LHC@home

Great; as you will have seen running flat out on intensity scans, one million turns max.
Over 100,000 tasks running! CERN side infrastructure is creaking at the seams.
Will run down in a week or two to introduce a new SixTrack version (with suitable
warning).

---

That sounds good :-) according the server status page the servers are doing good, no baglogs for validation and so on building up.

---

Christoph

I'm glad we can contribute to the science at CERN and were crunching so much work for you.

hey @ll,

at first i'd like to say that its cool to be a part from the LHC-project from my home computer. it's partwise a huge fun to computing workunits for several teams and a honour too. lhc@home is one of the teams too :)

what makes me just a little sad is, that i don't know what exactly my computer is computing for you... ok, i'am be able to have a look at the teamsites in the forums, but i think that the informations are too global defined, as that i could say "oh, ok i computing a wu for that project, with this expected result/s at the moment, and this graphic is so awesome man" (ok, a "screensaver" isn't neccessary, but a form with much more result would be).

altough you have informations into boinc like sixtrack and so on, but i wish there were some more specifications for me to know what i'am currently doing here.

the next part is, that your project didn't releases some gpu-computed wu's (or may be i've never reached someones), it's would be cool to do some cuda-operations on my graphicscard (yes, its a nvidia).


ok guys, may be you'll get some (user-experience)improvements of your delivered wu's. have a nice day and greetz to everyone in the team,
me :)

f0rkB0mb,
maybe you should read this :
http://lhcathomeclassic.cern.ch/sixtrack/contrib/sixtrack_description_files/sixtrack.htm

Happy crunching !
Manfred

f0rkB0mb,
maybe you should read this :
http://lhcathomeclassic.cern.ch/sixtrack/contrib/sixtrack_description_files/sixtrack.htm

Happy crunching !
Manfred

Made the link active.

---

Christoph

About the use of GPUs there is currently no application for that.
It is on the ToDo list but there are higher priorities first.

First GPUs; this NOT easy. SixTrack is in Fortran.
Almost all GPU support is in CUDA for C.
Nonetheless PGI is working on Fortran support.
The main loop in SixTrack is difficult to fit on the
Graphics card, but we are in the process of trying
a simpler application first.

We are currently studying the high Luminosity upgrade to the LHC.
Results will be published in due course and will be available to all.
The big concern is that at a high bunch charge, just a bigger bunch
probably, the counter circulating beams interact and the effect becomes
rather significant, but only after a large number of turns (like one million).
We are also considering how to run for more turns, a bit at a time.
It would be nice to have some graphics but they use resources (which could be
minimised mark you) but I am afraid it is not a priority.

Apart from the current production I am looking to explain some result
differences......this is immensely difficult. It is even thought that it might
be impossible to get identical results across different operating systems
and hardware. We are proving it is possible (well almost!).

---

From f0rkB0mb
the next part is, that your project didn't releases some gpu-computed wu's (or may be i've never reached someones), it's would be cool to do some cuda-operations on my graphicscard (yes, its a nvidia).

By now you know that LCH@Home has no GPU work.

I also do work for SETI@Home which generates a lot of work for GPU crunching. I simply set my project preferences so that I accept GPU work from SETI and unless I run out of GPU work from them I almost never get SETI CPU work. Meanwhile, the CPU is happily crunching away on LCH and Einstein.

Einstein@Home also does GPU work but their tasks are pretty long so my experience has been that they don't play well with the shorter Tasks of SETI. The work gets done but Einstein GPU work grabs the processor and since there is no Task switching on the GPU the Task has to run until it is complete.

Apart from the current production I am looking to explain some result
differences......this is immensely difficult. It is even thought that it might
be impossible to get identical results across different operating systems
and hardware. We are proving it is possible (well almost!).

I'm a bit confused. Are you saying you're getting different results from different OS/hardware, or are you saying the results match?

In a way he says both.

As I understand it the current oppinion in computing world is that it is not possible to get matching results over all available hardware and OS combinations.
That is also the reason why the validator is validating results which don't match 100% but are 'strongly similiar' to each other.

Eric is working to prove that it is possible.

---

Christoph

I am striving for identical results across OS and IEEE 754 conformant hardware
(which includes GPU by the way apart from some exception handling). This has
been working great but I now find a few cases which are not 0 ULP different
but which are acceptable. I must have a new problem in post-processing.
I shall publish the post mortem to the Message Boards when this issue is resolved.
The most likely explication is that I have messed up a recent change to the
code, because I am getting identical 0 ULP results with four different compilers.
Tough nut to crack! Eric.

---

First GPUs; this NOT easy. SixTrack is in Fortran.
Almost all GPU support is in CUDA for C.

What is in the Fortran, which is not in the C? Unique Scientific Library?
All other code translates to the C without any problems.

Don't use CUDA please. CUDA - for NVIDIA card only, not for ATI and other. Use OpenCL instead, it is the same for both.

I think Eric talked about using OpenACC, this is open standard for based off OpenMP which allows computations on a wide range of platforms.

The develop can write in Fortran if he likes and the code can run on CPU/GPU even huge compute clusters.

FORTRAN is an old scientific language. Most scientific programs were written in FORTRAN and are still used. There is no sense in translating them.
Tullio

Does this project really need the high horsepower computing capability of a GPU?

Just asking.

And what is the ROI of taking the time to develop and test GPU capability?

Does this project really need the high horsepower computing capability of a GPU?

Just asking.

And what is the ROI of taking the time to develop and test GPU capability?

Good questions and I believe the answer is clear from this project's history. We know that the only time it has work is when they need to recalibrate/refocus the magnets/beams for higher energies and/or different particles. It appears we (the crunchers) have never failed to get the job done in time using good ol' fashioned CPUs so yah... what's the sense in porting to GPU? I just don't see the need myself, don't see a positive ROI.

Let's not forget that recent updates to include use of sse2/3 and pni extended instruction sets are a significant optimization in and of themselves. Also, reduction of the min quorum to 2 has resulted in a huge speed up too compared to what was in place 5 years ago so again I see no pressing need for GPUs at this project.

Good old FORTRAN. It'll never die and for good reason... it's popular amongst people writing code for crunching numbers. I see no reason to stray from FORTRAN either. They say it compiles into code that is just as fast as C so why bother.

CUDA? Well, AMD users don't like to admit it but CUDA is the more mature platform and the more capable platform when it comes to crunching numbers. That's what the critics and reviewers say and I believe them. OpenCL will catch up one day, perhaps, I hope it does because I own an AMD GPU too.

Hey all,

Thank you for your explanations.
And with the people assiduous who answer our question!!!!
It is necessary to be conscious that takes a time of nutcase!!!

I hope that work by links will be more stable for the future.
At the end of this month I thought of connecting a large machine… 32 threads… on sistrack...

Eric McIntosh and Frank Schmidt continue to work hard to give units to us of lhc@home. And we want work! but you have the same defect as us all you have only 2 arms ;-) Thank you!

---