Firefly and PC GAMESS-related discussion club

Learn how to ask questions correctly

Re^5: Tesla c2050

Alex Granovsky


the program does not hang - it simply slows down because
of severe resource over-subscription. Actually, you can find
the warning message near the beginning of the output.

The reason for over-subscription is that you are using very
non-trivial input file which you in addition have tried to
modify for your hardware configuration. However, the original
input was designed to run on the specific hardware configuration
and to get the most power of it. In particular, it affinitizes
different classes of threads; and the scheme it uses together
with your changes causes severe core over-subscription.

I've attached the template input file which I hope will be
much more suitable for your hardware configuration.

Alex Granovsky

On Thu Nov 11 '10 12:09pm, GadievRadik wrote
>Hi, Dr. Granovsky, sorry for the late reply.

>We just finished large MD simulation, and we totally ready to start FF8 calculations.

>I first step we run simple input file. Firefly8 correctly detect all tesla card and run calculation, but after several steps, calculation froze without any error messages (see attach).  

>Regards, Gadiev Radik

>On Sun Nov 7 '10 11:57pm, Alex Granovsky wrote
>>Hi Radik,

>>any news on your experiments with updated Firefly binaries?

>>Alex Granovsky
>>On Wed Oct 27 '10 10:57am, GadievRadik wrote
>>>I disable TDR feature, but still have same error in out file.

>>>I tested system with latest hoomd-blue, it works perfect.
>>>I guess, this due to the altered architecture of Fermi card.
>>>Full support of this card start since cudatoolkit 3.1 release.

>>>Thanks, Radik

>>>On Tue Oct 26 '10 6:50pm, Alex Granovsky wrote

>>>>Teslas are labeled as not useable in the output because CUDA driver
>>>>reports that there is an OS imposed limit on the execution
>>>>of individual kernel. I do not think this is actually the case
>>>>as Tesla is not a graphics card. While this is most likely just
>>>>the driver's misreport, you have two options how to proceed.
>>>>Namely, you can either completely disable the TDR feature
>>>>via registry, or contact me off the list for the updated binaries.

>>>>Alex Granovsky
>>>>On Tue Oct 26 '10 1:10pm, GadievRadik wrote
>>>>>Does firefly 7.1G support calculation on latest Nvidia Fermi cards?
>>>>>When i try run firefly on 3x tesla c2050, in out file cards labled as unsupported.

>>>>>System Win Server 2009 r2
>>>>>2x intel Xenon
>>>>>3x Nvidia Tesla c2050
>>>>>Cuda toolkit 2.3 and 3.2 installed both 32 and 64 bit installed.
>>>>>All cuda tests passed.


This message contains the 2 kb attachment
[ gpu_2.inp ] Corrected input for triple Tesla, eight core system.

[ Previous ] [ Next ] [ Index ]           Thu Nov 11 '10 7:07pm
[ Reply ] [ Edit ] [ Delete ]           This message read 713 times