Seite 1 von 1

Unrealistic deadlines

Verfasst: 15.09.2017 12:57
von Peter Hucker
I understand you maybe don't know exactly how long a task will take, but why are there tasks being issued for example with 150 days estimated completion time, with a deadline of only 25 days? That just doesn't add up. Yes I know the deadline is automatically extended, but why not make it correct in the first place?

Re: Unrealistic deadlines

Verfasst: 15.09.2017 21:32
von yoyo
Because the Deadline is the point where the server checks if the workunit was returned or not and if not the server sends can resent the workunit to somebody else. If a user downloads the workunit and afterwards runs away, it takes until the deadline (25 or 150 days) to issue a new workunit.

Re: Unrealistic deadlines

Verfasst: 15.09.2017 22:13
von Peter Hucker
I see, that makes sense.

The user runs away, ROTFPMSL!

So er.... next question, why can't BOINC be written to allow the deadline to change and display the new deadline? It's not only confusing to the user, but it's making it put the tasks into urgent mode when they don't need to be.

Re: Unrealistic deadlines

Verfasst: 15.09.2017 23:08
von gemini8
Boinc is a project which was handed to the community which is now working on it. It might be nicer using a system that updates the information in the client, but someone would have to implement that.

Some thoughts about deadlines:
On projects which are issueing short work units in my opinion it's not useful to have too short a deadline. I think some projects are pushing their work through using short deadlines because the Boinc client's behaviour in this respect is predictable.
On longer work units I think it's ok to have the task going into urgent mode because long workunits should finish the sooner the better. Expecially the RNA tasks have to be handled with care, and, for the normal user, are in the way of e.g. security updates. Finishing them fast is preferable to having them idleing there for an even longer time.
Another thing would be the deadlines Climate Prediction is issueing. They have quite long work units (I think I've seen tasks for five to eleven days on my machines), and they have deadlines ending somewhen next year. There's no pressure to finish them quickly, but you're not pushed to complete them today, or even better the day before yesterday.

What is the PMS in ROTFPMSL?

Re: Unrealistic deadlines

Verfasst: 15.09.2017 23:13
von Peter Hucker
gemini8 hat geschrieben:What is the PMS in ROTFPMSL?
Made my jeans wet.

Re: Unrealistic deadlines

Verfasst: 16.09.2017 04:58
von gemini8
I see. :-D

Re: Unrealistic deadlines

Verfasst: 16.09.2017 10:00
von Jacob Klein
Hmm.. Here's my take.

If the deadline is allowed to be flexible, then it may get extended even when BOINC hasn't done much work on the task. Worst case scenario (depending on things like Resource Shares), is ... the task never gets worked on, and is basically STUCK at the client with an ever-increasing deadline.

Currently, the deadline is extended on the server, such that: If BOINC contacts the server within 2 weeks, and did some work on the task, the server will extend the deadline, and not reassign the task to someone else. This works well, even in cases where BOINC hasn't done MUCH work on the task yet.

Currently, the deadline is not extended in BOINC. And when BOINC goes past deadline on that task... if the task hasn't checkpointed yet, it is aborted. If the task has checkpointed, it goes into EDF (earliest deadline first) mode, and BOINC prioritizes the task, to work on it before other tasks. This also works well - it's time to get that task done punk, regardless of how much you worked on it before!

So, yeah. The system actually works nicely overall.... just not intuitively! RNA World tasks are sooooooo weird!

As a side note, there are some steps I take whenever I get an RNA World task:
- Allow the task to checkpoint
- Exit BOINC
- Edit client_state.xml

- Find that task's <result> block
- CAREFULLY edit <report_deadline> to be EQUAL to <received_time>
--- This puts it into EDF mode right away! We work on it FULL TIME NOW.

- Find that task's <workunit> block
- CAREFULLY edit <rsc_fpops_bound> to make it 6x larger (from 200000000000000000.000000 to 1200000000000000000.000000)
--- This allows it to run for 6 times its prior limit before aborting for "ran too long". At p_fpops of 4367549860, on my fastest rig, it allows 8.7 years instead of 1.45 years :)

- Inspect the file to make sure your changes are perfect, because if not, you may lose your entire workload!
- Save the file
- Start BOINC
- Verify the task properties shows that received time equals deadline


Re: Unrealistic deadlines

Verfasst: 16.09.2017 10:44
von Peter Hucker
Unfortunately I've switched of RNA just now as it isn't on the Gridcoin whitelist. Just experimenting to see if I can get money to fund more graphics cards....