[original message edited out by CC -- "alternative" UM isn't so great]

Seems to be running OK on 2GHz laptop/winxp sp1. You said you rebuilt the linux one, will you be releasing it?

cheers
reiteration.net networks

Hi, Carl,

Running okay on P4 3.0, WinXP, Dbox.

No problems on the M$ side.

Jim
________________________________________________
Video meliora, proboque; Deteriora sequor
I see the better way, and approve it; I follow the worse
-- Ovid (43BC-17AD)

Running okay on an AMD 2.6GHz XP-M - WinXP SP2 - Emma. 1.95 Secs/timestep looks exactly the same at the moment but she's only been running the new Beta a few minutes...

<a href="http://www.nmvs.dsl.pipex.com/"><img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=6&amp;team=off&amp;trans=off"></a>

I have made a new build (reflected in the original) post that uses P3 optimizations exclusively, so the Intel "trick" of having AMD's use generic IA32 shouldn't work. Since PCs running CPDN should be P3 or higher anyway, it should be OK, there are a few people running P1's &amp; 2's but it's probably not a good idea since it would take a year to finish a model!

I see pretty dramatic performance increases with this build (3 sec to 2.5 sec on a Pentium4 on Linux; 2.5 seconds to 2.2 seconds on my AMD64 in Windows); so hopefully it doesn't mess up the model calcs.

If anyone is near a "phase change" (i.e. near 33.33%, 66.66%, or completion) and is trying out this optimized UM please let me know, as I would like to get the *.gmts.* and *.rmts.* files in your dataout dir to see if the calcs are sensible.

Running fine on an Intel P4 Mobile 1.6GHz - Longhorn 4074 - EDINBURGH. Because I'm running BOINC as a service, I will get my first result at the next trickle (yep, my computer 'had trickled' 2 minutes ago).

&gt;If anyone is near a "phase change" (i.e. near 33.33%, 66.66%, or completion) &gt;and is trying out this optimized UM please let me know, as I would like to &gt;get the *.gmts.* and *.rmts.* files in your dataout dir to see if the calcs &gt;are sensible.

Must wait few days ... 15.211%

I installed the new one this morning with no problems, but I've just installed the new P3 optimised version and it's running fine but has restarted the model from timestep 0 :-(


<a href="http://www.users.globalnet.co.uk/~sykesm/cpdn.html"><img src="http://www.users.globalnet.co.uk/~sykesm/gfx/sig.jpg"></a>

Grim initial performance result (calculated from full Trickles):

Dbox, P4 3.0, WinXP SP1, was 2.89 &amp; 2.88 sec/TS, now 3.29 &amp; 3.36 sec/TS!

It went from ~8h40m per Trickle, to ~10 hours.

Groan; those numbers are as bad as my Linux numbers were to the good. (BOINC runs stand-alone on this machine; viz is used infrequently and only for seconds at a time.)

I hope this is not a typical CPDNboinc/WinXP result.


________________________________________________
Video meliora, proboque; Deteriora sequor
I see the better way, and approve it; I follow the worse
-- Ovid (43BC-17AD)

wow, thanks for trying it out. I think I'd better just stick with the "tried &amp; true" settings, it's probably too risky at this stage and it seems the gains of one processor are offset by the losses on the other! Plus for all I know it may turn into garbage results, like what happened on the Mac when I tried running every optimization under the sun (it ran fast alright, but everything went to an iceball after a few days).

&gt; wow, thanks for trying it out. I think I'd better just stick with the "tried
&gt; &amp; true" settings, it's probably too risky at this stage and it seems the
&gt; gains of one processor are offset by the losses on the other! Plus for all I
&gt; know it may turn into garbage results, like what happened on the Mac when I
&gt; tried running every optimization under the sun (it ran fast alright, but
&gt; everything went to an iceball after a few days).

Please keep trying though Carl! :-)

<a><img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=18"></a>

Weird - Emma has only slowed from 1.95 to 1.96 secs/timestep but her trickle times are a full hour longer. :-?

<a href="http://www.nmvs.dsl.pipex.com/"><img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=6&amp;team=off&amp;trans=off"></a>

As other P4 owner, no improvement for EDINBURGH. Result of experiment is <a href="//climateapps2.oucs.ox.ac.uk/cpdnboinc/trickle.php?resultid=12677">here</a>. The first eleventh trickles were processed w/o P3 optimizations (3.6750-3.8056 s/TS). Trickle sent 23 Aug 2004 03:21:32 were processed with the tweaked .exe file (3.8128 s/TS).

I will restore original .exe file to continue processing of this WU.

After two trickles on both my machines (P4 2.4 and P4 3.0) neither showed any significant change. Both slowed slightly, but this seems to be a general trend as a model progresses. Both are also in the middle of stage one. Maybe the benefits only appear when the model is run from the beginning using the optimized version? I'm reverting to the standard beta; we'll see if things change.

Andrew

&gt; wow, thanks for trying it out. I think I'd better just stick with the "tried
&gt; &amp; true" settings, it's probably too risky at this stage and it seems the
&gt; gains of one processor are offset by the losses on the other! Plus for all I
&gt; know it may turn into garbage results, like what happened on the Mac when I
&gt; tried running every optimization under the sun (it ran fast alright, but
&gt; everything went to an iceball after a few days).
&gt;

Well, running on an AMD64 3200+, the P3 optimized version speeded it up considerably. From about a 6:09 trickle to ~5:55.

Now why Nick would have seen longer trickle durations with his AMD XP-M, and I see shorter trickle durations for my AMD64 is a puzzler. Perhaps the P3 optimizations allow AMD64 to use SSE2, but that doesn't explain Nick's slowdown.

Now to see if it's stable...

George

&gt; Well, running on an AMD64 3200+, the P3 optimized version speeded it up
&gt; considerably. From about a 6:09 trickle to ~5:55.
&gt;
&gt; Now why Nick would have seen longer trickle durations with his AMD XP-M, and I
&gt; see shorter trickle durations for my AMD64 is a puzzler. Perhaps the P3
&gt; optimizations allow AMD64 to use SSE2, but that doesn't explain Nick's
&gt; slowdown.
&gt;
P3 optimalization and SSE2?
+3% is no big deal when stability is at risk (as reported so far).

I understand Carl's approach to leave optimalization aside before public launch and give a change to some optimalization on sulphur cycle or after public launch.
<img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=56&amp;trans=off">

Honza wrote:
&gt; P3 optimalization and SSE2?
&gt; +3% is no big deal when stability is at risk (as reported so far).
&gt;
&gt; I understand Carl's approach to leave optimalization aside before public
&gt; launch and give a change to some optimalization on sulphur cycle or after
&gt; public launch.
&gt;
Well the P4's are slower with this .exe, but other than Martin's model restarting, I've seen no posts about model instability for the *Windows* client. I certainly understand why one wouldn't want P4's running this version, and why a cautious approach should be taken for the public launch, just reporting what I have seen running it on my system.

&gt; Well the P4's are slower with this .exe, but other than Martin's model
&gt; restarting, I've seen no posts about model instability for the *Windows*
&gt; client. I certainly understand why one wouldn't want P4's running this
&gt; version, and why a cautious approach should be taken for the public launch,
&gt; just reporting what I have seen running it on my system.
&gt;

I would be also happy to crunch faster on my AMD64 :-)
Note that people (astroWX) reported significant (e.g. 3.5 -&gt; 2.9 sec/TS) boots under Linux on P4s or 3.4 -&gt; 2.7 (josti) and some reported no performance change (both under Linux and Win). And, well, some people reported chashes under Linux...

I belivie that our first priority is to have stable model/client so people will join and not leave after while. Then we can try more some code optimatization.
Of course decisions are made by CPDN team - we, as testest, have to test and provide valuable feedback.

Happy modelling...

btw, i would like to see some live number on BOINC project - models years so far, completed models etc.


<img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=56&amp;trans=off">

&gt; btw, i would like to see some live number on BOINC project - models years so
&gt; far, completed models etc.

I would be interested in this as well. :-)

<a><img src="http://boinc.mundayweb.com/cpdn/stats.php?userID=18"></a>