Linux kernel versions and F@H performance

[hiigaran]

Most recent topic related to this appears to be some page that references version 2.6 back in 2009. Curious to see if there is more up to date information here.

Basically, has anyone been able to test differences in performance between different kernel versions for the purposes of folding? There have been plenty of kernel updates since 2.6, and they do bring about performance changes, if benchmarks comparing them have any indication.

[bruce]

I suspect that any performance changes are due to changes in the drivers, not the kernel. You will see evidence of driver overhead, but I've never seen enough kernel overhead to matter.

[Nathan_P]

I think that kernel performance was a lot more relevant back then as cpu folding was still very much the way to go and hardware had nowhere near the performance of todays cpu's. Given the switch to gpu folding and the current power of cpu hardware I don't think its anywhere near as important as it was. I also know that most of the people who would have been willing and able to test have long since left the folding community

just my £0.02

[hiigaran]

That's a pity. I'll be honest here, I've got the time and the resources to make the comparisons, but I can't really justify diverting folding power to such tests for several hours. I guess a lot of other people are in the same boat as well.

Or perhaps I could try doing it as each kernel is updated. What would be the best way to get benchmark results that are relevant to F@H?

I suspect that any performance changes are due to changes in the drivers, not the kernel. You will see evidence of driver overhead, but I've never seen enough kernel overhead to matter.

I'd assume that from what little benchmark comparisons are out there, the testers would have kept everything the same and just rebooted with a different kernel for each set of tests. While I agree that the drivers are what make the biggest difference, I have to wonder if more powerful systems would have noticeable effects with different versions.

[JimboPalmer]

Almost what you want.

http://fahbench.github.io/

It measures GPUs rather than CPUs. (although it may actually work on CPUs, it is not tested)

[bruce]

I'd assume that from what little benchmark comparisons are out there, the testers would have kept everything the same and just rebooted with a different kernel for each set of tests. While I agree that the drivers are what make the biggest difference, I have to wonder if more powerful systems would have noticeable effects with different versions.

Why not run FAHBench when you update EITHER the kernel or the drivers?

[hiigaran]

That would skew the results though. I'd need to keep the drivers the same, or otherwise it would be difficult to tell whether the kernel or the driver update was the cause of performance changes.

In that case, I'd need to keep a 'control' version of the drivers, and switch to it before running any benchmarks.

[JimboPalmer]

As I understand this, the Linux kernal is about to get slower on intel CPUs.

https://arstechnica.com/civis/viewtopic ... &t=1415305

Mind you, so is Windows and OSX, etc.

[ChristianVirtual]

As I understand this, the Linux kernal is about to get slower on intel CPUs.

https://arstechnica.com/civis/viewtopic ... &t=1415305

Mind you, so is Windows and OSX, etc.

That is a nasty one ... the Intel bug ... thx for linking

[hiigaran]

Will it affect folding? Apparently it only affects sys calls, which means gaming for example will not be significantly affected. Though gaming and folding are two different things, so I'm not so sure.

[ChristianVirtual]

[wild guess]don't know the technical details; but it might impact the GPU drivers and subsequent its performance; for CPU folding I think its less impacting[/wild guess]
we will see next week after the patches are out ...

[JimboPalmer]

[wild guess] check pointing is going to be all syscalls, so there will be some impact, but for most of us, checkpoints are rare.

I would expect little impact to CPU folding on any platform.

I am less clear about GPU folding, I think Linux handles the GPU in user space and Windows in kernal space, so it is possible that effects vary. [/wild guess. which will be corrected next Tuesday]

[foldy]

Games using GPU were tested on Linux with kernel patch and performance is the same.
https://www.phoronix.com/scan.php?page= ... ming-Tests

Fileservers or databases with big I/O to the disk suffer from this.
https://www.phoronix.com/scan.php?page= ... 6pti&num=2

[foldy]

Will different Linux distributions like Ubuntu, Mint have different FAH performance? (With same drivers or kernel)

What about FreeBSD?

[hiigaran]

Given that it's a kernel and microcode update, it shouldn't have any effect. Unless you're using some sort of distro-specific kernel, in which case I have no idea.