Climate Prediction

[Rocoll]

Originally Posted by kasi

Yes could be the heat, but I think my error may have been caused by using Honza's instructions for changing from Version 4.19 to Version 4.22. It worked fine and fast for the actual processing but mucked up after sending a trickle. I thought I would be OK because I had a backup of the whole BOINC directory but once the server had marked the model as being in error the backup was of no use. I didn't only change for the extra speed but also for the challenge of seeing if I could do it and curiosity about how much faster the new version would be on my machine. Curiosity killed the kasi.
Oh well, I'm sure I'm not the only one getting errors even if this time mine was probably self inflicted. I would really like to complete some units though because so far I've only had one that completed successfully.

An A64 optimisation code may be the answer?
I think I know where I can get one.
Better check it out by a trickle though as we don't want another unit to abort do now do we?

[kasi]

This new 4.22 one has been OK for 8 trickles, so fingers crossed for the next 102. Nice improvement in the speed too, down from 2.97 sec/TS to 2.57 sec/TS.

It occurred to me that making a backup of these models is pretty useless unless you run with "Network activity suspended" and only allow contact with the server at trickle time. A bit inconvenient for me to do that at the moment though because my monitor is mucking up and requires a lot of swearing and banging on the side to get working. I must get it fixed one day but I'm reluctant to carry it anywhere because it weighs so much.

A64 optimisation code? Not sure what you are referring to there. No sense in hacking to bypass "Genuine Intel" CPU check because Version 4.22 has the "-ax" optimisations (SSE2) turned off in the Intel compiler. Although turning SSE2 back on may result in speed demon processing, the reason it was turned off was that using it resulted in instability. Not sure if that's correct, it's only my understanding of what I read.

[kasi]

Welcome to new team member aNt.

[Rocoll]

Originally Posted by kasi

This new 4.22 one has been OK for 8 trickles, so fingers crossed for the next 102. Nice improvement in the speed too, down from 2.97 sec/TS to 2.57 sec/TS.

It occurred to me that making a backup of these models is pretty useless unless you run with "Network activity suspended" and only allow contact with the server at trickle time. A bit inconvenient for me to do that at the moment though because my monitor is mucking up and requires a lot of swearing and banging on the side to get working. I must get it fixed one day but I'm reluctant to carry it anywhere because it weighs so much.

A64 optimisation code? Not sure what you are referring to there. No sense in hacking to bypass "Genuine Intel" CPU check because Version 4.22 has the "-ax" optimisations (SSE2) turned off in the Intel compiler. Although turning SSE2 back on may result in speed demon processing, the reason it was turned off was that using it resulted in instability. Not sure if that's correct, it's only my understnading of what I read.

Yes I'm referring to the SSE2 optimising for A64 CPUs.
I've got a mate running the optimiser on his A64.
So far so good. Works on the 4.22 client and looks like a 20% improvement.
He had his first trickle today and its still running without any probs.
It also works on other BOINC projects.
He's running SETI too.

I've got P4 and P3 CPUs and I don't think its a good idea to use it on them.

[kasi]

Zounds, 20% faster than Version 4.22, that patch is smokin'! What's the total speed up over Version 4.19?
After my little accident when changing from 4.19 to 4.22 in mid model, I'll stay with what I have for now and try and complete the model.
However I'm interested in how your friend progresses and the specs of his A64 and sec/TS figure. I suppose it's OK to post about this stuff, without the Intel Police knocking on the door?

[Rocoll]

Originally Posted by kasi

Zounds, 20% faster than Version 4.22, that patch is smokin'! What's the total speed up over Version 4.19?
After my little accident when changing from 4.19 to 4.22 in mid model, I'll stay with what I have for now and try and complete the model.
However I'm interested in how your friend progresses and the specs of his A64 and sec/TS figure. I suppose it's OK to post about this stuff, without the Intel Police knocking on the door?

Not sure how it would work on a 4.19 as it was released on the same day the 4.22 was released. I think it was written specifically for the 4.22 version to improve the lagging A64 crunchers?
I was particularly interested to see if it would trickle without failure after your problem.
I don't see a problem with optimising from Intel, afterall they developed the SSE2 instruction so it would work better with compatable applications and it does. If there's a problem it would be for AMD.
SETI has had optimising options for nearly a year now and that's been accepted by most SETI crunchers and I don't see reports in the forums to say it presents any problems for SETI except for too many units being returned for the servers to handle.
100% improvements in SETI completion times are not uncommon I've heard.

[RnR]

I guess its all over the news to some extent... here is a good rundown...

2005 Continues the Warming Trend.

[bundy33op]

Does anyone know if it is possible to request the smaller hasdsm units? I started on a sulphur unit but will take more than 6 months to complete, 0.39% per day its going to take a long time assuming no corruption or any other error

My celery 1300 ain't the fastest machine anymore and needs a little help in this area

Bundy33op

[kasi]

The smaller HadSM3 Slab models are currently unavailable. To enable the coupled model to start next year a large number of Sulphur models need to be completed first. Therefore the Sulphur models have been given the highest priority and Slab models are not currently being assigned.

I understand that this means a very long completion time for those like yourself do not have a faster computer, but as long as the model is completed before the deadline in 2006 then the information gained is still useful.

You are not alone, the current unavailability of Slab models is causing some problems for other project contributors with less speedy computers. You can read about this here.

[Rocoll]

Originally Posted by Rocoll

Yes I'm referring to the SSE2 optimising for A64 CPUs.
I've got a mate running the optimiser on his A64.
So far so good. Works on the 4.22 client and looks like a 20% improvement.
He had his first trickle today and its still running without any probs.
It also works on other BOINC projects.
He's running SETI too.

I've got P4 and P3 CPUs and I don't think its a good idea to use it on them.

Well that didn't work.
The optimiser recalculated the time to completion and promised a significant improvement with the completion time.
It didn't work. Every way you look at it the times for each trickle didn't alter.
The time to completion counted upwards instead of downwards.
O well at least it didn't wreck the workunit.
I'm guessing 4.22 is already optimised for SSE2?

[kasi]

Yes you can't tell much from the BOINC estimate of the time to completion, because I've noticed it varies between different BOINC versions, sometimes it underestimates by a large amount and so continues to increase for a long time as the model progresses and sometimes the opposite. I only ever use the sec/TS figure. The sec/TS figure is an average of the work completed though, so if I have made any changes and want to see if there is any change in speed I use the time between trickles as an approximate indication of performance.

The spinup models were reported to be using version 9 of the Intel compiler with the -ax optimisations (SSE2) turned off and I assumed that version 4.22 of the Sulphur model was also. The increased performance of 4.22 was said to be due to the improvements in ver. 9 of the Intel compiler even with SSE2 turned off. This would be consistent with AMD machines getting a much larger increase in performance than Intel machines when doing 4.22 models.

It was an interesting experiment even if it didn't work out. Some other projects benefit from optimisations to the BOINC program itself because of the way the scores are computed but for Climate Prediction the points allocated per model are fixed and the BOINC version used makes no difference as it is only used as the interface to the climate prediction program that does the actual crunching. This talk of A64 optimisations prompted me to read a bit about perl scripts and I have downloaded Active Perl to have a bit of a play. Good to exercise the brain with something new every now and again.

[bundy33op]

I want to move the cpdn experiment to a new pc, and the below seems to be the best info i can find on it, direct from the climate prediction website:

Creating & restoring from a backup

Ratopher wrote:
I have a short question concerning the THC-upgrade. I want to save my current experiment, because I want to be able to restore it if the upload of the data fails at the end of the experiment (since you and others have described this problem, and it is not solved yet, as I understand).

It is always a good idea to backup the \climate prediction folder before applying any upgrade just in case the install process screws things up completely - it does happen occasionally.

Quote:
So, which subfolders of the CPDN-folder do I have to save? Is that all of them in the main folder except the archive data of the old experiment, or can I exclude some of them?

I store the contents of each \climate prediction\archive\experiment_x folder elsewhere - that reduces the maximum size of the whole \climate prediction folder to a more maneagable size for speedy & easy regular backups. If I were you, I would do exactly the same rather than trying to 'slice up' the \climate prediction folder's contents - ie.:

1. Edit / select all / cut & paste the whole contents of each \archive\experiment_x folder eleswhere, eg. I use x:\CPDN Archive\experiment_x - run ID#xxxxxx
2. Create a backup folder, eg. I use x:\CPDN Backup
3. Copy the complete \climate prediction folder into the backup folder.

Note: I also do (1) & (3) at the same time whenever I create a new backup copy, so that my separate archive always matches the state of my most recent backup copy. The current archive only changes at phase transition - so there's nothing to do mid-phase when you make a backup.

To restore from a backup, all you would have to do is:
4. Delete the original \climate prediction folder eg. c:\program files\climate prediction
5. Copy x:\CPDN Backup\climate prediction back into the same place as in (4) & restart it - there's no need to do anything else.

Is there anything else i need to know?

Current WU is a sulphur 4.22 @ 6.2% complete

Bundy33op

[kasi]

I think the backup and restore strategy that you described is used to recover from an error. If you want to transfer CPDN from one machine to another then you can just copy the whole BOINC folder from one machine to the other and it should work OK as long as the operating system is the same (or similar, Carl says Win2K>WinXP is OK). This is according to this thread here: http://www.climateprediction.net/boa...pic.php?t=2368
Good luck and I hope your new machine is a bit faster than your Celeron 1300.

[kasi]

Welcome to new team members GOC and Rob.

[bundy33op]

Thanks Kasi, that is all one has to do, install boinc on new machine, do not attach to project, and then overwrite the new boinc folder with the old folder.

The Athlon xp-m 2600 is a wee bit faster than the celeron 1300 :-)

Bundy33op