I have a dual boot Athlon2000+ with 512MB RAM. When I benchmark under Win2K i get 1396 Whetstones & 2347 Dhrystones, but under Kubuntu 5.10 I get 767 Whetstones & 1330 Dhrystones. Does anyone know if I have some sort of Linux installation/slowdown problem, or is this speed difference to be expected.

---

I have a dual boot Athlon2000+ with 512MB RAM. When I benchmark under Win2K i get 1396 Whetstones & 2347 Dhrystones, but under Kubuntu 5.10 I get 767 Whetstones & 1330 Dhrystones. Does anyone know if I have some sort of Linux installation/slowdown problem, or is this speed difference to be expected.

It\'s a Linux BOINC client vs. Windows BOINC client thing. Everyone sees it unless they compile their own client. As for the benchmarks, they matter not at all to climateprediction and are not a good predictor of climateprediction performance.

But if the pure CPU benchmarks are not relevant to CPDN, why do I also have big discrepancies in the \"to completion\" time, namely approx 2700hrs under Kubuntu but only 1492hrs under Win2K ?

Are these \"to completion\" times going to be \"not a good predictor of climateprediction performance\" or should I believe them ?

(The reason for all this is to work out if I can escape from using Windows without getting a massive CPDN performance hit in terms of WUs completed and gained credits)

---

The \"to completion\" times are worked out by BOINC.
BOINC just isn\'t very good at it for cpdn.
However, the newest 5.* versions will, apparently, gradually learn the correct amount, but only after fully processing several models.
It learns with all the projects; cpdn just takes a bit longer.

The best way to get a good time estimate, is to DIY. Use the cpu time, and look at the number of years processed on the graphics, then do some maths.

One more thing. There are also some \"curves\" in the actual prediction of competion time (which usually for me for CPDN drifts up till about 50% done ...).

There was a formula change in the 5.2x versions to be closer, but, current evidence suggests that this one is low in early days, improving, then over-shoots and then gradually corrects. Since CPDN work takes forever, well, the numbers are not highly accurate. They are an estimate, and on my systems slabs take about a month of CPU time with the Sulfurs taking about 90 days (13 more days before I complete my first).

=== Edit

Les, the DCF is only going to be used to apportion the work estimate when fetching new work. But does not affect the \"To Completion\" numbers.

More interesting is the fact that BOINC View and 5.2.13 can show different numbers or so a post on the BOINC View forums said ...

---

Paul
Thanks for the info.

The best way to get a good time estimate, is to DIY. Use the cpu time, and look at the number of years processed on the graphics, then do some maths.

Thanks to Les & Paul for help so far. I\'m going to manually time how many steps occur over a few minutes for each OS - This should help me make the ultimate Win2k of Kubuntu decision. I\'ll post my results when I\'ve done them.

---

[quote
I\'m going to manually time how many steps occur over a few minutes for each [/quote]

Benchmarking of two Sulphur WUs, both of which are at 1.xx% completed (with graphics turned off as only the Win2K machine has got OpenGL nVidia drivers installed):-
Under Win2k, 74 TS in 7 min, indicated 5.84 s/TS
Under Kubuntu, 59 TS in 7 min, indicated 6.65 s/TS

So it would seem that (on these 2 particular WUs):-
1 - BOINC has a lot of \"time prediction learning\" yet to do before it gets its estimates anywhere near correct.
2 - Win2K is faster than Linux by between 12% and 25% depending on whether I believe the TS over 7 minutes, or the displayed s/TS

Therefore regarding my \"Win2K or Kubuntu decision\", I think I\'m going to have to stick with Win2K if I want the best speed :-(

.....unless of course someone tells me that I have missed some important factor, or some tweak, or some optimised client for Linux.


(Les / Paul, many thanks for your help)

---

5.2.13 has a fix in the \'to completion\' estimates where the importance of the initial estimate * DCF is reduced more as the result progresses than in previous versions. The importance of CPU time * percent complete rises faster in proportion.

---

BOINC WIKI

BOINCing since 2002/12/8