LHC@home

this msg withdrawn.

We know about this. It is caused by to library functions: errf and exp/log.
They return different results on the two processor types. We have contacted the compiler guys, but they basically say that they do not give more than accuracy than to the 15th decimal. We are hoping that NAG will finish their cross platform compatible libraries soon.

We are at the moment only running some of our studies, because of this problem. Unfortunately most of the work needed done at this time is of this type of work. Which studies the effect of interaction between the two beams going around the ring in opposite direction.


Christian S�ttrup
LHC@home Developer

Greetings Christian.

Sorry, if I caused Your tension to go up a bit, but my alert is withdrawn and I can see from Your post that I was right in my assumption that some code was to blame.

I have now run the exact same wu on 4 of my computers using the following processors: XP1800+, XP2800+, XP3000+ and P3Mobile.

All results match to the bit. :)

(Regrettably I don't own a P4, I am itching to see how much the results differ..)

This at least proves that my computers agree on what result to get. Now I know that I do not have a faulty processor, RAM-Block, Heat-problems, etc.

While I have the 'You on the line':

Are the results useable?

It botheres me to think that my work may be for the garbish-can only. (I hope not :))

If You think it is a good idea, I will post as a new topic, how I went about making the test.

Here is how:
Watch the BOINC work-screen, wait until one WU is just 2 minutes from completing. (Make sure You still have WU's ready to process in Your cache ;))Exit BOINC and ZIP the entire BOINC.DIR (preserve paths) and save it as WU-TEST.ZIP.
De-activate Your Network and restart BOINC and wait for the next complete WU to finish.
Stop BOINC and locate the result-file and make a copy to somewhere safe.
Re-enable Your Network and BOINC starts sending the result.
The saved result is our reference file. Rename to say RESULT-ref.txt.

To test a machine:
Exit BOINC and rename the BOINC-dir to BOINC-ORG.
Unzip WU-TEST-ZIP preserving paths.
De-activate Network and start boinc.
Now the new machine carries on from the frozen moment. In 2 mins it will start crunching the same 'Test' WU.
Wait for it to finish and exit BOINC.
Save the result, naming it something easy to remember, f.eks the name of the machine.
Delete the BOINC dir.
Enable Network.
Rename the BOINC-ORG dir back to BOINC and start Boinc.

Apart from the time used to test, Your work will continue without problems.

Now You can compare the 2 result-files and hopefully they match.

A much better idea is a LHC Approved Testprogram cosisting of a small shell, a copy of sixtrack and a 100,000 turn WU. Include the correct result and the alowed slack.

If users wonder why their results does not get credited or if they think their equipment may be faulty, they can just run the test and see if they have problems.

This test could also serve as a benchmark. (I know that a lot of crunchers want to see just how fast their computer is compared to other participants').

What do You think?

ChrisD

Danish stuff:

Undskyld! Jeg ser f�rst nu at du er dansk, men det jo mest korrekt at skrive p� engelsk.
Jeg er bare nysgerrig, hvilken processor rammer 'best'? Og kan I overhovedet bruge resultaterne fra den 'forkerte' processor?
Du kan svare mig privat, s� vi ikke for�rsager nogen ulykker. :)
Er det ikke muligt at holde de kritiske routiner i ren x86 kode. K�retiden bliver selvf�lgelig l�ngere, men I f�r nogen resultater I kan bruge.

Tak for din tid, hav en god week-end, og held og lykke med projectet.

Christian S�rensen
box1@lundbykrat.dk