CyberShake MareNostrum Training

From SCECpedia
Jump to navigationJump to search

This page provides training for running the CyberShake TEST site on MareNostrum (MN4)

Videos

Recordings from the live training sessions are available below:

February 2, 2020 (1:04, 157 MB)

Training overview

The goal of this training is to get you to run by hand all the steps involved in a CyberShake run. Here are the basic steps involved in the training:

  • Set up needed files
  • Initialize database with run information
  • Create Strain Green Tensors
  • Create synthetic seismograms and intensity measures
  • Create final data products

Actions you need to take will be in bold.

Terminal commands and output will be in this font.
My username is pr1ejg10 and my project is pr1ejg00.  Replace <username> or <working dir> with your username or your working directory, respectively.

Set up needed files

  1. Create a directory to work from. I recommend something in scratch.
  2. pr1ejg10@login2:~> cd /gpfs/scratch/pr1ejg00/<username>
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10> mkdir TEST
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10> cd TEST
    

    An overview of the code involved in CyberShake is provided here. For this test, we are using the SGT-related codes, the PP-related codes, and the Data Products codes, but not the Stochastic codes.

    Since MN4 does not permit outgoing connections, it is impossible to install CyberShake on MN4 directly from the repository. For the purposes of this training, I suggest you use my install directly. My CyberShake installation is located at:

    /gpfs/projects/pr1ejg00/CyberShake
    

    Due to the outgoing connection problem, on MN4 we are using a local SQLite database rather than a remote MySQL database, which is what SCEC's CyberShake install uses. Each trainee should work from their own database.

  3. Copy in my training database from /gpfs/projects/pr1ejg00/CyberShake/database/training.sqlite.
  4. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> cp /gpfs/projects/pr1ejg00/CyberShake/database/training.sqlite .
    
  5. In order to look at this database, we need sqlite. Add the SQLite module to your environment.
  6. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> module load sqlite
    
  7. Let's examine this database. Use sqlite3 to investigate the tables.
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> sqlite3 training.sqlite
    SQLite version 3.20.0 2017-07-10 19:08:59
    Enter ".help" for usage hints.
    sqlite> .tables
    CyberShake_Runs                      IM_Types                           
    CyberShake_Site_Regions              PeakAmplitudes                     
    CyberShake_Site_Ruptures             Rupture_Variation_Scenario_IDs     
    CyberShake_Site_Types                Rupture_Variation_Scenario_Metadata
    CyberShake_Sites                     Rupture_Variations                 
    ERF_IDs                              Ruptures                           
    ERF_Metadata                         SGT_Variation_IDs                  
    ERF_Probability_Models               Studies                            
    Hazard_Curve_Points                  Velocity_Model_Metadata            
    Hazard_Curves                        Velocity_Models                    
    Hazard_Datasets     
    sqlite> .schema Velocity_Models
    CREATE TABLE Velocity_Models
    (
    Velocity_Model_ID integer primary key AUTOINCREMENT not null ,
    Velocity_Model_Name varchar(50) not null,
    Velocity_Model_Version varchar(50) not null
    );
    sqlite> select * from Velocity_Models;
    1|CVM-S4.26|4.26
    
    

    I have pre-inserted some of the setup you'll need, such as Ruptures, Rupture_Variations, and CyberShake_Site_Ruptures. For queries and insertions, SQLite uses practically identical syntax to MySQL. To quit, type .quit.

    Note that not all the tables defined in the full CyberShake schema are in this test database.

Initialize database with run information

CyberShake keeps track of what we call 'runs' in the database. A run is a full CyberShake calculation for a single site.

When we run large CyberShake studies, we have scripts which automatically create runs as part of our workflows. For this test, we will populate the database ourselves.

  1. Create a database entry for our test run.
  2. We will only specify the fields in the CyberShake_Runs table which are required. Some of them make less sense not in the workflow context. First, we will determine the site, velocity model, ERF, rupture variation scenario, and SGT variation IDs needed.
    sqlite> select * from CyberShake_Sites;
    1|CyberShake Verification Test - USC|TEST|34.0192|-118.286|1
    sqlite> select * from Velocity_Models;
    1|CVM-S4.26|4.26
    sqlite> select * from ERF_IDs;
    35|WGCEP (2007) UCERF2 - Single Branch|Mean UCERF 2 - Single Branch Earthquake Rupture Forecast FINAL|1|1
    36|WGCEP (2007) UCERF2 - Single Branch 200m| Mean UCERF 2 - Single Branch Earthquake Rupture Forecast FINAL, 200m|1|1
    sqlite> select * from SGT_Variation_IDs;
    1|AWP_ODC_SGT|SGTs generated with AWP-ODC-SGT with Qp=Qs=10000
    2|AWP_ODC_SGT GPU|SGTs generated with AWP-ODC-SGT GPU
    sqlite> select * from Rupture_Variation_Scenario_IDs;
    1|36|genslip-v3.3.1b|Graves & Pitarka (2014) with uniform grid down dip hypocenter location, modified rupture variation constant
    insert into CyberShake_Runs(Site_ID, ERF_ID, SGT_Variation_ID, Velocity_Model_ID, Rup_Var_Scenario_ID, Status, Status_Time, Last_User, Max_Frequency, Low_Frequency_Cutoff, SGT_Source_Filter_Frequency) values (1, 36, 1, 1, 1, "SGT Started", "2020-02-14 12:00:00", "<username>", 0.5, 0.5, 1.0)
    
  3. Determine the run ID, which you'll use for your test.
  4. sqlite> select Run_ID from CyberShake_Runs;
    1
    
  5. Exit sqlite. We're done with it for now.
  6. sqlite> .quit
    

Create Strain Green Tensors

As outlined here, there are 7 jobs we need to run to generate SGTs for our test site.

PreCVM

Details about this stage are available here. This stage may be modified for Iceland, since each site may end up using the same volume. Note that the volume dimensions must be evenly divisible by the number of cores in that dimension.

  1. Copy over my PreCVM batch script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/precvm.slrm .
  2. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> cp /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/precvm.slrm .
    
  3. Edit the batch script. Change the sqlite path (--server sqlite:...) to point to your sqlite file instead of mine.
  4. Submit the job, using sbatch. When complete, make sure there are no errors in precvm.e .

UCVM

PreCVM must finish before you can run this stage.

Details about this stage are available here. This stage will not be part of the Icelandic processing - you won't use UCVM to create a velocity model - but some other code will provide one.

  1. Copy over my UCVM batch script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/ucvm.slrm .
  2. Submit the job. When complete, make sure there are no errors in ucvm.e .This is a parallel job, and may wait in the queue for some time before running.

PreSGT

PreCVM must finish before you can run this stage, but it can run concurrently with UCVM.

Details about this stage are available here.

  1. Copy over my PreSGT batch script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/presgt.slrm .
  2. Submit the job. When complete, make sure there are no errors in presgt.e .

PreAWP

UCVM and PreSGT must finish before you can run this stage. Note that we're skipping the Smoothing step, since we're only using a single velocity model.

Details about this stage are available here.

  1. Copy over my PreAWP batch script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/preawp.slrm .
  2. Edit the script. Change '--run-id 1' to use your correct run id, which might not be 1.
  3. Submit the job. When complete, make sure there are no errors in preawp.e .

SGT

PreAWP must finish before you can run this stage.

Details about this stage are available here.

We run two SGTs, one for each horizontal component. They can run concurrently. PreAWP set up the input files needed for both horizontal components. If you'd like to run the vertical also, you'd need to make changes to PreAWP.

  1. Copy over my SGT x and SGT y batch scripts from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/awp_x.slrm and awp_y.slrm .
  2. Submit the job. When complete, check awp_x.e and awp_y.e .
  3. We usually see lines like:
    Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
    

    This is OK and is not an error.

    This job runs on 1700 cores, so you may need to wait in the queue for a while before it runs.

NaNCheck

We run two of these as well, one for each horizontal component. The SGT for that component must finish before you can run this stage.

Details about this stage are available here.

  1. Copy over my check batch scripts from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/check_x.slrm and check_y.slrm .
  2. Submit the jobs. When complete, make sure there are no errors in check_x.e or check_y.e .

PostAWP

We also run two of these. The SGT for the component must finish before you can run this stage, but it can run concurrently with NaNCheck.

Details about this stage are available here.

  1. Copy over my post batch scripts from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post_x.slrm and post_y.slrm .
  2. Change the RUN_ID variable to be set to your run ID.
  3. Submit the jobs. When complete, make sure there are no errors in post_x.e or post_y.e .

Congratulations, you have generated a pair of SGTs!

Post-processing

The post-processing steps are illustrated in this diagram. Since you're not moving the SGTs between systems, we can skip the md5sum check stage, so we only have one stage to run.

Details about DirectSynth, the seismogram synthesis stage, are available here .

This step requires a bit more setup than the previous ones. We must provide a list of all the ruptures which need their seismograms synthesized, and typically I symlink both the rupture geometry files (from which rupture variations are generated), the SGT files, and their headers.

  1. Create a directory for the post-processing inside your working directory.
  2. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> mkdir post-processing
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST> cd post-processing
    
  3. Copy in the rupture file list from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing/rupture_file_list_TEST .
  4. Symlink the *.sgt and *.sgthead files you created earlier into the post-processing directory.
  5. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> ln -s ../TEST_fx_<run id>.sgt
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> ln -s ../TEST_fx_<run id>.sgthead
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> ln -s ../TEST_fy_<run id>.sgt
    pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> ln -s ../TEST_fy_<run id>.sgthead
    
  6. Create symlinks to all the rupture geometry files. You can use my helper script, /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing/make_lns.py.
  7. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing/make_lns.py rupture_file_list_TEST /gpfs/projects/pr1ejg00/CyberShake/ruptures/Ruptures_erf36
    
  8. Copy in the batch script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing/run_ds.slrm .
  9. Edit run_ds.slrm so RUN_ID uses the correct run id.
  10. Run the job. This one may also take a while in the queue.
  11. Check two places for errors. Check both the end of ds.e, and also the end of log.84. At the end of log.84 you should see something like:
  12. Jan 24 20:55:45.870309> Sending complete message to process 83.
    Jan 24 20:55:45.870318> Sending message to processor 83.
    Jan 24 20:55:45.870326> Shutting down.
    

    When complete, you will have 395 Seismogram, PeakVals, RotD, and Duration files. File formats are described in detail here.

Data Product generation

Once the seismograms and IM files are generated, the IMs can be put into the database and a variety of data products generated. The jobs involved in the typical CyberShake workflow are illustrated in this diagram.

For this test, we aren't going to run all the jobs, just the Load Amps, the Check, and the Curve Calc for the RotD files.

Load Amps

Details about this stage are available here.

We will point to the SQLite database for database population. I've created a helper script for this purpose.

  1. Change directories back to your top-level working directory, not the post-processing one.
  2. pr1ejg10@login2:/gpfs/scratch/pr1ejg00/pr1ejg10/TEST/post-processing> cd ..
    
  3. Copy over my helper script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/db_insert.sh .
  4. Edit the script. Change '-server sqlite:...' to point to your training.sqlite file that you used earlier in the training.
  5. Run the script. It will print output as it goes.

Check DB

Details about this stage are available here. Basically, this task verifies that all the IMs were correctly inserted into the database.

  1. Copy over my helper script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/db_check.sh .
  2. Edit the script. You'll need to change both the sqlite file it points to, and also the run id (the -r argument).
  3. Run the script. It will print output as it goes. Check for errors.

Curve Calc

Details about this stage are available here .

Before we can calculate curves, we need to have a corresponding Hazard_Dataset in the database. This only needs to be done once for a given combination of velocity model, SGT variation, rupture scenario, and ERF.

  1. Create an appropriate Hazard Dataset. First, get the IDs you'll need from your CyberShake_Runs table, then use it to populate Hazard Datasets.
  2. sqlite> select ERF_ID, Rup_Var_Scenario_ID, SGT_Variation_ID, Velocity_Model_ID from CyberShake_Runs where Run_ID=<run id>;
    36|1|1|1
    sqlite> insert into Hazard_Datasets (ERF_ID, Rup_Var_Scenario_ID, SGT_Variation_ID, Velocity_Model_ID, Prob_Model_ID, Time_Span_ID, Max_Frequency, Low_Frequency_Cutoff) values (36, 1, 1, 1, 1, 1, 0.5, 0.5)
    
  3. Copy over my helper script from /gpfs/scratch/pr1ejg00/pr1ejg10/TEST/curve_calc.sh .
  4. Edit the script. Replace the first argument, the path to the sqlite file (/gpfs/projects/pr1ejg00/CyberShake/database/test.sqlite in mine) with the path to your database. Also, replace the --run-id with the correct one for your test.
  5. Run the script. It will print output as it goes.

Congratulations! You have calculated CyberShake hazard curves! You can access them either by looking at the *.pdf and *.png files you've created, or by using sqlite to examine the values in the database directly.

Verification

To confirm that you got the right answers, compare one of your hazard curves with the one I generated.

For the 3 second RotD50 curve, the reference values (the values I got when I ran the code) are:

X value Y value
0.005 0.00669446347113578
0.007 0.00669446347113578
0.0098 0.00668048983780867
0.0137 0.00665296801989335
0.0192 0.00651297322733557
0.0269 0.00630546899968387
0.0376 0.00576024330858993
0.0527 0.00509651492213958
0.0738 0.00416060395812046
0.103 0.00279691975696716
0.145 0.00166295938324768
0.203 0.000834190556129544
0.284 0.000362477682089857
0.397 0.0001266222803048
0.556 2.71102060581674e-05
0.778 4.75466199389984e-06
1.09 1.86439212512823e-07
1.52 0.0
2.13 0.0