Just joined the project a few days ago and received my first 2 WUs. 43 hours done, roughly 3100 hours to go (14,4 s/TS).

In MPEG-2-Encoding my "thingie" here performs roughly like a P4/1800. Therefore I can hardly believe that a single P4/2400 or P4/3000 ist almost 10 times faster than my machine.

Most of the time there was nothing running except the 2 climateprediction-WUs via BOINC 4.19.

Screensaver ist set to blank.

Regards Gerhard.

What is the speed of your processors? Is it a dual P3 of some sort? 100 or 133 front side bus and memory?

Yes it's a dual P3/1000 with 512 MB RAM (PC 133), CPU peed ist 1000 MHz and FSB speed is 133 MHz.

Something must be wrong here.

For example:

host# 105333 is a single Celeron at 2400 MHz. It gives 1221.6 million ops/s and 3644.9 million ops/s and needs 3.2 s/TS

host# 110294 (dual P3@1000 MHz; my machine) gives 883.84 millon ops/s and 2168 million ops/s and needs 14.4 s/TS.

I don't untderstand, how 50% more CPU-Power generates 450% faster results.

Looking at one of your machines (host# 79505) I get really confused. This machine ist slower (MIPS) than mine and produces much faster results, how can that be?

There must be a bottleneck here or I made a mistake setting up BOINC and CP.

Is ist possible tu run one WU on both CPUs?
At the moment there are two workunits running.

Regards Gerhard

> Something must be wrong here.
>
I think so.

> Looking at one of your machines (host# 79505) I get really confused. This
> machine ist slower (MIPS) than mine and produces much faster results, how can
> that be?
>
The benchmarks for BOINC, are meaningless in measuring CPDN performance. In fact, the difference in benchmark scores between the same processor in BOINC for Linux and Windows is ridiculous. For my P4, the Linux benchmark sucks compared to the Windows one, but CPDN processes faster in Linux. The benchmark is definitely a problem with BOINC, which is well known.

> Is ist possible tu run one WU on both CPUs?
> At the moment there are two workunits running.
>
No, the code isn't parallelized, but eventually they may have to with dual core processors becoming more numerous over the next couple years.

Yes, you definitely have a problem. With a single 800 MHz Duron with 100 MHz front size bus and PC133 memory, I was getting 9 sec/TS. With a 1 GHz P3 (and BOINC CPDN is optimized for Intel processors), you should definitely be getting better than that. I would hazard a guess in the 6 to 8 sec range but I could be off. You may want to limit yourself to run one model (you can change in preferences) and see what your trickle durations are.

With a dual P3 1Ghz you should be running two models and averaging around 9 or 10 s/TS per model. This is what I was getting on a pair of D6VAA's . Unfortunately there seems to be a penalty of about 25% per processor when running dual PIII's .

---

[information on benchmarks]

Hmmm, I hope that the other parts of the software are more reliable than the benchmarks. ;-)

> > Is ist possible tu run one WU on both CPUs?
> > At the moment there are two workunits running.
> >
> No, the code isn't parallelized, but eventually they may have to with dual
> core processors becoming more numerous over the next couple years.

Quite surprising because there are already lots of hyperthreading CPUs (poor mans SMP) out in the wild and they are still spreading. intel gave the SMP-supporting software a real push with the introduction of hyperthreading.

> Yes, you definitely have a problem. With a single 800 MHz Duron with 100 MHz
> front size bus and PC133 memory, I was getting 9 sec/TS. With a 1 GHz P3 (and
> BOINC CPDN is optimized for Intel processors), you should definitely be
> getting better than that. I would hazard a guess in the 6 to 8 sec range but
> I could be off. You may want to limit yourself to run one model (you can
> change in preferences) and see what your trickle durations are.

Running only one model accelerated it a little bit (13.5 s/ZS) and after resuming the second model (now both are running) the speed didn't decrease.

CPDN seems to need a lot of memory bandwith and the memory interface of the VIA 694X (chipset of my board) is quite slow.

I will leave it as it is and keep on running two models.

In a few weeks time I will have a new machine (dual Xeon2667, 2 GiB Memory) which shold be a "little" faster (I expect 2,5x - 3x). Then I will try to transfer (should be possible) the models to the new machine and see what happens to the speed of CPDN.

Regards Gerhard

> I will try to transfer (should be possible) the models to the new machine and see what happens to the speed of CPDN.
>

Should be possible. You will need to transfer the whole BOINC folder and reinstall BOINC over the top. But beware if you ever have to use CD/DVD to do a transfer, as the read only flag will have to unset.

By then BOINC 4.2x may be released and we are not yet sure how upgrade to that version will be done.

---

[Transferring BOINC an CPDN-models]

> Should be possible. You will need to transfer the whole BOINC folder and
> reinstall BOINC over the top.

I will try it in this manner. The new machine will use the old harddisks so no transferring is necessary. Just reinstall BOINC.

> By then BOINC 4.2x may be released and we are not yet sure how upgrade to that
> version will be done.

I'm no version junkie. :-)

Regards Gerhard.

> ...
> In a few weeks time I will have a new machine (dual Xeon2667, 2 GiB Memory)
> which shold be a "little" faster (I expect 2,5x - 3x). Then I will try to
> transfer (should be possible) the models to the new machine and see what
> happens to the speed of CPDN.
> ...


If you want to use the full power of those Xeons, use them with HT. Of course your models will loose some speed from that but you will be able to run 4 at once.
____

The P3/1000 : a penalty of about 25% (rather a bit more on Coppermine with only 256k cache) is normal for those systems, a big CPU cache can fix some of it but there's still some loss compared to a single system. Upgrading to Tualatin P3s would help some as those have 512k cache, but still it would not be twice the speed of a single CPU system.

If you want to keep the Coppermines, you could try to use a different project on one of the CPUs, my P3/1266 has much better trickle times running only one CPDN and something else on the other CPU.

...
> > In a few weeks time I will have a new machine (dual Xeon2667, 2 GiB
> > Memory) which shold be a "little" faster (I expect 2,5x - 3x).
...

> If you want to use the full power of those Xeons, use them with HT.

I'm still considering the upgrade from w2k to WinXP. The activation "feature" gives me little headache because until now I found no statement from Microsoft what will happen when the product "lifetime" is over. When I buy a product then I want to have the decision how long I use it.
____
> If you want to keep the Coppermines, you could try to use a different project
> on one of the CPUs,

After changing this things (*) accelerated the models from 14,4 s/TS to 13.3 s/TS (and its still slowly accelerating).

(*) model1 from CPU 0,1 to CPU 0
model2 from CPU 0,1 to CPU 1
Global preferences:
- Write to disk at most every: from 60s to 10s
- Leave applications in memory while preempted? from no to yes

Regards Gerhard

I have a server board with two S603 Xeon 2200 / 1G RIMM here but not running yet, my problem is the OS too (plus they need a less noisy cooler). I haven't made a decision yet but afaik. neither Win2k WS nor XP WS will support the 4 virtual CPUs - so I currently try to catch a cheap Win2k server (SBS) licence at EBay. If they refuse to leave one for me for little money, it will be some NetBSD or Linux I guess.


> ____
> ...
> (*) model1 from CPU 0,1 to CPU 0
> model2 from CPU 0,1 to CPU 1

I will have to try that too, sounds like a good idea, thanks :-)

> I have a server board with two S603 Xeon 2200 / 1G RIMM here but not running
> yet, my problem is the OS too (plus they need a less noisy cooler). I haven't
> made a decision yet but afaik. neither Win2k WS nor XP WS will support the 4
> virtual CPUs

Windows XP WS supports two physical CPUs with Hyperthreading. It will use all 4 logical processors with their full potential. There is no need for purchasing a server version of Windows XP (written somewhere on microsoft.com).

- so I currently try to catch a cheap Win2k server (SBS) licence
> at EBay.

IMO no good idea because W2k doesn't support Hyperthreading at all. It will treat the 4 logical CPUs like physical ones. This ist often slower than 2 Xeons with HT disabled.
...
[affinity]

It works here. The last 8 hours the models accelerated slightely from 13.3 s/ST to 13.2 s/TS. I will complete them with tis settings ("only" 2700 hours left).


Regards Gerhard

Thanks for those hints. I tried RedHat 7 now but the kernel seems to be too old, it shows only 2 CPUs *sigh* no Halloween Linux for me ;-)

I will receive a 2003 server trial version very soon so I can check how it handles the board (it's a "slightly" outdated IWill DP 400).

Beeing busy with a SETI Classic Gauntlet (probably the lastone) for 3 weeks I will not test CPDN on it for some time though.