Hello, BOINC fellows !

I try to understand the following:

I have to machines, which produce nearly the same BOINC benchmarks (715/740 whetstones, 1730/1780 drystones),
but computing a CPDN model takes on the first ca. 1600 hours (uff, long time, I know), on the other (that with a little better benchmark results) expects to take ca. 900 hours.

Why are the great differences ? And what can I do to make it faster (please no advices to buy a faster machine ;-))

Grettings, Frank

> Hello, BOINC fellows !
>
> I try to understand the following:
>
> I have to machines, which produce nearly the same BOINC benchmarks (715/740
> whetstones, 1730/1780 drystones),
> but computing a CPDN model takes on the first ca. 1600 hours (uff, long time,
> I know), on the other (that with a little better benchmark results) expects to
> take ca. 900 hours.
>
> Why are the great differences ? And what can I do to make it faster (please no
> advices to buy a faster machine ;-))
>
> Grettings, Frank
>
Hard to know for sure without knowing more about your systems. What are the PCs you are referring to?

The BOINC benchmarks aren't really anything like the CPDN model and are not an indicator of performance for CPDN. If the processors are two of the same type (P4, Celeron, AthlonXP, Athlon64) and same speed, then the difference could be due to motherboard/memory differences. CPDN performance is highly dependent on memory bandwidth and latency. Anything that increases memory bandwidth and/or decreases memory latency while still being stable, will result in better performance.

Otherwise, if performance ratings are approximately the same, and one processor is an Intel and the other AMD, the Intel will likely be faster because of optimizations used in compiling CPDN.

Hello, Geophi !

> Hard to know for sure without knowing more about your systems. What are the
> PCs you are referring to?
Thanks for your reply.

Both Intel Pentium 3 processors, one with 800 MHz, the other 1 GHz.
But I think, the processor isn't the reason. There are complete differences in mainboard, memory etc.
For CPDN I looked at the time/model time step: 3.4 s for pc1, 7.9 s for the other. So it seems to me, that I have to wait a long time to complete the model. Mmh...

Greetings, Frank
>
> The BOINC benchmarks aren't really anything like the CPDN model and are not an
> indicator of performance for CPDN. If the processors are two of the same type
> (P4, Celeron, AthlonXP, Athlon64) and same speed, then the difference could be
> due to motherboard/memory differences. CPDN performance is highly dependent
> on memory bandwidth and latency. Anything that increases memory bandwidth
> and/or decreases memory latency while still being stable, will result in
> better performance.
>
> Otherwise, if performance ratings are approximately the same, and one
> processor is an Intel and the other AMD, the Intel will likely be faster
> because of optimizations used in compiling CPDN.
>

memory speed is a big factor, the MetOffice "Unified Model" used is very memory I/O intensive, so the same exact CPU but with different memory configs (i.e. motherboard & memory bus speed etc) will show vastly different completion times. CPDN also does not use the "official" BOINC benchmarks, credits are based on the trickle points (something like 90 per 10,802 timestep trickle I believe)?

Hello, Carl !

Are you already in Potsdam ?

> completion times. CPDN also does not use the "official" BOINC benchmarks,
> credits are based on the trickle points (something like 90 per 10,802 timestep
> trickle I believe)?

I really don't care about credits, simply try to make the runs faster.

Greetings from Berlin,
Frank

> Are you already in Potsdam ?
Sort of, I came, I saw, I ran away (long story).

>I really don't care about credits, simply try to make the runs faster.

The 3.4s/TS on a P3 is very fast, the 7-8s is probably more typical for a P3 (the complicated model seems to run best on a P4 or AMD64).