This might be answered somewhere I missed, but I got a hadcm3n ocean model in the last couple days with a shorter deadline.

The estimated runtime is about 550 hours which is about 5 times the runtimes of other models. This runtime might actually be a little low based on current progress (about 6 hours for 1%). The deadline that came with it is June 30, 2011.

I calculate that my machine can make the deadline with a couple weeks to spare without going into high priority (I run several projects on a dual core), but I'm curious why the deadline has been shortened by about 9 months over the normal models?

---

CPDN deadlines have usually been long. In part this is because the project does not use many of the BOINC features that relate to job management: the reasons for this are various but include the investment of effort required to finish a long model (you can't take away someone's credits for slipping over a year-long deadline and expect them to run another model), the difficulty in validating a complex model running in different platforms (again, the user's effort must be acknowledged and a year down the line is a bit late to re-issue a model), etc.

There are consequences of this policy. It fails - as you point out - to fit CPDN properly within a multi-project schedule at the PC end; it's inefficient for the shorter models the project has now made available; it clearly encourages some users to think that it's appropriate to take years to process a model that can actually be finished in weeks.

The HADCM3N deadline is therefore a more realistic one. I don't know whether this represents a welcome improvement in the project's exploitation of BOINC. At any rate the scientific team behind this particular set of models has started with a great communication effort (here and here) - so I'm assuming they're interested in getting results back quickly.

I have one of these: it's been in "High Priority" since I got it downloaded... I'm 92 hours in, showing 18% complete, and showing 588 hrs (24.5 days running 24/7) to completion: and BTW, the deadline is June 27. Even given that you want these in faster, I think a little more leeway on the deadline is advisable.

My CDPN system is an overclocked Intel Core 2 quad Q9300, with 4Gb memory...

... looks like the 588 hours is something of an overestimate: if it's done 18% in 92 hours then it should have about 420 hours to go, making about 21 days for the whole model. The deadline allows it 91 days to finish - i.e. it would have to run 21/91 = 23% of the time to meet the deadline. That's a bit high for a model that's supposed to be running 'in the background'.

The accuracy of the to-completion estimate should improve and the high priority will then be relaxed. (Does that happen as the model goes along or at the end? As a single-project cruncher who uses machines that are on 24/7, I've never attempted to understand the combinations of FLOP estimates, machines, DCFs, model mixes etc.)

I asked the project team, and got this reply from Ed:

Unfortunately it took far longer than expected to get this high resolution version of the model to work on CPDN. [The project had no programmers for a few months. - G] Some epic work from the CPDN team has finally made it possible.

... The RAPIT team are contributing to an international review of an Atlantic ocean monitoring system (see e.g. http://www.nerc.ac.uk/press/releases/2007/29-rapid.asp) - the question is whether the monitoring should be continued.

Our stats team have some work to do on the raw results from this experiment to allow us to contribute effectively, so the tight deadline is welcome for us! Hopefully people can get these runs turned round quickly!

Hope this helps.

Thanks for the replies. Like I said, no real worry for my machine, but maybe I can suspend a project or two to help get an extra 1 or 2 of these done. After all I'm about to refurbish my daughters' computer and add Boinc, so I can spare a few extra cycles! :P

---

The pool is already empty. :(
Perhaps due to grab-and-crash computers.

---

Backups: Here

The accuracy of the to-completion estimate should improve and the high priority will then be relaxed. (Does that happen as the model goes along or at the end? As a single-project cruncher who uses machines that are on 24/7, I've never attempted to understand the combinations of FLOP estimates, machines, DCFs, model mixes etc.)

The "to completion"-estimate is at start fully determined by the flops-estimates and machine-parameters including DCF. During a run the estimate should put more and more weight on actual done for this run, and less on the initial estimates.

DCF is only updated at the end, with a successful finish. Also, it will increase fast if neccessary, but take slower to decrease. With a mix of models to run, the DCF can constantly fluctuate so the estimates will never be very accurate, especially if the flops-estimate is too low so takes longer to run than expected.

Irrealistic deadline !
I have received hadcm3n WUs with respectively 1760 and 2940 hours to complete within 3 months (90 days). Except working at 100% and 33 hours per day, my PCs will never meet this deadline.
By comparison, hadam3p WUs have an 11-month deadline for 200-300 hours. That sounds more realistic for a multi-project application !

---

The tight deadline is required for scientific reasons.

7 years ago an international scientific collaboration deployed a system of observational instruments to help provide early detection of change in the Atlantic Meridional Overturning Circulation (MOC) (see Unique monitoring system for the Atlantic circulation proves its worth).

The HadCM3 application is part of an international review assessing whether this monitoring should be continued. The experiment is being run as a series of 40 year tasks and 3 of them must be completed by this autumn to meet research deadlines (see 2 posts made by researchers yesterday).

The experiment strategy is detailed here.

Bernard

You do realise, I hope, that "deadlines" only apply to other projects, and NOT to cpdn.
For this project the time limit is just a BOINC thing, that gets ignored. As do a lot of BOINC "features". The only purpose of the "deadline" is to keep BOINC happy. The climate data will be accepted long after it passes.
BUT ...
For this experiment, only those models that get completed within the time limit will be used to seed the next batch of models. the others will get added into the data pool, for use later on.

If your computer can't finish them in time, don't sweat it. Others will provide the early model data, and you, along with a lot of other slow running machines, will provide the bulk of the data for later stages of the research.

---

Backups: Here

So the WUs will continuously run in high priority until complete (i.e. during 4 months for the 2940h one if my computer is 100% available for BOINC - respectively 8 months if only 50% or 12 months if only 33%). Sorry, this is not acceptable. I'll probably have to deselect the acceptance of other applications when no work is available for the selected ones.

---

You'll probably see the completion-estimate drop about two-to-one as the Task continues. That is, they'll complete in about half the original estimate -- if my machines are representative.

---

"We have met the enemy and he is us." -- Pogo
Greetings from coastal Washington state, the scenic US Pacific Northwest.

Bernard:

what astrowx says is correct - i received one of the cm3n work units with an estimate of 2400 hours to completion...it actually is doing 2 percent per day, so it will finish in about 50 days...