Valid XHTML 1.0!


 

(last updates 24. Apr and 9. Jan 2005: Sun Fire v20z/Solaris 10 resp. HP-UX rx2600 added)

Caution: since the second half of April 2005 most machines are running now BOINC applications, from which seti is just one, and so only a few without a BOINC client available are still running the legacy SETI@home client for a limited time.

Since I have run the Berkeley SETI@home software now for a considerable time and on a notable multitude of platforms (several UNIX systems and in some insane MS Windows environments), I want to provide you with the essential facts of these experiences. You can easily see, that the variation in CPU time per work unit is much bigger with the 3.0 client than with the 2.4 client and at average it needs longer — but sometimes despite this it's even quicker! All times given are in hours; the numbers mean minimum, arithmetic mean and maximum respectively.

Because the screensaver and graphical clients waste CPU time with displaying instead of crunching, all listed client number are exclusively text client values — I want to get results quickly and haven't the time nor the desire to watch the familiar SETI@home display regularly.

Concluding remark: due to the tremendous slowdown of the 3.0 client compared to the 2.4 client on the rather old SUN SPARCstation 5 (from 1996) I decided meanwhile, that the additional work no longer pays off the very low throughput; so I use it no more. As you can easily see, the 3.03 client is again much slower, this time generally, but not by a factor 2, as one could expect… The next old computer retired was the RS/6000 from 1996, but since some time ago I could use a new, bored LINUX server with faster CPU for the project too, see below. Meanwhile I use additionally two more SUN computers, see table below, then a new IBM pSeries workstation joined the bundle of active machines, another Linux PC, an older was retired and an HP-UX server as fastest at least temporarily available. At now a new p615 AIX 5L system is clearly on top, though it has only one, but pretty fast CPU. I hope to add another real fast dual CPU machine any time soon...

Finally: the version 3.08 is merely a bugfix release of 3.03 regarding http communication and will not change runtime behaviour of data processing of the 3.03 version.

  Currently active machines are in bold letters below!

platformclient 2.4client 3.0client 3.03–3.08
PIII 733 MHz 7.5 ns RAM / LINUX 2.4.18/20–24 glibc2.2.5 (Debian Woody)not used thennot used then7.5/8.8/9.5
PIII 550 MHz 10 ns RAM / LINUX 2.2.16/19/2.4.17 glibc2.1not used thennot used then10.0/11.8/13.8
PII 450 MHz 10 ns RAM / LINUX 2.4.10/20–24 glibc2.2.4 (SuSE 7.3)not used thennot used then11.6/14.4/16.6
PII 400 MHz 10 ns RAM / LINUX 2.4.18 glibc2.2not used thennot used then12.2/14.5/15.65
PII 333 MHz 10 ns RAM / LINUX 2.2.16 glibc2.110.6/11.4/12.18.8/12.0/14.515.8/19.2/22.0
PII 333 MHz 10 ns RAM / LINUX 2.4.19 to 2.4.23 glibc2.2.5 (Debian Woody)not used thennot used then15.8/20.0/28.7
PII 333 MHz 10 ns RAM / LINUX 2.4.12 to 2.4.17 glibc2.2.4 (SuSE 7.3)not used thennot used then16.75/20.9/26.7
PII 266 MHz 10 ns RAM / LINUX 2.2.14 glibc2.111.8/12.9/13.810.4/14.3/16.918.1/21.8/24.9
Sun Fire V20z AMD 244 Opteron (2x) 1.8 GHz / Solaris 10***not used thennot used then4.0/4.5/4.9 ****
Sun Fire V20z AMD 242 Opteron (2x) 1.6 GHz / Linux 2.4.19 SLES 8 x86***not used thennot used then2.6/3.0/3.45
Sun Fire V20z AMD 242 Opteron (2x) 1.6 GHz / Linux 2.6.5 Suse 9.1 x86_64, x86_64 client***not used thennot used then3.35/3.6/3.9
Sun Fire V20z AMD 242 Opteron (2x) 1.6 GHz / Linux 2.6.5 Suse 9.1 x86_64, i686 client***not used thennot used then2.6/3.25/3.8
Apple PowerMac G4 7455 (2x) 1 GHz 7.5 ns RAM / OS X 10.1.2–5***not used thennot used then5.9/7.735/9.0
Apple PowerMac G4 7455 (2x) 1 GHz 7.5 ns RAM / OS X 10.2.3–8 ***not used thennot used then5.7/7.3/8.6
Apple PowerMac G4 7455 (2x) 1 GHz 7.5 ns RAM / OS X 10.3.1–2 ***not used thennot used then6.0/7.4/8.7
Apple PowerMac G4 7455 (2x) 1 GHz 7.5 ns RAM / Linux 2.4.21 glibc2.3.1 (Mandrake 9.1) ***not used thennot used then6.38/8.0/9.71
IBM RS/6000 Model 850 PowerPC (1996) / AIX 4.233.7/35/36never released!51.5/57.4/61
IBM pSeries (RS/6000) Power 3 375 MHz 44P170 / AIX 4.33not used thennever released!7.4/9.0/10.45
IBM pSeries (RS/6000) Power 4+ 1.2 GHz P615 / AIX 5L 5.2not used thennever released!2.5/3.05/3.6
HP LX3000 PA-RISC (2x) 550 MHz / HP-UX 11.00 ***not used thennot used then6.75/8.05/9.06
HP 9000 model 800 rp2405 PA-RISC (1x) 800 MHz / HP-UX 11.00not used thennot used then5.25/6.2/7.15
HP 9000 model 785 PA-RISC (1x) 500 MHz / HP-UX 11.00not used thennot used then6.5/7.8/9.0
HP model rx2600 Itanium2 (1x) 1 GHz / HP-UX 11.23not used thennot used then1.16/1.42/1.7
SUN SPARCstation 5 (1996)/ Solaris 2.644/47/4957/72/7998/106/115 *
unspecified SUN Solaris **not used thennot used then12.15/14.1/15.45
unspecified SUN Solaris **not used thennot used then9.04/10.8/12.7
PII 266 MHz 10 ns RAM / MS Windows NT 4.012.6/13.6/14.6no data available!no data available!
PII 400 MHz 10 ns RAM / MS Windows 959.8/10.6/10.9no data available!no data available!

 

* only three values taken, they scatter more for sure!

** for some legal reasons I'm not allowed to reveal the exact kind of these machines — I decided to put up the numbers, which are not helpful for benchmarking without it; because you could derive the numbers (the average at least) approximately by my sums and the other active machines!

*** on SMP machines these are too per CPU values: keep in mind, that the throughput scales with the number of processors (only!) then, when at least as many SETI@home client instances are running, as the machine has CPUs…

*** very disappointing numbers, I know, but with a self compiled BOINC SETI client I gained about twice the performance compared to this rather outdated Solaris 2.6 32 bit i686 version

General closing remark: when you use more than one CPU with different mean times for the SETI@home project, than you can calculate your current (keep in mind, that the official value contains OVERFLOW values and often a mixture of different fast client versions!) mean CPU time not with the arithmetic, but with the harmonic mean, presumed you let them run all equal times (usually continuously!): i.e. for four CPUs and mean times t1 to t4 this makes an overall harmonic mean of

th = 4/(1/t1 + 1/t2 + 1/t3 + 1/t4)
and the reason is, that a faster CPU produces more results in equal time and thus weighs stronger, than a slower. Due to this effect the harmonic mean has to be used, which gives lower values than the arithmetic one, i.e. in favor of the faster CPUs you use.
 

back to SETI hints  back to SETI main  back to main

remarks etc. to: stefan.urbat@apastron.lb.shuttle.de

(URL:  http://www.lb.shuttle.de/apastron/setiTab.htm)