climateprediction.net home page
New work discussion - 2

New work discussion - 2

Message boards : Number crunching : New work discussion - 2
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 39 · 40 · 41 · 42

AuthorMessage
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4530
Credit: 18,669,877
RAC: 14,865
Message 70072 - Posted: 22 Nov 2023, 8:24:29 UTC

You may not need to go to the trouble of setting this up because I'm working on an updated Linux version of weather at home, which we'll use as the code for the windows version.


That would be really good. As I remember, the logic behind making WAH Windows only was having models be single platform improved consistency across batches. Will there be testing to check that all other things being equal the results from a given task will be the same whichever platform it is on?
ID: 70072 · Report as offensive
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1044
Credit: 16,196,312
RAC: 12,647
Message 70073 - Posted: 22 Nov 2023, 10:43:42 UTC - in response to Message 70072.  

That would be really good. As I remember, the logic behind making WAH Windows only was having models be single platform improved consistency across batches. Will there be testing to check that all other things being equal the results from a given task will be the same whichever platform it is on?
Good question. In practise the spread in forecast differences from running on different hardware/OS, should be considerably less than the spread from perturbations applied by the scientist to the model's starting state. I know this was looked at in a published paper years ago in the original long climate runs. People at CPDN will know more about this than me as I come from a weather forecast background, where the aim is to get the spread of forecast results about the same as the known errors in the data.

As an aside, in the recent news about Google DeepMind's new AI weather forecast model, they stated they can't yet do ensemble forecasting. That's important because ensemble forecasting is key to understanding how good your 'best' forecast is.
---
CPDN Visiting Scientist
ID: 70073 · Report as offensive
Profile Iain Inglis
Volunteer moderator

Send message
Joined: 16 Jan 10
Posts: 1084
Credit: 7,729,948
RAC: 3,139
Message 70075 - Posted: 22 Nov 2023, 11:42:31 UTC - in response to Message 70073.  

[Glenn Carver wrote:] As an aside, in the recent news about Google DeepMind's new AI weather forecast model, they stated they can't yet do ensemble forecasting. That's important because ensemble forecasting is key to understanding how good your 'best' forecast is.


Thanks for mentioning that — the reports seemed very positive but without much context.
ID: 70075 · Report as offensive
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1044
Credit: 16,196,312
RAC: 12,647
Message 70100 - Posted: 8 Jan 2024, 12:25:08 UTC

(Moderators: should we have a new thread of 'New work discussion' for 2024?)

Forthcoming batches

The following batches are planned for Jan (or early Feb).

a/ Weather@Home (Windows)*
    NZ25 - New Zealand 25km grid, natural forcings.
    EAS25 - East Asia 25km grid, range of different forcings.


b/ HadAM4 (Linux)
N216 climatological runs producing high frequency northern-hemisphere output.

c/ OpeniFS (Linux)
Low resolution batch to look at variation of model results across different hardware

*We'll also roll out updated versions of the apps for Weather@Home, HadAM4, & HadSM4 to fix issues with the models failing, particularly on restarts. Although we hope to get these out before the Weather@Home batches it may not happen due to time pressure from the projects funding these batches.


---
CPDN Visiting Scientist
ID: 70100 · Report as offensive
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4530
Credit: 18,669,877
RAC: 14,865
Message 70101 - Posted: 8 Jan 2024, 13:42:42 UTC

I will open a new thread and copy your message.
ID: 70101 · Report as offensive
Previous · 1 . . . 39 · 40 · 41 · 42

Message boards : Number crunching : New work discussion - 2

©2024 cpdn.org