Steward Observatory Computer Benchmarks These benchmarks compare a variety of machines seen around Steward Observatory. The actual benchmarks are given in tabular and graphical form. To get benchmark binaries or source code for Windows NT, Linux, or Solaris, contact me. Anybody got a spare DEC Alpha or PIII/550 I can borrow? :-)

This test diagnoses the floating point unit and some integer performance. Here is the rendered test image:

machine:name	CPU ID	OS	rendering time	remarks
Sparc 1:newton	20 MHz Sparc	Solaris	440 sec	(1)
Sparc IPX:loke	40 MHz Sparc	Solaris	220 sec	(1)
Sparc LX:bobafett	30 MHz uSparc	Linux!	174 sec	(2)
Sparc 10:kingsholm	40 MHz SSprc	Solaris	102 sec	(1)
Sparc 20:ethel	50 MHz SSprc	Solaris	72 sec	(1)
HyperSparc:chinadoll	150MHz HSprc	Solaris	39 sec	(1)
UltraSparc:sirius	170 MHz USprc	Solaris	27 sec	(1)
UltraSparc:saguaro	200 MHz USprc	Solaris	22 sec	(1)
UltraSparc:lithops	300 MHz USprc	Solaris	15 sec	(1)
486PC:stupid	66 MHz 486DX2	Linux	186 sec	(2)
486PC:skaro	100 MHz 486DX4	Linux	123 sec	(2)
586PC:syrinx	83 MHz P24T	Linux	99 sec	(2)
586PC:aspc119	133 MHz P5	Linux	42 sec	(2)
686PC:draco	200 MHz P6	Linux	23 sec	(2)
686PC:agave	300 MHz K6-2	Linux	15.5 sec	(2)
686PC:gromit	333 MHz mobile PII	Linux	10.2 sec	(2)
686PC:narya	350 MHz PII	Linux	10.4 sec	(2)
686PC:clavelina	400 MHz Celeron	Linux	10.2 sec	(2)
686PC:loke	500 MHz PIII	Linux	7.5 sec	(2)

Links:

• The POV-Ray Home Page


• Sparc Solaris 2.5.1 binary distribution of POV-Ray 3.01


• x86 Linux (glibc) binary distribution of POV-Ray 3.02


• Compile flags and benchmark instructions

This test diagnoses the floating point unit and some integer performance. Here is the rendered test image:

machine:name	CPU ID	OS	rendering time	remarks
Sparc 1:newton	20 MHz Sparc	Solaris	2480 sec	(1)
Sparc 2:loke	40 MHz Sparc	Solaris	1190 sec	(1)
Sparc LX:bobafett	50 MHz uSparc	Linux!	890 sec	(1)
Sparc10:kingsholm	40 MHz SSprc	Solaris	442 sec	(1)
Sparc20:ethel	50 MHz SSprc	Solaris 2.5.1	319 sec	(1)
HyperSparc:chinadoll	150MHz HSprc	Solaris	180 sec	(1)
UltraSparc:sirius	170MHz USprc	Solaris	130 sec	(1)
UltraSparc:saguaro	200MHz USprc	Solaris	105 sec	(1)
UltraSparc:lithops	300MHz USprc	Solaris	73 sec	(1)
486PC:stupid	66 MHz 486DX2	Linux	1070 sec	(2)
486PC:skaro	100 MHz 486DX4	Linux	642 sec	(2)
586PC:syrinx	83 MHz P24T	Linux	456 sec	(2)
586PC:aspc119	133 MHz P5	Linux	214 sec	(2)
686PC:draco	200 MHz P6	Linux	108 sec	(2)
686PC:agave	300 MHz K6-2	Linux	66 sec	(2)
686PC:gromit	333 MHz mobile PII	Linux	45.8 sec	(2)
686PC:narya	350 MHz PII	Linux	46.4 sec	(2)
686PC:clavelina	400 MHz Celeron	Linux	41.5 sec	(2)
686PC:loke	500 MHz PIII	Linux	32.7 sec	(2)

Links:

• The POV-Ray Home Page


• Sparc Solaris 2.5.1 binary distribution of POV-Ray 3.01


• x86 Linux (glibc) binary distribution of POV-Ray 3.02


• Adopted compilation flags and benchmark instructions

This test diagnoses the floating point unit. Linux machines are at a distinct disadvantage as the CLOUDY F77 code is run through f2c and compiled by gcc. Using g77 will help, but it wasn't available on all machines, and it's far from optimized right now.

Links:

• The Cloudy Home Page


• Sparc Solaris 2.5.1 binary of Cloudy 90.04 (hotfixes to 13 May 1998)


• x86 Linux (glibc) binary of Cloudy 90.04 (hotfixes to 13 May 1998)


• Also required are two text files: parispn.in and c84.ini.


• Compile flags and benchmark instructions

This test diagnoses integer performance. The times are evaluated on the third iteration through LaTeX, when all files are neatly within the I/O cache.

Links:

• LaTeX file, minus figure references


• Benchmark instructions

This test diagnoses integer, floating point, and disk I/O performance. A 1024x1024 image is rotated 45 degrees a total of 8 times, sequentially. The resulting frames are then imcombined. The two steps are tallied separately, as shown.

Links:

• The IRAF Home Page


• 1024x1024 pixel test image


• Compile flags and benchmark instructions

Looking at these benchmark results, it would seem that PC's have caught up with the personal workstation market in terms of power and performance. This becomes extremely enticing when one compares a $5,000 USparc machine with a similarly-equipped K6-2/350 machine that not only matches the Sparc in performance, but costs 1/3 as much!

With flexible 32-bit multitasking, networking, multiuser operating systems like Linux available for PC's, you can do anything you ever wanted to do on a workstation or server-class computer and perform it faster and cheaper on a modern PC.

We have benchmarked our systems with Linux primarily, however one may wish to compare how the most popular PC operating system (DOS) compares to Linux. The answer: almost jokingly. A number of machines capable of dual-booting between DOS and Linux were compared for a few of these benchmarks:

Summary: DOS sucks. Linux rules! But what about OS's like Windows 95 and NT? Benchmarks using these OS's, compiled using Visual C++ 1.1, were in fact 10-20% faster than Linux on the same machine for EVERY test! But some concern about the optimizations was raised when the reported floating point roundoff errors in FLOPS20 were consistently 100+ times worse using Visual C++ than Linux's GCC/EGCS compiler.

The scores for Syrinx are a bit weird, especially if you compare floating-point and integer performance. Under FP, Syrinx ties with a P5-100, but only with a 486-100 for INT manipulations. A single test narrowed down this issue:

Last modified: Sat Nov 20 17:00:04 MST 1999