pause/resume WU - 25% of credit gone

[MarcvdM]

Hi,

it happens I really need to pause the folding GPU work, for instance when I need to work with Libreoffice, which becomes near impossible to do when the GPU is at 100%. So, pausing a WU, and resuming it as little as 5 or 10 minutes later - 25% or so of estimated credit: gone. Not complaining, but that seems rather silly?

[HaloJones]

so firstly the chances are you will have a checkpoint that will lose you a little of the actual process. secondly when it first restarts it can get a little confused as to how fast it's working.

But yes, the gpu is very sensitive to how quickly the unit is finished and returned.

If you are having a big impact in LibreOffice, it suggests that it is using the gpu for acceleration. Try this

[bruce]

Actually, that link sounds like it's having LibreOffice use MORE of the GPU. (I only read it briefly.) The opposite is probably true. Many apps assume the GPU is rarely used and it is more powerful then your CPU so it's always a benefit. When you run FAH the opposite is often true. The GPU is NEVER unused and the compute blocks generally take relatively long slices of the resources. If there's a way to NOT use accelerated processing (another word for GPU processing) and use the CPU instead, it's almost always a benefit for people running a GPU FAHCore. Disable accelerated processing and see how it feels.

[HaloJones]

Code: Select all

Crash on program start

LibreOffice comes with OpenGL rendering enabled by default and this configuration might cause problems with certain combinations of graphics hardware and their drivers, which can lead to crashes. If LibreOffice crashes immediately on startup, it will not be possible to open the Options dialog and disable OpenGL there, so you have two options:

    LibreOffice 5.3 and newer:
        Start LibreOffice in Safe Mode (open "LibreOffice (Safe Mode)" in the Windows start menu or run libreoffice --safe-mode in the Linux terminal).
        Select "Configure" -> "Disable Hardware Acceleration (OpenGL, OpenCL)"
        Click "Apply Changes and Restart"
    LibreOffice 5.2 and older:
        On Windows:
            Download this registry fragment and import it in the registry editor, which will disable the UseOpenGL parameter on Windows.
        On Linux:
            Open the registrymodifications.xcu file found in LibreOffice's user profile folder and change UseOpenGL and ForceOpenGL entries to false:

            <item oor:path="/org.openoffice.Office.Common/VCL"><prop oor:name="UseOpenGL" oor:op="fuse"><value>true</value></prop></item>
            <item oor:path="/org.openoffice.Office.Common/VCL"><prop oor:name="ForceOpenGL" oor:op="fuse"><value>false</value></prop></item>

For more details, please see tdf#93546.

[bruce]

FAH does not use OpenGL, it uses OpenCL and they are very different. We can't help you with OpenGL or LibreOffice.

In gneral, FAH prefers the later versions of the proprietary drivers that you can download from the website of the GPU's manufacturer. Those drivers generally support both OpenCL and OpenGL.

[bruce]

FAHCore_22 running on your GPU will write checkpoints periodically at an interval set by the scientist that designed the project. Unless the project is really, really short, the checkpoints should be much more frequent than every 25% but after a pause or a shutdown, it always restarts from the previous checkpoint.

If the applicable work files are still on you system, you can look in the "science" log and scroll down a page or so for something that looks like this:

Code: Select all

  Total number of steps: 4000000
  XTC write frequency: 50000
  Checkpoint write frequency: 50000 (1.3%)

The numbers will most likely be different, but the only one that you're interest in is (1.3%) ... which is actually a rounded value of 1.25% if you divide 50000 by 4000000

[MarcvdM]

In my experience Libreoffice does NOT use openGL, but it uses OpenCL if you enable that (where it should really be using OpenGL: rendering of even text, let alone graphs is horrid when other processes are using the GPU at 100%). I see evidence in nvidia-smi it does allocate a type "C" (OpenCL) slot, not type "G".

It really does not matter if a process (F@H mostly) is using the GPU at 100% in type "C" or type "G" mode - the GPU is busy, period. At first firefox screen draw was almost pixel by pixel when F@H running, until I forced firefox to use the GPU (type "G"), after which it runs as if F@H was not using te GPU at all.

As to the original inquiry - I know F@H takes time to get back up to speed (PPD for the WU) after WU pause, and I know it drops back slightly in progress percentage when un-pausing. The "Estimated Credit" falls back to a value not in relation to the progress percentage at all. Even when the PPD is back up to what it was before pausing the WU the credit estimate for the WU is sure to be much lower as it was when the WU first started running. That simply does not make sense to me.

[HaloJones]

The original run from download assumes the progress will be consistent and estimates the credit based on that. If you pause the unit, let's say for an hour, the new estimated time to finish is of course at least one hour later than before and the Quick Return Bonus is affected. The QRB is not linear. So the estimated points are now lower. The longer you have paused it for the greater the impact on the QRB.

[Neil-B]

The QRB rewards "Quick Return" no "Quick Process" … to follow on from what HaloJones wrote … If you processed 50% of a 1day timeout WU in say half an hour, then paused the slot for 23 hours, then un-paused the slot and complete the rest of the WU in another half hour the total time elapsed is 24hr so between when you first paused the slot (when estimated points will include a significant QRB) and when you un-pause the slot (when there will be no QRB only base points in the estimate) you will see a significant drop in the estimated points.

[MarcvdM]

hmmm, ok then. Still rather puzzling seen the fact I mostly receive WU's with an estimated time somewhere between 6 and 19 hours. Assumiing I would actually pause a WU for a week, but end it before "you have 7 days to run the WU" the original credit of let's say 100K will by reduced to 1K or less?
update - that would probably be the "base credit"?

[PantherX]

Here's a quick breakdown for points:
Before the Timeout date: base + bonus
After Timeout but before Expiration date: base
At Expiration date: WU is deleted and you get 0 points.

[TPL]

That 7 days is time to expiration. Time out is something much less, 1 or 2 days typically. If you don't return WU before time out, it goes to assignment queu to reassigned to someone else. If that someone returns it before you, you will not be credited at all. You will always get base credit if successfully return WU before expiration(between timeout and expiration before anyone else) but QRB is a reward for fast return. The faster the more credits. Everything that comes over base credit is a bonus.

Why it is so difficult to understand?

[Neil-B]

"If that someone returns it before you, you will not be credited at all."

TPL may have misunderstood the points system a little (or maybe I have) … If you return a WU completed after time out but before expiration date you will get the base points irrespective of whether anyone else has returned it before you.

Timeouts, Deadlines, Base Points for any Project can be found https://apps.foldingathome.org/psummary

[TPL]

Ok I may mistake here. I had a believe WU will be dumped if it is already returned successfully.

[TPL]

There is another thing I want to mention. I have two computers folding but from old one I have removed GPU slot because it just is too slow. It was also recognized wrong to be Radeon HD8600 Series as it is R5 M330 in real. No reason to beat that dead horse any more. But anyway, in the beginning I was using it and still got some QRB returning WUs between timeout and expiration. Not that much of course but clearly over base points. Maybe up to double the base. But as it always go over timeout it is not good. CPU I am still using as it is capable before timeout.

It has been told it should only be base points then but I am extra rewarded.