Project 10476

[GPU timpster]

"Project 10476

Disease Type: Cancer

These are Core18 projects for Windows+NVIDIA only."
http://fah-web.stanford.edu/cgi-bin/fah ... ed?p=10476

Why do OPERATING SYSTEM and on top of that, GPU vendor dependent work units exist? What in the world could be so different about Windows / Mac / Linux, AND AMD vs Nvidia cards that you CHOOSE to only run these projects on Windows / Nvidia cards.

Is that why AMD is not performing as fast in FAH? Is Nvidia paying you to limit what projects run on specific cards? It seems very very strange.

[bruce]

FAH is NOT paid by GPU vendors or by Opernating Systems developers and makes no recommendations about your system.

Code compiled for one of Windows / Mac / Linux will not run on the other two. Separate code must be developed for each one. In some cases, the development is relatively easy, while in other cases, it's essentially impossible. For example BigAdv was Linux-only because the Developers couldn't get it to work on Windows.

The internal designs of AMD and NV GPUs are significantly different and equal performance on one project does not imply there will be equal performance on another project.
During the early testing of some projects, such as the one you referenced, performance on AMD was very significantly inferior to performance on NV and rather than be overwhelmed with complaints about low PPD, restrictions were added to avoid being overwhelmed with complaints.

[GPU timpster]

Alright, that's very informative. I do understand the difference between GPU hardware, that ought to be a challenge, and I respect the developers for working with both.

I'm still not really clear on the operating system differences, and I just don't think it should be in the way of folding. I look at it as, you give the CPU a task, and the CPU does it, or you give the GPU some stuff to do and if the GPU understands how to, it completes it.

I completely 100% understand the Linux BigAdv stuff because I don't want to buy a windows server license just to fold on more than one 10+ core CPU or something. But with regular projects, how in the world does the operating system affect how the project will run? If you have more detailed information, I'd like you or anyone else who is familiar with it, to please expand!

[bruce]

An operating system supports various programs with a bunch of common routines. The code written by a programmer doesn't actually do I/O or manage memory allocations or intercept errors or perform a large number of other basic services. Those requests are passed from the application to the operating system which does the basic work before returning control to the App. Each OS has its own way of doing those things.

See also articles like Wikipedia's discussion of an API (Application programming interface).

[7im]

Also note this project was released 18 months ago while the GPU fahcores were undergoing rapid development from Core 17, to Core 18, and now the project is using Core 21 for new research. As mentioned above, it is not standard practice to exclude one type of hardware or operating system, unless absolutely necessary, and with the absolute understanding that efforts would continue to support all types of hardware and operating systems.

It is kind of hard to accuse the project of favoring Nvidia when the very first GPU client was released on ATI hardware. Folding is completely agnostic.

[GPU timpster]

An operating system supports various programs with a bunch of common routines. The code written by a programmer doesn't actually do I/O or manage memory allocations or intercept errors or perform a large number of other basic services. Those requests are passed from the application to the operating system which does the basic work before returning control to the App. Each OS has its own way of doing those things.

See also articles like Wikipedia's discussion of an API (Application programming interface).

WOW! I had no idea really. I thought programming was doing the same thing on all systems... very interesting, and I'll read the wiki soon. So, now what I'd like to know is: what does the Pande group think is the most efficient platform for this type of work? Linux, Mac, Windows etc?

[JimboPalmer]

Here is a thread where Windows XP lacks support that Core 21 needs, even though newer versions of Windows have that support.

viewtopic.php?f=38&t=28710

An excellent example of an application needing OS routines to get to the drivers to get to the actual hardware.

[bruce]

what does the Pande group think is the most efficient platform for this type of work? Linux, Mac, Windows etc?

This is a WU for a GPU. Since GPUs are not supported on the Mac, your answer is either Windows or Linux.

There are small differences in efficiency between WIndows and Linux, but the best answer is USE WHATEVER YOU HAVE.