydro2d	 22.3		 64.2		234.6
093.nasa7	 27.3		 59.2		 81.5
094.fpppp	 16.9		 71.3		 70.3

geo mean	 21.0		 63.4		109.0

The interesting results are those that show how much parallelism there
is in some SPECfp92 codes. The CPU module is identical in the SS10 and the
SC2000 and is a 40MHz SuperSPARC with 1 Mb external cache. The SPARCClassic
(and the SPARCstationLX) use a 50MHz MicroSPARC.

SPECrate values

Machine	CPU's		SPECrate_int92	SPECrate_fp92
SC2000	8 M41		8047		10600
SC2000	4 M41		4539		 5627
SC2000	2 M41		2326		 2953
S600MP  4 Cypress	1847		 1930
SS10	1 M41		1263		 1544
SClassic 1 MicroSPARC	 626		  498

HP 890	4 HP-PA		4301		 4685
HP 890	2 HP-PA		2253		 2360
HP 897  1 HP7100	1857		 3359

IBM 980 1 RS6000	1404		 2960

For a wider look at scalability there are results for DP Linpack 1000x1000
using the same compiler and measuring with varying numbers of processors.
Again this is for a single parallelised process.

CPU's		DP Linpack 1000x1000 MFLOPS
2			 46.7
4			 91.2
6			130.1
8			159.6
10			189.0
12			210.9
14			224.4
16			249.5
20			268.5


For a multiuser workload there are AIM III results

Machine		CPUs		AIM III jobs/minute
SC2000		 8 M41		 1631
SC2000		 4 M41		 1014
SC2000		 2 M41		  570
S600MP-140	 4 Cypress	  503
SS10		 1 M41		  422
SClassic	 1 MicroSPARC	  148

Pyramid S12/12	12		 1414
Pyramid S12/8	 8		 1112

Sequent S2000	12		 1455
Sequent S2000	 8		 1027

IBM RS6000-970	 1		  489

All this info comes from "SPARCcenter and SPARCserver Performance Brief"
published November 1992. Any typo's are mine. SPEC figures may change
slightly before being published in the SPEC newsletter.

Regards Adrian

---
Adrian Cockcroft -  adrian.cockcroft@uk.sun.com or adrian.cockcroft@sun.co.uk
Sun Microsystems, 306 Science Park, Milton Road, Cambridge CB4 4WG, UK
Phone +44 223 420421 - Fax +44 223 420257	Sun Mailstop: ECBG01