Folding Forum

Hi
Now when we got GPU folding on linux I was wondering if someone has measured the performacne difference between GPU folding on those 2 system (includeing SMP)

In general, there are too many possible permutations of hardware (HW) to come to a usable conclusion for everyone. You are going to have to test the two OS's yourself on your HW to compare. At least then you have some usable data for your own setup.

GPU speeds are the same in Win vs Lin, with the exception that it's way easier to overclock the GPU in Win, so Win tends to win that question. SMP is still faster in Lin, so overall it's a TIE.

Plus no AMD folding in Lin yet,so that's one strike.

Run what best suits you.

Linux tends to run things faster in general, one reason being it has less load from bloatware, a faster kernel, etc. I don't know if this would effect the GPU though.

Windows bloat is a Linux fanboy myth.

In the past, when Windows and Linux used the exact same fahcores, their performance was exactly the same on the same projects.

Since then, the Linux fahcores have slightly better SMP optimizations, so they run faster, NOT because Windows is bloated. The kernel version may have a very small impact, but not really worth mentioning when so many other factors have much larger impact to performance.

As a general statement, while you may find that a clean Linux system runs faster than a Windows system that is running a lot of "free" (adware-infested) software, the differences are actually quite small. You're likely to find bigger between "identical" Windows systems than between Linux and Windows.

My personal experience with Windows 8 64-bit and Ubuntu 12.04.2 64-bit is that while I can overclock higher in Windows, the same overclock is unstable in Linux. However, even with lower overclock, I was getting a slightly higher PPD when compared to Windows. For the GPUs, I found out that Linux is faster than Windows (this is due to the difference of drivers so can change with ever release of a new driver version) but overclocking in Windows is much more easier with the GUI but in Linux, the only way to overclock is my flashing the GPU. Moreover, you can't get some readouts of the GPU in Linux but you can get them in Windows. Personally, I find that if I want to run a stock system, Ubuntu is the best. If I want to tweak, I will experiment with Windows to find the stable zone and then verify it with Linux. YMMV.

7im wrote:Windows bloat is a Linux fanboy myth.

In the past, when Windows and Linux used the exact same fahcores, their performance was exactly the same on the same projects.

Since then, the Linux fahcores have slightly better SMP optimizations, so they run faster, NOT because Windows is bloated. The kernel version may have a very small impact, but not really worth mentioning when so many other factors have much larger impact to performance.

As a general statement, while you may find that a clean Linux system runs faster than a Windows system that is running a lot of "free" (adware-infested) software, the differences are actually quite small. You're likely to find bigger between "identical" Windows systems than between Linux and Windows.

I don't claim that the bloat has any significant impact, it was just an assumption. But to say that Windows' bloat is totally a myth just doesn't seem right... A totally fresh install of Windows seems to use quite a bit more CPU while idle or while doing a light task than Linux in the same situation and on the same computer, even a relatively "bloated" distro like Ubuntu or Fedora.

What you assume and what is real are not the same.
As I already stated, I have personally tested and verified that fresh installs of windows xp and Ubuntu perform at identical speeds when they are dedicated to folding. This was a few years ago when both windows and Linux still used the exact same fahcores.

Thanks for testing that 7im. It's good to know that they are the same between XP and Ubuntu on that machine, but I wonder how they do overall. Clearly more tests would be needed. Maybe your results are accurate across the entire board, I don't know. Wine makes the v6 GPU client think that it's in XP.

Linux has background processes too of course, but they are usually minimalistic. I've noticed Stonecold's observation too: Windows has antivirus/defragmentation/indexing/updating background jobs that use CPU and thus can compete with F@h. Vista and 7 run these tasks automatically, but in XP they are started manually most of the time.

Windows and Linux also use different CPU schedulers.

Much has changed since I did my testing. SMP, GPUs, etc. SMP optimizations are clearly better in linux at this time. That wasn't always the case though. The performance pendulum swings back and forth, and the operating system has one of the least impacts on folding performance as compared to so many other things.

And if you're not running AV and routinely defragging on linux as well, just like Windows, then that's a problem too.

All things being equal, Windows and Linux fold the same speed. There's nothing left to debate. The question has been answered a dozen times over in this forum. Speaking of which, I'm going to ask a mod to merge this topic in with one of those existing Win v. Lin topics. We don't need so many duplicates.

As for the current state of affairs, it's about even when mixing GPUs and CPUs. Linux SMP wins, Windows GPU wins. Tie game. See PantherX's post above as explanation.

7im wrote:Much has changed since I did my testing. SMP, GPUs, etc. SMP optimizations are clearly better in linux at this time. That wasn't always the case though. The performance pendulum swings back and forth, and the operating system has one of the least impacts on folding performance as compared to so many other things.

And if you're not running AV and routinely defragging on linux as well, just like Windows, then that's a problem too.

All things being equal, Windows and Linux fold the same speed. There's nothing left to debate. The question has been answered a dozen times over in this forum. Speaking of which, I'm going to ask a mod to merge this topic in with one of those existing Win v. Lin topics. We don't need so many duplicates.

As for the current state of affairs, it's about even when mixing GPUs and CPUs. Linux SMP wins, Windows GPU wins. Tie game. See PantherX's post above as explanation.

You don't need to run AV (unless you're running a mail server and want to do it) or defrag on Linux (for ext filesystems, btrfs, etc) unless your free space is less than 10% or so. But I was talking about a fresh install, e.g. Windows w/o AV, an unfragmented disk, etc (this was just my personal experience, I'm sure Phoronix has better, more controlled benchmarking results).

Also yeah, you're probably right, the optimizations probably play a larger roll in speed than the comparatively small increase or decrease in operating system overhead for the two.

Another thing to consider, is that modern x86-64 architectures are all NUMA these days. How well optimized is the OS for NUMA? We know that Linux introduced significant NUMA performance improvements with the 3.13 kernel, and even more NUMA optimizations since then. These performance optimizations are in effect for every Linux distribution, there is no distinction between "Desktop" and "Server" kernels in the Linux world.

Windows is a different story. How well optimized is Windows Desktop edition for NUMA? I would guess not very well, based on Microsoft's history of supporting different CPU/Memory feature sets in their Desktop editions vs. Server editions. For example, PAE was supported in all 32 bit Server editions, allowing them to use up to 64 GB of RAM with a 32 bit OS. Windows Desktop editions however, were all crippled, no proper PAE support, so they could not address more than 4 GB (more like ~3.3 GB in practice due to device addressing). I would imagine Microsoft still cripples their Desktop editions in some way, for licensing reasons, as they've always done in the past. I know for a fact that XP and Vista have no NUMA optimization whatsoever. I do not know about NUMA support in newer Desktop editions, like Windows Seven and Eight.

Perhaps the only way to know for sure, is to benchmark FAH on Windows Desktop, then later install Windows Server on the exact same machine and benchmark again. Obviously the performance difference, if any, will be more apparent on higher socket-count systems. I've got a four socket G34 Opteron test system here, but no access to Microsoft media or license keys.

The fact that M$ considers licensing in monetary terms should surprise no one. It is no different than the airlines charging separately for baggage, meals, preferred seat assignments, etc. that used to be included.

Perhaps your perspective is different than mine, but I contend that the typical home computer is a single-socket device that runs a 64-bit home version of Windows7. Main RAM is unified except for various levels of cache between the RAM and other parts of the system. The value of NUMA support, if it is absent from these OSs, would be minimal, at best, although I'd certainly prefer having that support than not. For most people, a more important question is what better ways can be found to get data to and from the GPU's processing array. I'm not aware of ways NUMA would support better access to the GPU's VRAM although that seems to be a visible concern for the hardware developers.

A bigger concern for me is how updates to Linux manage the installation and updates required for the proprietary GPU drivers that FAH requires if a person expects to use the system for other things through a GUI. See Jesse_v's how-to and some pretty widespread bug reports whenever a new kernel is generated by the updates.

Windows 7 and newer have NUMA support. Win 7 supports NUMA on up to 64 logical processors.

And unless you were running the now retired BigAdv work units in Linux, the average folder isn't going to benefit from NUMA because the average folder just doesn't have that many CPU cores, those cores are all on one socket with the same access to memory, and fah work units don't use large data sets.

I do agree that testing is the best method. While operating systems and hardware changes often, the need for testing hasn't really changed in the last 2 years since the thread was started.