freebsd stable - Apr 2020 - HOWTO donate CPU to the fight against the Corona-virus

Quoting Stefan Ehmann <shoesoft at gmx.net> (from Sat, 21 Mar 2020  
11:38:26 +0100):
> On Thursday, March 19, 2020 8:57:45 AM CET Alexander Leidinger via freebsd-
> stable wrote:
>> Hi,
>>
>> if someone wants to donate some FreeBSD based CPU resources to the
>> fight against the Corona-virus, here is a quick HOWTO in terms of
>> installing the Folding at Home client on FreeBSD:
>>
>>
https://www.leidinger.net/blog/2020/03/19/fighting-the-coronavirus-with-free
>> bsd-foldinghome/
>>
>
> Unfortunately, (using a CPU slot for the same work unit) TPF is 2-3 times
> slower than on Ubuntu for me. Much of the speed difference seems to  
> be related
> to libOpenCL. If remove libOpenCL on Ubuntu, it's still 20-30% faster
than on
> FreeBSD.
The pure CPU based code should be the same. Someone would have to  
trace / reverse engineer what is going on.
> Don't know how stable the TPF numbers are, so numbers may be bogus.
>
> Will a CPU slot also use the GPU with libOpenCL or is it just using better
> optimized code? I tried to install libOpenCL but all I get is:
No idea. Just an assumption: either full CPU or full GPU.
> OpenCL: Not detected: clGetPlatformIDs() returned -1001
Depending on what clGetPlatformIDs is doing and if it is not  
using/requesting some GPU support, it may be worth to check the code  
of it so see if we can improve something.
> Since there's no CUDA support for FreeBSD, I guess there is no point  
> in trying
> getting GPU slots to work.
I assume the same. First step would be to get CUDA support in FreeBSD.  
I think I remember somewhere on the X11 mailinglist someone told that  
he got some NVidia GPU compute part working... not really sure about  
that part. It would help to get a big player to request it from  
NVidia. As Netflix seems to go the "encode videos on CPU" (assumption
based upon their Intel av1 codec support/benchmarks) way, and other  
players in the video / graphics business more oriented towards linux,  
I do not have much hope in this regard ATM.

Bye,
Alexander.

-- 
http://www.Leidinger.net Alexander at Leidinger.net: PGP 0x8F31830F9F2772BF
http://www.FreeBSD.org    netchild at FreeBSD.org  : PGP 0x8F31830F9F2772BF
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: Digitale PGP-Signatur
URL:
<http://lists.freebsd.org/pipermail/freebsd-stable/attachments/20200321/5f69bb01/attachment.sig>

On Saturday, March 21, 2020 12:07:55 PM CET Alexander Leidinger
wrote:
> Quoting Stefan Ehmann <shoesoft at gmx.net> (from Sat, 21 Mar 2020
>
> 11:38:26 +0100):
> > On Thursday, March 19, 2020 8:57:45 AM CET Alexander Leidinger via
> > freebsd-
> >
> > stable wrote:
> >> Hi,
> >>
> >> if someone wants to donate some FreeBSD based CPU resources to the
> >> fight against the Corona-virus, here is a quick HOWTO in terms of
> >> installing the Folding at Home client on FreeBSD:
> >>
> >>
https://www.leidinger.net/blog/2020/03/19/fighting-the-coronavirus-with-f
> >> ree bsd-foldinghome/
> >
> > Unfortunately, (using a CPU slot for the same work unit) TPF is 2-3
times
> > slower than on Ubuntu for me. Much of the speed difference seems to
> > be related
> > to libOpenCL. If remove libOpenCL on Ubuntu, it's still 20-30%
faster than
> > on FreeBSD.
>
> The pure CPU based code should be the same. Someone would have to
> trace / reverse engineer what is going on.
I'm pretty sure now that libOpenCL is only relevant for GPU slots.

I couldn't reproduce that the presence of libOpenCL.so has any effect on CPU
slots. Didn't make much sense anyway, something else must have been going
on.
So there's probably no point in getting OpenCL to run on FreeBSD until we
have
GPU rendering.

The numbers displayed by FAHControl are rather strange:
* There is no discernible difference in speed if 1 or all CPU cores are used
(but top shows that 600% CPU cycles are burned) - happens on both Ubuntu and
Linuxolator
* According to the progress bar, Ubuntu completes 1% per minute, but
Linuxolator only 0.1% (for the same work unit)

Don't know if the numbers displayed are bogus or there is really that much
of
a difference. Maybe the issue is only related to a specific WU or to AMD-CPUs.

I've also tried https://fahbench.github.io/ but it's mainly targeted at
GPUs
and uses a different Core.

On Sunday, March 22, 2020 11:38:31 AM CEST Stefan Ehmann
wrote:
> On Saturday, March 21, 2020 12:07:55 PM CET Alexander Leidinger wrote:
> > Quoting Stefan Ehmann <shoesoft at gmx.net> (from Sat, 21 Mar
2020
> >
> > 11:38:26 +0100):
> > > On Thursday, March 19, 2020 8:57:45 AM CET Alexander Leidinger
via
> > > freebsd-
> > >
> > > stable wrote:
> > >> Hi,
> > >>
> > >> if someone wants to donate some FreeBSD based CPU resources
to the
> > >> fight against the Corona-virus, here is a quick HOWTO in
terms of
> > >> installing the Folding at Home client on FreeBSD:
> > >>
> > >>
https://www.leidinger.net/blog/2020/03/19/fighting-the-coronavirus-with
> > >> -f
> > >> ree bsd-foldinghome/
> > >
> > > Unfortunately, (using a CPU slot for the same work unit) TPF is
2-3
> > > times
> > > slower than on Ubuntu for me. Much of the speed difference seems
to
> > > be related
> > > to libOpenCL. If remove libOpenCL on Ubuntu, it's still
20-30% faster
> > > than
> > > on FreeBSD.
> >
> > The pure CPU based code should be the same. Someone would have to
> > trace / reverse engineer what is going on.
>
> I'm pretty sure now that libOpenCL is only relevant for GPU slots.
>
> I couldn't reproduce that the presence of libOpenCL.so has any effect
on CPU
> slots. Didn't make much sense anyway, something else must have been
going
> on. So there's probably no point in getting OpenCL to run on FreeBSD
until
> we have GPU rendering.
>
> The numbers displayed by FAHControl are rather strange:
> * There is no discernible difference in speed if 1 or all CPU cores are
used
> (but top shows that 600% CPU cycles are burned) - happens on both Ubuntu
> and Linuxolator
> * According to the progress bar, Ubuntu completes 1% per minute, but
> Linuxolator only 0.1% (for the same work unit)
>
> Don't know if the numbers displayed are bogus or there is really that
much
> of a difference. Maybe the issue is only related to a specific WU or to
> AMD-CPUs.
Just a short update:
I've tested the port with a different WU and everything seems normal. Speed
is
comparable to Linux and multi-core also works as expected.

My previous problems can probably be ignored, not sure what the problem
actually was.