GPU time left is wildly inaccurate for projects with more than one GPU app

[hucker75]

Describe the bug
For projects such as Primegrid with many GPU apps, if Boinc is used to seeing one of them, then gets a different one, the "time left" estimate is wildly inaccurate.

For example, I currently have a genefer_extreme task which has reached about halfway in 2 days. So sensibly, we'd say the other half would take two days aswell. But it has an estimate of 36 days. This completely screws up scheduling for other projects on that GPU, as it believes the Primegrid task will not reach the deadline, and enters panic mode.

Steps To Reproduce

1. Run Primegrid on small tasks on GPU, such as genefer 15 (the smallest GPU task).
2. Switch Primegrid to do genefer_extreme (the largest GPU task).
3. Let it run for a bit and observe the crazy estimate of remaining time.

System Information

• OS: Windows 11
• BOINC Version: 7.20.2

Additional context
Perhaps Boinc could concentrate less on assuming the task could be non-linear in progress, and concentrate more on how long this one has taken so far? Or perhaps record how long each app takes? It would appear currently Boinc is not allowing for multiple GPU apps from one project.

[RichardHaselgrove]

If progress is linear, use <fraction_done_exact/> in an app_config.xml file.

And tell the project to pay attention to <rsc_fpops_est> when setting up the task template file - it's a project problem, not BOINC's.

[hucker75]

I thought Boinc didn't support multiple apps under GPU for time estimates? Has this changed?

Should I tell every project then? Because they all estimate badly, anything with more than one GPU app - Einstein, Primegrid, ....

But I'll try fraction done exact, sounds useful. EDIT - works perfectly, thanks.

[RichardHaselgrove]

It was changed in October 2010.

Any project which shows - and uses - an "Average processing rate" in a host's Application details (i.e. most of them, but excluding Einstein) is equipped to do that.

Note that boundary conditions apply: a new application or app_version won't show realistic estimates until the host has returned 11 "successful" tasks. Successful requires a normal exit plus validation.

[hucker75]

Surely until it has 11 successful tasks, it should estimate according to the fraction done system, or even better, just use the 1 or 2 it's completed? Whatever it uses by default is wildly wrong. And 11 tasks is far too long to wait for week-long genefer_extreme tasks.

[AenBleidd]

I don't think we have an issue here but it worth to keep it open, so I convert this to discussion

[hucker75]

It's clear to me, fraction_done_exact works perfectly every time, I turn it on for everything now. At least make an option to let the user set that globally.

[RichardHaselgrove]

It doesn't work for every app, for every project. Only use it on the ones which do work.

[hucker75]

Which does it not work on? I'm just adding it whenever I see wildly daft predictions (which is about 50% of the time!)