I'm running my first beta model now and it's reporting 3.25 seconds per timestep ( 18.5 ts/min ), which is slightly faster than the old model used to give me of about 15-16 ts/min.

Great, but the odd thing is, I'm still running the old model as well. According to task manager, they are sharing the single CPU almost equally. According to CPFarmView, the old model is running about 8 seconds per timestep which is half it's previous speed as I'd expect.

But why is the new model running so fast?

Between them I'm now getting 25.7 ts/min which must be a record for a standard 2.4GHz Desktop.

PS. Exiting the old model doesn't make the new model run any faster

From the behaviour and performance it looks like a real Dual CPU machine :-)
I'm getting 26 ts/min with on 3.2GHz P4 or AMD64 3000+

I think it's because sec/ts calculated from CPU time and timesteps of each model. I got about 1.8 sec/ts during alpha testing and i guess that if i haven't reseted the project and continue running old and new one, the behaviour would be the same...
I would also expect both your models decreasing it's speed, resp. figured going down.

Yep, same performance improvement for me, although I’ve halted my V 3.0.1 model for the time being. My 2.6GHz Celeron Laptop has increased from ~15 TS/Min to ~20 TS/Min under BOINC. Maybe the FORTRAN compiler is providing some performance improvement, as per http://www.climateprediction.net/board/viewtopic.php?t=1237. I thought I recalled carl saying something like a~25% performance improvement, but I can’t seem to find this commentary.

Isn't it true that the OS is time-slicing the two models on a single CPU? In which case, the models would behave, internally, as though they had the world to themselves -- and they would, while they were allowed to run by the OS. Hence, the TS should resemble stand-alone operation. (Some loss to additional overhead, though.)

Dual CPUs or P4 HT virtual CPUs will each have separate runs, and for a P4, it will show in elongated TS. But, as Honza noted, expect at least 15% improvement on throughput compared to two sequential runs. (According to manual calculations in Alpha, and what may be an erroneous assumption, I had 29% & 38% throughput improvement over 'classic' CPDN -- which makes my assumption suspect. [P4 3.0 & P4 2.8, respectively, on SuSE Linux 9.0.])

Edit: Besides a supposedly better Fortran Compiler for M$ & Linux, Carl makes fewer disk writes, which should also improve performance.

Cheers.
________________________________________________
Washing one's hands of the conflict between the powerful and the powerless means to side with the powerful, not to be neutral.
-- Paulo Freire (1921-1997), educator, author.

Hi Martin,

I guess it may be confusing since the "old CPDN" was "wall clock" time, and with BOINC I am now using actual CPU time? So even if you have 50 things running on your PC, the actual CPU time from one timestep to the next is an average of 3.8 seconds CPU time or whatever; but in "real time" that may be 10 seconds between timesteps! So exiting the old model doesn't change your CPDN/BOINC timestep timings as it still is 3.8 "CPU-Time" seconds per timestep

It should allow us to give more accurate comparisons of different computer configurations from trickle information etc (especially since BOINC's code for machine types is a lot better than the old CPDN common designation for CPU type as "AMD or Intel Processor" ;-)

> From the behaviour and performance it looks like a real Dual CPU machine :-)
> I'm getting 26 ts/min with on 3.2GHz P4 or AMD64 3000+
>
> I think it's because sec/ts calculated from CPU time and timesteps of each
> model. I got about 1.8 sec/ts during alpha testing and i guess that if i
> haven't reseted the project and continue running old and new one, the
> behaviour would be the same...
> I would also expect both your models decreasing it's speed, resp. figured
> going down.
>
>

I get 26 TS/ min with a AMD 64 3200+ in 32 Bit mode. That's less than in CPDN client version 3.0.0.1. There i had 30/ min

Founder of

<a href="http://climateapps2.oucs.ox.ac.uk/cpdnboinc/team_display.php?teamid=36"><img src="http://www.ug-abi.de/2008/projekt1.jpg"></a>