home page


Questions and Answers : Unix/Linux : sec/TS
Message board moderation

To post messages, you must log in.


Send message
Joined: 27 Jan 07
Posts: 288
Credit: 2,332,206
RAC: 0
Message 27485 - Posted: 24 Mar 2007, 19:18:29 UTC
Last modified: 24 Mar 2007, 19:41:04 UTC

How do I figure out how many seconds per time-slice my hadcm3 model is using? I.e., how fast is it calculating?

I\'m very interested in speeding this model up, as it\'s 543 CPU hours elapsed but only 9.8% completed.

Edit: Found that I\'m crunching 4.81 s/TS. What\'s the formula for estimating how long this WU will take?
ID: 27485 · Report as offensive     Reply Quote

Send message
Joined: 12 Feb 07
Posts: 9
Credit: 51,485
RAC: 0
Message 27523 - Posted: 26 Mar 2007, 14:47:23 UTC - in response to Message 27485.  

No doubt others will correct me if I\'m wrong but my understanding is as follows:
Assuming that you started with a new model and didn\'t pick one up part way through, then the model simulates 160 years from 1920 to 2080. There are 72 timesteps per model day so the total number of timesteps is 160 * 365.25 * 72 = 4207680.
Your machine is taking 4.81 s/TS so your model will take 4207680*4.81 seconds in total or roughly 234 days assuming the machine is on all the time.

You should be able to confirm this from the BOINC manager which should show the time taken so far and the time to go, both in hours.

Your percentage and elapsed CPU time seem to tie in with mine allowing for the differences in s/TS.

There is some more about this on the FAQ under:

1.3.2 \"So how long do these Trickles take and what are: a Model, a Phase, a Trickle and a Timestep?\"

ID: 27523 · Report as offensive     Reply Quote

Send message
Joined: 30 Oct 05
Posts: 15
Credit: 96,265
RAC: 0
Message 27524 - Posted: 26 Mar 2007, 15:30:23 UTC

There are actually 4,147,560 timesteps in a full model, which at 4.81 s/TS equates to 5541.6 hours (just under 231 days) crunching.

They simplify a year a bit to make the modelling easier, so the maths actually is

72 timesteps per day * 30 days per month * 12 months per year * 160 years which makes 4147200 timesteps. The remaining 360 timesteps are to do with the way the model is programmed I believe, but I can\'t remember the exact reason. I think it is to ensure that the final data transmission point is reached, but I stand to be corrected on this.
ID: 27524 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7253
Credit: 23,154,247
RAC: 0
Message 27531 - Posted: 26 Mar 2007, 19:49:56 UTC

These are TCMs, which stands for Transient Coupled Models.
The Coupled means that, unlike the original \'slab\' models, which had a fixed approximation for the ocean values, these new models have a dynamic ocean.

Because of this, ocean values have to be calculated as well, which is at one hour intervals, unlike the half hour intervals of the atmosphere.
So each day\'s calcs includes an extra 24 TimeSteps for the ocean calcs.

If you watch the text to the left of the globe for a while, you\'ll see that first off the Atmosphere values are claculated.
Then, at midnight, the day changes for a moment to the next number. But the day then chnages back to the one just calculated, and the \'title\' changes to Ocean. At this point, the values get calculated very fast. (The ocean\'s dynamics don\'t change as rapidly as do the atmosphere\'s.)

So, for each day, there are 48 atmosphere TimeSteps , and 24 ocean TimeSteps, making 72 TimeSteps per day. (Times 360 days per year, = lots.)

ID: 27531 · Report as offensive     Reply Quote

Send message
Joined: 27 Jan 07
Posts: 288
Credit: 2,332,206
RAC: 0
Message 27570 - Posted: 27 Mar 2007, 19:29:00 UTC - in response to Message 27531.  

Thanks for the extra info. I think my PC is under specs for this project....

It\'s too bad the TCM executable doesn\'t use multiple threads, one for the atmosphere and one for the ocean or something like that. I could get a work unit done in half the time (I have 2 physical CPUs)!
ID: 27570 · Report as offensive     Reply Quote

Questions and Answers : Unix/Linux : sec/TS