Retrieve larger data sets from MARS on ECgate or HPCF
Back to ECMWF overview
Updated by nik: 12 March 2013
To retrieve large numbers of files from MARS you need to log in to the ECgate server and issue jobs in form of scripts.
The retrieval of data is done through a submission of a shell-script.
When running a program on the UNIX system use the batch system (not interactive mode). That means you submit the job with explicit commands so that the job is run unattended under Unix.
nohup is a way to submit a job to run unattended on a Unix system, but there exists more sophisticated batch systems for handling jobs.
The batch system currently available on ECgate and HPCF is called LoadLeveler and jobs are submitted with the command llsubmit.
OBS OBS! From June 2013 the batch system on ECgate will change from Loadleveler to SLURM and jobs are submitted with the command sbatch.
This page should be updated to contain the commands for the new batch system when it is up running.
CREATING A SCRIPT
Log in to ECgate.
Create a ksh script on your home directory on ecgate.
(The default shell is either Korn (ksh) or C-shell (csh)).
In the beginning of the script, set the Batch System keywords:
#@ shell = /usr/bin/ksh (Specify the shell) #@ job_type = serial (Indicates that this is a serial job) #@ job_name = request_to_MARS (Name of job) #@ initialdir = /scratch/ms/no/sb9/test/ (Pathname to the initial working directory. OBS: Do not user environmental variables like $USER or $SCRATCH in these keywords!) #@ output = $(job_name).$(jobid).out (*.out file) #@ error = $(job_name).$(jobid).err (Error file) #@ class = normal (indicates the priority of the job, usually this is "normal") #@ notification = complete (Sends notification on completion) #@ notify_user = <userId>@ecmwf.int (change to your userID, by default your userID ) #@ account_no = spnoflex (FLEXPART account) #@ queue (indicates the end of the keywords, mandatory)
Then add your request information which might look like this:
retrieve, class = od, ("Operational archive") stream = oper, ("operational Atmospheric model", for analysis data this would be "an") expver = 1, ("Experiment version", always use 1) date = 1, ("Specifies the Analysis/Forecast base date", n is the number of days before today) time = 00:00, ("Specifies the Analysis/Forecast base time" step = 0/to/72/by/6, type = cf, levtype = pl, levelist = 100/150/200/250/300/400/500/700/850/925/1000, param = 129.128/130.128/131.128/132.128/133.128, grid = 0.5/0.5, area = 65/0/55/20, target = "ecmwf_data.grib" (Output file containing the data)
A summary of MARS keywords can be found here:
To transfer files to zardoz add the following to your ksh script:
ectrans -remote my_ms_association@genericFtp \ -source ecmwf_data.grib \ -mailto firstname.lastname@example.org \ -onfailure \ -remove \ -verbose
RETRIEVE, LIST, READ, COMPUTE and WRITE CALLS
Instead of the retrieve keyword in the above script you can use other keywords to list, read, compute or write data.
As a rule - you should have as few retrieve calls as possible, but rather have each retrieval call collect as much data as possible.
This is to limit the number of calls to the archive considering the architectural structure of the archive. Some data are stored on tape, and this is collected "manually".
Accessing the same tape multiple times in form of many retrieval commands is a large effort.
You can collect a lot of data in one single retrieve call and subsequently change the output file and create multiple files of the data with the read/compute/write commands.
However, to retrieve e.g. ERA-INTERIM data for a full year, it would be most appropriate to split the retrieval calls up in 12 (one for each month) considering
- the architecture of MARS - volume of data extracted - restrictions on disk space - restart availability - queuing (the larger the request the slower)
By optimizing your request you can jump forward in the queue!
SUBMIT YOUR JOB
Submit your job as a batch job to LoadLeveler.
To submit your script:
MONITOR YOUR JOB
llq -u <UserId> To view where your job is in the queue
llq -l <jobId> To get a detailed description for a job
llq -s <jobId> To determine why the job is not running
llcancel <jobId>To cancel your script
See man llq for more options