LHC@home

I was wondering if it is possible to get a 64-bit client of the sixtrack software under linux.

I already figured out that it is possible to compile boinc to 64-bit but if the client is not 64-bit it will not give an performance increase.

Any comments on this?

p.s. yes I know that 64-bit it self does not give an increase in performance but that the architectural changes will (16 registers instead of 8 and some other modifications) and it might even increase the accuracy of floating points numbers although I’m not sure about that.

I'm not sure whether the fortran compiler we use supports 64-bit architectures, but even if it doesn't we most likely don't want to do so because it might be yet another source of different results for same workunit.


Markku Degerholm
LHC@home Admin

64 bit is the future, and you'll have to compile to i64AMD sooner or later to get decent performance from modern architectures. You might as well find potential problems with 64 bit early and get started soon.

Future generations of Intel desktop processors from will also support iAMD64.

I would not mind doing some 64 bit benchmarks / test units to see if there is a large difference in the results.

If however the results are more accurate it would be nice to know about it in advance so people that have 64 bit capable machines can give more accurate readings. It would not be nice if an atom would hit the wall because of a miss calculation by a 32 bit machine.

If 64 bit would be more accurate it can also be used to validate 32 bit results and give a higher change of success.

> I would not mind doing some 64 bit benchmarks / test units to see if there is
> a large difference in the results.
>
> If however the results are more accurate it would be nice to know about it in
> advance so people that have 64 bit capable machines can give more accurate
> readings. It would not be nice if an atom would hit the wall because of a miss
> calculation by a 32 bit machine.
>
> If 64 bit would be more accurate it can also be used to validate 32 bit
> results and give a higher change of success.
>

Does the SixTrack software use integer calculations for the simulations? I figured that it would use floating point calculations, but please correct me if I assume incorrectly. If it uses the FPU, then it would be using 80 bit precision even on 32 bit integer cores and has since at least the i387 coprocessor days.

>> I figured that it would use floating point

Absolutely!

It's the difference between floating point maths on Intel and AMD processors that is the root of the validator problems.


Giskard - the first telepathic robot.