wrf hardware requirements

Looking for new hardware to run WRF? Intel or AMD? Check this forum.
Post Reply
papad1p
Posts: 4
Joined: Thu Apr 09, 2015 8:07 am

wrf hardware requirements

Post by papad1p » Thu Apr 09, 2015 8:18 am

Hi,

I am interested in setting up a Wrf configuration and I would very much appreciate inputs on hardware required. A comparison of dedicated single server and a rented cluster, in terms of pricing, would be most welcome. My requirements are to run the model within 2-3 hours and a resolution of 2-6km for the whole Europe.

Best,
P.

meteoadriatic
Posts: 1510
Joined: Wed Aug 19, 2009 10:05 am

Re: wrf hardware requirements

Post by meteoadriatic » Thu Apr 09, 2015 8:26 am

You will find a lot of information and examples here in this subforum.

As a short summary, it looks that it is better to stay away from any kind of network clustering and run simulations on single workstations whenever possible. Also, Intel platforms gives much better performance than AMD.

papad1p
Posts: 4
Joined: Thu Apr 09, 2015 8:07 am

Re: wrf hardware requirements

Post by papad1p » Mon Apr 13, 2015 9:00 am

Many thanks meteoadriatic,

Would you please be able to throw some light into the specs/pricing required for a single workstation to satisfy my requirements?

meteoadriatic
Posts: 1510
Joined: Wed Aug 19, 2009 10:05 am

Re: wrf hardware requirements

Post by meteoadriatic » Mon Apr 13, 2015 12:58 pm

Sure but lot of required information is missing in your question.

1) Really whole Europe including Russia?
2) 2-6km resolution means huge range of required resources (like 5-10 times more resources for 2km in comparision with 6km run)
3) How much hours in future you need to forecast?
4) Do you have specified core (NMM or ARW)?
5) Completing big job in only 2-3 hours is in my opinion needless requirement when it comes to hardware specifications. Can you explain why do you need to complete runs in such short time?

papad1p
Posts: 4
Joined: Thu Apr 09, 2015 8:07 am

Re: wrf hardware requirements

Post by papad1p » Tue Apr 14, 2015 6:56 am

Great, thank you very much meteoadriatic for your response.

I am not keen to include Russia, but I would like to plan for scalability. I have seen that 3km can be sometimes optimal in terms of accuracy. It would be much appreciated if you could provide an estimate for all Europe/3km and another one for Western Europe/6km.

My interest mainly lies in wind/solar power; typically wind/solar power projects have a lifetime of ~20 years, but I understand that a prediction for ~1year may suffice and additional data can be estimated. Your input would be valuable also here in terms of core.

I have in mind to somehow automate my system for day ahead predictions, hence the requirement for for 2-3 hours execution. Do you think this is over-ambitious?

meteoadriatic
Posts: 1510
Joined: Wed Aug 19, 2009 10:05 am

Re: wrf hardware requirements

Post by meteoadriatic » Tue Apr 14, 2015 7:41 am

Hello,
papad1p wrote:My interest mainly lies in wind/solar power; typically wind/solar power projects have a lifetime of ~20 years, but I understand that a prediction for ~1year may suffice and additional data can be estimated. Your input would be valuable also here in terms of core.
I don't understand this one.
papad1p wrote:I have in mind to somehow automate my system for day ahead predictions, hence the requirement for for 2-3 hours execution. Do you think this is over-ambitious?
So, you need one day ahead of forecast? Something like ~36 hours?

papad1p
Posts: 4
Joined: Thu Apr 09, 2015 8:07 am

Re: wrf hardware requirements

Post by papad1p » Wed Apr 15, 2015 6:52 am

Hello,

Yes the requirement is to be able to run (i) a day-ahead prediction for 24-48 hours but also have the ability to run (ii) a 1-year prediction.

The first point is concerned with the application; I need the prediction to translate the weather forecast to renewable power plant output.

Many thanks again for your input.

meteoadriatic
Posts: 1510
Joined: Wed Aug 19, 2009 10:05 am

Re: wrf hardware requirements

Post by meteoadriatic » Wed Apr 15, 2015 1:30 pm

OK,

you need something like 4000x3000 kilometers area, that with 3km resolution gives 1.300.000 points + outer domain = ~2.000.000 points which is a lot.

Give you need only about ~36h of forecast, I guess it could be done using 4-socket workstation with new line of Intel Xeons and something like 96GB RAM. If that is expensive solution, then you might want option to divide area into 4 zones and run 4 different domains on cheaper hardware, maybe with single CPU each workstation.

It is hard to predict exact the needs, not many of us has experience with such large amount of data. Still don't believe if mentioned solutions will be able to complete runs in 2-3 hours; but I personally do not see this as something to be very much important. If those are dedicated computers for this job, you will want them to run almost all of the time, so optimizing runs to run for ~5 hours (for 4 runs/day), and 10 hours (for 2 runs/day) is something to really consider. If you run 2 hours and then 4 hours idle your equipment, in my opinion that is big waste of money.

Also I don't have experience with climate runs, for your point (ii).

Post Reply

Who is online

Users browsing this forum: No registered users and 1 guest