What I mentioned, discussion on the HCC+64bit performance. It's been posted about more than a few times.

Memory understated the number... not 20% but 25% gain for HCC: http://www.worldcommunitygrid.org/forums/wcg/viewthread?thread=22295#192355

Curiously Jean is not specifying using what xbit client. Intriguing how a 32 bit client on a 64 bit OS (never occurred one could install a 64 bit OS on a 32 bit machine), would allow for the science to use 64 bit OS/hardware optimizations. Anyway, HCC runs faster on 64 bits OS as was tested.

:-D

Kind of think the 64 bit CPU is a must.

My understanding is this: x86-64 has several operating modes. With a 32 bit OS, it will run in legacy mode. This is just like a 32 bit CPU.

With a 64 bit operating system, it can run in 64-bit mode or in compatibility mode (not legacy mode). The science application will run in compatibility mode (it doesn't matter what mode the BOINC client or manager runs in, since they aren't CPU bound). In compatibility mode, the processor can still take advantage of the 64-bit architecture. This is where we see our 25% advantage for integer operations.

IA-64 is completely different, and I don't think we will see the same advantage there.

For the record the Boinc client is a 64-bit one, but like Didactylos I don't believe that it has anything to do with the improved performance. I think the benefit comes from the fact that the 64-bit OS allows the HW to be used at its optimum. Which as you recalled is better demonstrated with integer computations.

For Didactylos: the benefit for HCC is clearly visible. Even now that the page faulting of HCC is back to its former tremendous rates (I have yet to find or to be told why) a HCC WU needs about 20 minutes less under Linux 64-bit than under XP SP3 32-bit. At its best (before I "upgrade" to Ubuntu 8.10 + Boinc 6.2.15) the difference was about 50 minutes.

Right now I am running HFCC alone when under Linux (I have a gold badge to catch soon!) and for this project there seems to be shorter runtimes too. But it is less clear and I have to do some more results analysis before being able to size a possible benefit.

Cheers. Jean.