UT theoretical chemistry code forum

It seems that we are back to this problem with the new BETA v4.00 application.

Awarded credit is jumping all over the place and now down to a new low of 1.25 cr/h, over 2,800 seconds for less than 1 credit.
Linux Result
Linux Result
Windows Result
Windows Result
Windows Result
This is on a 64 bit Linux machine that was getting around 31 cr/h. Windows 32 bit is worse as it was lower to start with at around 18 cr/h, but has dropped well below this now (to under 5 cr/h) with the new application.

Also the progress bar does not work and the estimated run times has jumped to over 2,500 Hours on my Windows machines and over 70 hours on my Linux machines.

Can these issues please be addressed ?

Conan

Conan wrote:It seems that we are back to this problem with the new BETA v4.00 application.

Awarded credit is jumping all over the place and now down to a new low of 1.25 cr/h, over 2,800 seconds for less than 1 credit.
Linux Result
Linux Result
Windows Result
Windows Result
Windows Result
This is on a 64 bit Linux machine that was getting around 31 cr/h. Windows 32 bit is worse as it was lower to start with at around 18 cr/h, but has dropped well below this now (to under 5 cr/h) with the new application.

Also the progress bar does not work and the estimated run times has jumped to over 2,500 Hours on my Windows machines and over 70 hours on my Linux machines.

Can these issues please be addressed ?

Conan

Well the run time estimates are getting better, back to almost normal run time now. Thanks for that, as every job went straight to High Priority and was stopping all other work.
Progress bar still not working.

The credit issues are still around thought.
Host 20489 is now getting back up to normal type credit awarded with BETAeOn v4.00, while it still varies a lot it was only about 6 or so WUs that had the 1 cr/h issue.
However Host 1063 is still having ALL results since 3:20:50 UTC on the 8/2/12 giving 5 cr/h or less on every BETAeOn v4.00 WU processed.
It was working OK when the new v4.00 WUs started to be issued, then credits just dropped and have not returned.

Conan

What the crap is the programmer doing for compiling the Windows version of this project? The solution is simple, get off your butt and learn how to compile code correctly. You can't just compile for Intel and decide to give AMD machines a middle finger and think you did a job well done. I won't be surprised if this project is being compiled with Intel C+ Compiler on Windows but using the GCC compiler on Linux (This is why AMD processors suddenly become 8x + faster on Linux in comparison to Windows). I have a Windows machine and an Ubuntu machine, and I'd rather not have to install this project under an Ubuntu virtual machine just to speed of computations. I've had this argument with Folding@home and some other projects doing this same hideous mistake. You should want to get the most out of all your resources available, not cripple a good percentage of your resources. Having to install this on an Ubuntu virtual machine under Windows as a workaround just because I use AMD processors is not a good workaround. Under Ubuntu a task completes in ~200 seconds per core, but in Windows it completes in ~1600 seconds.

A factor of 8 is pretty bad! We're using the mingw 4.6 compiler for windows. It would be interesting to know with which processor you are finding this difference. We'll try a similar test and see if we can reproduce it. Then we can try other compilers or compiler setting to improve the AMD+Windows combination.

And as an aside, we are big fans of AMD processors. In my experience (on linux) the intel compilers do a very good job for AMD processors -- better in fact than PGI (and gnu).

I also had to run Dotsch x64 in a VM to get some decent credits a day for this project. Running on i7 2600k Win7 x64, the Linux VM almost doubled the score. I sincerely hope the Windows client can be recompiled to perform like the Linux client.

Just to add my 2 cents worth.
I'm trying Puppy Linux on a machine with Windows 7 64 and it's nearly twice as quick with Linux. If you want to check for yourselves see my profile Concrete-mixing Moose at the Knights who say Ni! team. Computer id 28285 and computer id 25843 are the same machine with different OS.