Regarding run time/start time.

Forum dedicated to older versions of EMS package (WRFEMS v3.2, v3.1 or older). Support is user-to-user based, so please help others if you can.
Post Reply
delinend
Posts: 65
Joined: Wed Oct 28, 2009 7:07 pm

Regarding run time/start time.

Post by delinend » Mon Jan 21, 2013 8:25 pm

Hi.

I know it is possibel to start WRFEMS, with at shifted time. And not allways from point zero.

But has anyone tryed this, and is it possible?

Normal setup:
One machine running WRFEMS from +00 to +48 hours.

Thinked setup:
Two machines.. One running from +00 to +24 and the other machine from +25 to +48. They are two fysicaly machines. And then merge the two GRIB outputs, when both runs are finish.

My Q.: Is the +25 to +48 output from machine two, the same, as if I run all 48 hours on one machine ?
Do the WRFEMS output change, if I chose not to start from the normal WRF start, but chose to shift it ex. 24 hours, when starting ?

Best regards

meteoadriatic
Posts: 1574
Joined: Wed Aug 19, 2009 10:05 am

Re: Regarding run time/start time.

Post by meteoadriatic » Mon Jan 21, 2013 9:20 pm

It is not the same.

If you start simulation at +24h, then that forecast grib will be used for initial conditions and model will start from that point. If you start at 00h then 00h grib will be used for starting model. It can never be the same.

I actually use forecast grib +06 for starting model on some setups. I didn't found drawbacks with doing that. If you have analyse grib and forecast gribs in same resolution it is probably nothing much to lose if you initialize model with forecast grib and you spare some time. Of course, simulation WILL be different. I don't say worse but different. Maybe worse, maybe better, probably just case dependent.

delinend
Posts: 65
Joined: Wed Oct 28, 2009 7:07 pm

Re: Regarding run time/start time.

Post by delinend » Tue Jan 22, 2013 8:41 pm

Thanks meteoadriatic, for your answer.

I see you point, and it was just an idea, to bring my simulation running faster.

But I still can't figur out, how to make a WRF run faster, if I have many CPU's :-(
I'm trying to make a simulation 72 hours, with a 240x240 15km. parent, with a 100x100 5km. nest. And I can't bring it down under 4 hours, with 1-way nesting.

I have tryet many hardware tunings, on my 48 core machine. HPC/Performance modes, SSD disk in RAID0 and 1, Node interleaving, high MEM speed, and alot of Memory. L2 and L3 cache adjustments. But noting makes it run faster. I allso tryet compiling WRF 3.2.1 with Intel Frontran ++, and got it working, but this gave me not mutch smaler run time :-(

Thats why is was thinking of splitting the simulation, into 2x48 core machine (they are located same place).

How to make it run faster ? I know, many want answer to that Q :-)

Best regards.

meteoadriatic
Posts: 1574
Joined: Wed Aug 19, 2009 10:05 am

Re: Regarding run time/start time.

Post by meteoadriatic » Tue Jan 22, 2013 11:01 pm

Can you post contents of your static/namelist.wrfm file in domain run directory?

delinend
Posts: 65
Joined: Wed Oct 28, 2009 7:07 pm

Re: Regarding run time/start time.

Post by delinend » Wed Jan 23, 2013 9:06 am

Sendt you a PM, with the namelist file.

I use the following commands to run:
ems_prep --dset gfs:NFS --length 72 --cycle 00 --domains 2
ems_run --domains 2

Use full GFS local, that i download from NCEP.

My best run time, from this setup, is with 44 cpu's out off 48 cpu's, is about 4 hours. I think is god to have extra cpu's (4), to handel the I/O :-)

Thanks.
Best regards.

meteoadriatic
Posts: 1574
Joined: Wed Aug 19, 2009 10:05 am

Re: Regarding run time/start time.

Post by meteoadriatic » Wed Jan 23, 2013 6:30 pm

I had an idea that you might omitted to set domain decomposition, but you did.

However, you don't use numtiles, this could be very efficient in some setups. Is that EMS v3.1 or v3.2? If it is v3.2 you can set numtiles > 1 in conf file, but if it is v3.1 you will need to do little change in ems files to boost numtiles value.

I would try with numtiles = 5, then if your run time shorten try = 10, then if you again get more speedup try = 20, etc... then when you get slowdown (or crash) then fine tune number by lowering it to the point where you get minimal running time. Every integer number is allowed.

delinend
Posts: 65
Joined: Wed Oct 28, 2009 7:07 pm

Re: Regarding run time/start time.

Post by delinend » Thu Feb 14, 2013 8:38 am

Thanks meteoadriatic.

The decomp and numtiles gave me a small % better runtime. :-)
But I think my main problem, is Numa on my AMD. And not mutch to do with that :-(

Best regards.

Post Reply