pycbc_make_offline_search_workflow: A workflow to search for gravitational waves

Introduction

The executable pycbc_make_offline_search_workflow is a tool used to search for gravitational waves using data from a network of gravitational-wave detectors. It performs all of the necessary steps in the workflow, including data management, template bank construction (if required), matched filtering and signal-based vetoes, multi-detector coincidence, data-quality cuts, background estimation, and software injections to test the search. Its ultimate task is to determine whether or not a compact binary coalescence is present in the given data. The output is a webpage containing the plots that can be used to understand the results of the analysis

Configuration file

The behavior of the workflow is controlled by a configuration file (also known as an ini file) that is made up of three types of sections: workflow, the pegasus profile and the executable options. The workflow sections control how different parts of the the workflow hang together. The pegasus profile sections are equivalent to lines you would have in a condor_submit file (e.g. requirements, storage size etc). Anything you would do in condor you would do here. The third section type maps the options to an executable.

[pegasus_profile]
condor|accounting_group=ligo.dev.o1.cbc.bns_spin.pycbcoffline

This is used for keeping an account of the pycbc usage

[workflow]
; https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/workflow/initialization.html
h1-channel-name = H1:GDS-FAKE_STRAIN
l1-channel-name = L1:GDS-CALIB_STRAIN
file-retention-level = all_triggers

You tend to put things in here which will be referred to later (but you can leave it empty). This is a nice way to keep options the same without the need to repeatedly define them. Here the L1/H1 data channel name are given. We also add an option to keep all the triggers/plots the analysis produces

[workflow-ifos]
l1 =
h1 =

Set up which detectors you are going to run over. A blank space after an equals sign denotes True.

[workflow-datafind]
; See https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/workflow/datafind.html
datafind-method = AT_RUNTIME_SINGLE_FRAMES
datafind-h1-frame-type = H1_ER_C00_AGG
datafind-l1-frame-type = L1_ER_C01_L1
datafind-check-segment-gaps = update_times
datafind-check-frames-exist = raise_error
datafind-check-segment-summary = no_test

This section defines which frames we are going to use and employs different levels of checks to see whether the data exists, there are gaps etc.

• ‘datafind-method’ states how we are going to find the frames. The ‘AT_RUNTIME_SINGLE_FRAMES’ means the executable returns a list of single frame files. You can however provide a cache file, but then the options need to be changed.

• ‘datafind-h1-frame-type’ refers to the frame type the H1 channel name will be found in for the time specified. Same for L1.

• ‘datafind-check-segment-gaps’ option checks to see if there are gaps in the segments from the segment database and the option ‘update_times’ will change the analysis times to skip over these gaps.

• ‘datafind-check-frames-exist’ checks to see if the frames you are looking at actually exists, and if they don’t the ‘raise_error’ option will stop the workflow.

• ‘datafind-check-segment-summary’ Checks the segment summary table and makes sure that the frames exist for all times that the segments are known

[workflow-segments]
; See https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/workflow/segments.html
segments-method = AT_RUNTIME
segments-h1-science-name = H1:DMT-SCIENCE:1
segments-l1-science-name = L1:DMT-ANALYSIS_READY:1
segments-database-url = https://dqsegdb5.phy.syr.edu
segments-veto-definer-url = https://www.lsc-group.phys.uwm.edu/ligovirgo/cbc/public/segments/ER6/H1L1V1-ER6_GDS_CALIB_STRAIN.xml
segments-science-veto = 1
segments-veto-groups =
segments-final-veto-group = 1

This section does a series of checks to the segment database for the segments you need for your analysis.

• ‘segments-method’ option should not change.

• ‘segments-h1-science-name’ option specifies the segment name at LHO we consider to flag science time. The same is given for L1.

• ‘segments-data-url’ specifies the url for the segment database we want to query.

• ‘segments-veto-definer-url’ is the url for the veto definer file we want to use for the search.

• ‘segments-science-veto’ species which category of veto you want to eliminate from your search before it is performed to consider the data science. In this instance, 1 denotes that all the times of Cat 1 vetoes. Time vetoed here is not used in any part of the analysis, and is treated as if it were not collected.

• ‘segments-veto-groups’ is an option you can populate with different veto categories and diagnostic plots will be made after each veto is employed.

• ‘segments-final-veto-group’ is an important option as the vetoes defined here will be used to remove triggers from the search before coincidence is performed. An option of 1 will remove all Cat 1 veto times from the analysis before it is performed. If you want to add cat 2 then the option is 12.

[workflow-tmpltbank]
; See https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/ahope/template_bank.html
tmpltbank-method=PREGENERATED_BANK
tmpltbank-pregenerated-bank=/home/jveitch/projects/mdc/spin/tmpltbanks/nonspin/BNS_NonSpin_30Hz_earlyaLIGO.xml

This section specifies which template bank to use

• ’tmpltbank-method’ option specifies whether you want to use a regenerated bank or to make it on the fly. In O1 we will be us a pregenerated bank.

• ‘tmpltbank-pregnerated-bank’ specifies the location of the xml with the pregenerated bank. Note that this exact location is only valid for SUGAR, and that in general one must provide their own template bank.

[workflow-splittable]
splittable-method = IN_WORKFLOW
splittable-num-banks = 2

This section sets the options for splitting the bank to help with computational costs.

• ‘splittable-method’ tells you the method by which to split the bank, in this instance it is IN_WORKFLOW. If you do not want to split the bank, change this option to NOOP

• ‘splittable-num-banks’ specifies how many banks to split the original bank into.

[workflow-matchedfilter]
; See https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/ahope/matched_filter.html
matchedfilter-method=WORKFLOW_INDEPENDENT_IFOS
min-analysis-segments = 5
max-analysis-segments = 5
output-type = hdf

This section defines how the matched filter is going to be performed. Whether it is going to be independent for each detector, and also how the analysis is actually going to be separated in to chunks given the data available.

• ‘matched-filter-method’ defines where the data is going to be separated and searched over, in this instance the data for each IFO will be considered independently and in the workflow

• ‘min-analysis-segments’ defines the minimum number of overlapping chunks you separate the data in to to analyze. This is a proxy for segment length. In this instance 5 has been stated. Therefore if the data cannot be split in to 5 overlapping chunks the code skips over the data. To understand how much time this is you need to look in the [inspiral] options and consider the segment-length and padding options specified. ‘max-analysis-segments’ is the same but for the maximum number of overlapping chunks. Be aware if you lower/raise either of these numbers you will affect the psd estimation.

• ‘output-type’ is the format of the output trigger files from the matched filter search

[workflow-coincidence]
; See https://ldas-jobs.ligo.caltech.edu/~cbc/docs/pycbc/ahope/coincidence.html
parallelization-factor = 10

This part of the workflow looks for coincidence between templates between detectors. All coincidences are kept. If you have a large template bank you probably want make the ‘parallelization-factor’ large

[workflow-injections]
injections-method=IN_WORKFLOW

This section deals with software injections. Here you are specifying whether to use either pregenerated injections sets or ones made within the workflow itself. In this case, we will use one that is created within the workflow.

[executables]
; setup of condor universe and location of executables
inspiral          = ${which:pycbc_inspiral}
injections = ${which:lalapps_inspinj}
splittable = ${which:pycbc_splitbank}
segment_query = ${which:ligolw_segment_query_dqsegdb}
segments_from_cats = ${which:ligolw_segments_from_cats_dqsegdb}
llwadd = ${which:ligolw_add}
ligolw_combine_segments = ${which:ligolw_combine_segments}
bank2hdf = ${which:pycbc_coinc_bank2hdf}
hdfinjfind = ${which:pycbc_coinc_hdfinjfind}
coinc = ${which:pycbc_coinc_findtrigs}
statmap = ${which:pycbc_coinc_statmap}
statmap_inj = ${which:pycbc_coinc_statmap_inj}
plot_sensitivity = ${which:pycbc_page_sensitivity}
plot_foundmissed = ${which:pycbc_page_foundmissed}
plot_snrifar = ${which:pycbc_page_snrifar}
page_foreground = ${which:pycbc_page_foreground}
page_injections = ${which:pycbc_page_injtable}
hdf_trigger_merge = ${which:pycbc_coinc_mergetrigs}
plot_snrchi = ${which:pycbc_page_snrchi}
plot_coinc_snrchi = ${which:pycbc_page_coinc_snrchi}
plot_segments = ${which:pycbc_page_segments}
results_page = ${which:pycbc_make_html_page}

This section defines where each of the executables live; it tells the workflow which files to process. It might be worth checking you can find all of these paths before you set the code running.

The following options are those associated to a given executable.

[llwadd]
[datafind]
urltype=file

This is the format for the return of the data find executable - you want a file.

[segments_from_cats]

Some sections are left empty. That is fine, but you have to define each option otherwise the code will complain

[ligolw_combine_segments]

[splittable]
; options for splittable job
random-sort =

This option randomly sorts the bank to be split up before processing

[injections]
waveform = SpinTaylorT4threePointFivePN

Define the waveforms you want to use for injections

[injections-bnslininj]
f-lower = 20
min-distance = 1000
max-distance = 150000
d-distr = uniform
l-distr = random
i-distr = uniform
min-mass1 = 1.0
max-mass1 = 3.1
min-mass2 = 1.0
max-mass2 = 3.1
m-distr = componentMass
min-mtotal = 2.0
max-mtotal = 6.2
disable-spin =
time-step = 89.155
time-interval = 10
seed = 1234

These are the injections parameters you want to define. Only defining ones which aren’t so obvious

• f-lower = low frequency cut off

• min-distance = (kpc)

• max-distance = (kpc)

• d-distr = the distance distribution of the injections

• l-distr = the distribution of injections in the sky

• i-distr = inclination of the injection

• time-step = time between injections. This can be whatever time you want, but remember if the injections are too close together you can screw up your psd estimation. ~90s seems ok.

• time-interval = time interval to inject the signal. It will not always be exactly at time-step, but at a time of time-step +/- random_number(0,time-interval)

• seed = random seed, choose different numbers to get different realizations of the same background distribution

[inspiral]
; inspiral analysis parameters -- added to all inspiral jobs
chisq-bins = 256
snr-threshold = 5.0
approximant = SPAtmplt
order = 7
cluster-method = window
cluster-window = 1.0
segment-length = 512
segment-start-pad = 64
segment-end-pad = 16
psd-estimation = median
psd-segment-length = 16
psd-segment-stride = 8
psd-inverse-length = 16
strain-high-pass = 30
pad-data = 8
processing-scheme = mkl
sample-rate = 4096
filter-inj-only =
low-frequency-cutoff = 40

These are the parameters you want to define for the inspiral search

• chisq-bins = number of chisq bins for the standard Bruce Allen chisq

• snr-threshold = SNR threshold

• approximant = approximation you want to use. SPAtmplt is stationary phase approximation template which is a fast implementation of Taylor F2.

• order = PN order, the numbers are double the order. So 7=3.5PN

• cluster-method = method over which to identify the loudest trigger - in this case a window

• cluster-window = take a 1 second window around the loudest trigger

• segment-length = the length of a segment you want to analyze. Remember previously we mention we want 5 overlapping segments

• segment-start-pad = the amount of time we want to pad the start of the data by. In this instance we want to not use the first 64 seconds of data, as it will contain errors from filtering. This takes in to account the length of time we lose due to PSD corruption (16s) and the wrap around effect we have due to the template (48s)

• segment-end-pad = the amount of time we want to pad the end of the data by. See above.

• psd-estimation = the method by which we want to estimate the psd

• psd-segment-length = length of time used in each psd calculation

• psd-segment-stride = time spacing between each psd calculation. 16s length with 8s stride implies a 50% overlap

• psd-inverse-length = time length used to truncate the inverse FFT (that is, the time domain realization) of the psd

• strain-high-pass = high pass filter applied to strain data before psd estimation

• pad-data = 8 second padding added to beginning of data to account for filter corruption for resampling and high-pass before data is broken up into chunks

• processing-scheme = indicates which software to use for processing (MKL = math kernel library made by Intel)

• sample-rate = sample rate of data (will be down sampled in workflow)

• filter-inj-only = Use only segments with injections in them for matched filter

• low-frequency-cutoff = low frequency limit for the matched filter search

[inspiral-h1]
; h1 specific inspiral parameters
channel-name = ${workflow|h1-channel-name}

Specify the name of the channel you want to run the inspiral analysis over for H1. Here we are referring back to the name in the workflow module

[inspiral-l1]
; l1 specific inspiral parameters
channel-name = ${workflow|l1-channel-name}

[bank2hdf]
[trig2hdf]

[coinc]
coinc-threshold = 0.000

Here we are doing exact match coincidence. So we take the light travel time between detectors and look for triggers which are coincident within this time window. The threshold defines if you want to extend the window.

[coinc-full]
decimation-factor = 1000
loudest-keep = 200
timeslide-interval=1.1

This section concerns time slides without injections, and its purpose is to keep a small number of timesmlide triggers for background estimation. Time slides are done at all relative offsets that are multiple of the ‘timeslide-interval’, which is defined here to be 1.1 seconds. We don’t store all the coincident triggers due from time slides. We keep 200 of the loudest triggers from each template time slide, given by the second option, which gives a good estimation of the background at low FAR. The top option specifies for which timeslides we will keep all triggers, to get an overall estimation of background (not just the loudest). In this instance we would keep the triggers from 1000th, 2000th, 3000th timeslide.

[coinc-injfull&coinc-fullinj]
timeslide-interval={coinc-full:timeslide-interval}
loudest-keep-value = 8.5
cluster-window = {statmap|cluster-window}

This section concerns time slides with injections in the data. We assume only one injection will be coincident with a timeslide (done every 1.1 seconds - see first option) trigger and we keep its coincidence if its ranking statistic (newSNR) > 8.5 as specified in the second option. This is to limit storage of unimpactful triggers only.

[coinc-injinj]

[pegasus_profile-statmap&pegasus_profile-statmap_inj]
condor|request_memory = 20GB

This is the amount of memory the jobs might take

[statmap&statmap_inj]
veto-window = 0.050
cluster-window = 10.0

This controls the final clustering after all coincidence testing. The cluster-window indicates the time window used for clustering. The veto-window is used to remove all coincident zero-lag triggers so that they aren’t included in background estimation

[hdfinjfind]
injection-window = 1.0

The rest of the config file concerns plotting formats

[page_foreground]
[plot_snrifar]

[plot_snrchi]
[plot_coinc_snrchi]
[plot_coinc_snrchi-inj]
[plot_coinc_snrchi-bkg]
background-front=
[plot_coinc_snrchi-inj&plot_coinc_snrchi-bkg&plot_snrchi]
newsnr-contours =  6 8 10

[plot_sensitivity]
sig-type = ifar
sig-bins = 1 3 10 30 100 300 1000 3000 10000 30000 100000

[plot_sensitivity-mchirp]
bin-type =  mchirp
bins = 0.89 1.31 1.74 2.17 2.60
min-dist = 40
max-dist = 120
dist-bins = 50

[plot_sensitivity-mtotal]
bin-type =  total_mass
bins = 2 2.4 3.2 4 6
min-dist = 40
max-dist = 120
dist-bins = 50

[plot_sensitivity-spin]
bin-type =  spin
bins = -0.4 -0.2 0.2 0.4
min-dist = 40
max-dist = 120
dist-bins = 50

[plot_sensitivity-mchirp_binless]
bin-type =  mchirp
bins = 0.89 1.31 1.74 2.17 2.60
min-dist = 40
max-dist = 120

[plot_sensitivity-mtotal_binless]
bin-type =  total_mass
bins = 2 2.4 3.2 4 6
min-dist = 40
max-dist = 120

[plot_sensitivity-spin_binless]
bin-type =  spin
bins = -0.4 -0.2 0.2 0.4
min-dist = 40
max-dist = 120

[plot_foundmissed]
[plot_foundmissed-mchirp]
axis-type=mchirp
dynamic=
[plot_foundmissed-chirpdistmchirp]
axis-type=mchirp
dynamic=
distance-type=chirp_distance
[plot_foundmissed-time]
axis-type=time
dynamic=

[plot_foundmissed-mchirp_static]
axis-type=mchirp
log-distance=
[plot_foundmissed-chirpdistmchirp_static]
axis-type=mchirp
distance-type=chirp_distance
log-distance=
[plot_foundmissed-time_static]
axis-type=time
log-distance=

[hdf_trigger_merge]
[pegasus_profile-hdf_trigger_merge]
condor|request_memory = 10GB

[page_injections]
[plot_segments]

[results_page]
analysis-title="PyCBC Coincident Analysis"
analysis-subtitle="..."

Generating the workflow

The workflow is generated by running the script pycbc_make_offline_search_workflow. This program takes the command line arguments

$ pycbc_make_offline_search_workflow --help
No CuPy
No CuPy or GPU PhenomHM module.
No CuPy or GPU response available.
No CuPy or GPU interpolation available.
usage: pycbc_make_offline_search_workflow [-h] [-v] [--version [VERSION]]
                                          [--config-files CONFIGFILE [CONFIGFILE ...]]
                                          [--config-overrides [SECTION:OPTION:VALUE ...]]
                                          [--config-delete [SECTION:OPTION ...]]
                                          --workflow-name WORKFLOW_NAME
                                          [--tags TAGS [TAGS ...]]
                                          [--output-dir OUTPUT_DIR]
                                          [--cache-file CACHE_FILE]
                                          [--plan-now] [--submit-now]
                                          [--dax-file DAX_FILE]

Program for running offline analysis through event finding and ranking then
generate post-processing and plots.

options:
  -h, --help            show this help message and exit

PyCBC common options:
  Common options for PyCBC executables.

  -v, --verbose         Add verbosity to logging. Adding the option multiple
                        times makes logging progressively more verbose, e.g.
                        --verbose or -v provides logging at the info level,
                        but -vv or --verbose --verbose provides debug logging.
  --version [VERSION]   Display PyCBC version information and exit. Can
                        optionally supply a modifier integer to control the
                        verbosity of the version information. 0 and 1 are the
                        same as --version; 2 provides more detailed PyCBC
                        library information; 3 provides information about
                        PyCBC, LAL and LALSimulation packages (if installed)

Configuration:
  Options needed for parsing config file(s).

  --config-files CONFIGFILE [CONFIGFILE ...]
                        List of config files to be used in analysis.
  --config-overrides [SECTION:OPTION:VALUE ...]
                        List of section,option,value combinations to add into
                        the configuration file. Normally the gps start and end
                        times might be provided this way, and user specific
                        locations (ie. output directories). This can also be
                        provided as SECTION:OPTION or SECTION:OPTION: both of
                        which indicate that the corresponding value is left
                        blank.
  --config-delete [SECTION:OPTION ...]
                        List of section,option combinations to delete from the
                        configuration file. This can also be provided as
                        SECTION which deletes the enture section from the
                        configuration file or SECTION:OPTION which deletes a
                        specific option from a given section.

Options for setting workflow files:
  --workflow-name WORKFLOW_NAME
                        Name of the workflow.
  --tags TAGS [TAGS ...]
                        Append the given tags to file names.
  --output-dir OUTPUT_DIR
                        Path to directory where the workflow will be written.
                        Default is to use {workflow-name}_output.
  --cache-file CACHE_FILE
                        Path to input file containing list of files to be
                        reused (the 'input_map' file)
  --plan-now            If given, workflow will immediately be planned on
                        completion of workflow generation but not submitted to
                        the condor pool. A start script will be created to
                        submit to condor.
  --submit-now          If given, workflow will immediately be submitted on
                        completion of workflow generation
  --dax-file DAX_FILE   Path to DAX file. Default is to write to the output
                        directory with name {workflow-name}.dax.

The configuration files can either be passes as local files, or given as URLs to specific configuration files managed for an analysis. For example, to generate a workflow to search two weeks of S6D data and place the results in your public_html directory, run the command:

pycbc_make_offline_search_workflow --workflow-name s6d_chunk3 --output-dir output \
  --config-files https://code.pycbc.phy.syr.edu/ligo-cbc/pycbc-config/download/master/S6/pipeline/s6_run_pycbc_er8_pre_release.ini \
  https://code.pycbc.phy.syr.edu/ligo-cbc/pycbc-config/download/master/S6/pipeline/executables.ini \
  https://code.pycbc.phy.syr.edu/ligo-cbc/pycbc-config/download/master/S6/pipeline/injections.ini \
  https://code.pycbc.phy.syr.edu/ligo-cbc/pycbc-config/download/master/S6/pipeline/data_S6.ini \
  https://code.pycbc.phy.syr.edu/ligo-cbc/pycbc-config/download/master/S6/pipeline/gps_times_s6d_big_dog_two_weeks.ini \
  --config-overrides "results_page:output-path:${HOME}/public_html/s6/s6d-big-dog-weeks"

The configuration results_page:output-path can be changed appropriately to set the output web page location.

Note

To use released executables for production analysis, you should specify the URL to an executables.ini file from the PyCBC Software repository.

Planning and Submitting the Workflow

Pegasus is used to plan and submit the workflow. To involve Pegasus to plan a PyCBC workflow, you use the command pycbc_submit_dax which takes the command line arguments

$ pycbc_submit_dax --help
usage: pycbc_submit_dax [-h] [optional arguments]

optional arguments:
  -h, --help              show this help message and exit
  -p, --pegasus-properties FILE use the specified file as
                               the pegasus properties file
  -P, --append-pegasus-property STRING add the extra property
                                          specified by the argument
  -n, --no-submit         Plan the DAX but do not submit it
  -l, --local-dir         Directory to put condor files under
  -Q, --no-query-db       Don't query the pegasus DB.

If the environment variable TMPDIR is set then this is prepended to the 
path to the temporary workflow execute directory passed to pegasus-plan.
If the --local-dir option is not given.

If the environment variable PEGASUS_FILE_DIRECTORY is set then the
script will look there for configuration, 
otherwise the script will look for this directory by querying the
pycbc.workflow module.

Note that you are running on a resource that mandates accounting, then you will also need to add a valid tag with the --accounting-tag command line argument. Please see the LDG accounting page. to determine the correct tags. These can be applied by adding the following line to your submit invocation.

For example, to plan and submit the workflow in the example above, change to the directory that you specified with the --output command line option to pycbc_make_offline_search_workflow and plan and submit the workflow:

cd output
pycbc_submit_dax --accounting-group ligo.dev.o1.cbc.explore.test --dax s6d_chunk3.dax

Note

The above example uses the accounting tag ligo.dev.o1.cbc.explore.test which should not be used in practice.

You can monitor the status of the workflow with Pegasus Dashboard, or the other Pegasus tools described below.

If the workflow runs successfully, the output will be place under the directory specified by results_page:output-path when the workflow is complete.

Monitor and Debug the Workflow (Detailed Pegasus Documentation)

To monitor the above workflow, one would run:

pegasus-status /usr1/ahnitz/log/ahnitz/pegasus/weekly_ahope/run0011

To get debugging information in the case of failures.:

pegasus-analyzer /usr1/ahnitz/log/ahnitz/pegasus/weekly_ahope/run0011

Pegasus Dashboard

The pegasus dashboard is a visual and interactive way to get information about the progress, status, etc of your workflows.

The software can be obtained from a separate pegasus package here <https://github.com/pegasus-isi/pegasus-service>.

Pegasus Dashboard is currently installed on sugar. To view your Pegasus Dashboard, in a browser go to:

https://sugar.phy.syr.edu/pegasus/u/albert.einstein

This shows a page that has a table of all your workflows that were submitted from sugar. You can view the details of a workflow by clicking on the link in the Workflow Details column of the table.

Clicking on the Workflow Details link will take you to a webpage that gives a high-level overview of the workflow, telling you how many many jobs succeeded, fail, the submit directory, etc. There is a table with tabs at the bottom of the page. If you click the tabs Failed, Running, and Successful the page will generate a table that lists all the failed, running, and successful jobs in the workflow respectively. You also have the ability to search the table for a particular kind of job using the Search bar.

You can view the details of a job by clicking the link in the Job Name column. This will take you to a Job Details page. This page will tell you where to find stdout files, stderr files, how much wall clock time the job took to run, etc. There is a table at the bottom of the page with a Failed and Successful tab. If you click on the respective tab, it will list all invocations of that job. You can click on the link in the Invocations column for more information.

On the Invocation Details page there is information about the command line arguments, executable path, CPU time, wall clock time, etc.

In certain cases, the pegasus monitor daemon may crash and this could result in invalid or nonsensical information on the dashboard (e.g. a cumulative computing time of None). This problem can be solved by running pegasus-plots on the workflow directory: the command should tell you what to do. Typically this will be running pegasus-monitord in replay mode (see its man page).

Pegasus Analyzer

The pegasus analyzer is a command-line tool for reporting sucessful and failed jobs.

To run pegasus_analyzer on your workflow, type:

pegasus-analyzer /usr1/ahnitz/log/ahnitz/pegasus/weekly_ahope/run0011

pegasus_analyzer will display a summary of suceeded, failed, and unsubmitted jobs in the workflow. After the summary information, pegasus_analyzer will display information about each failed job. An example would be:

************************************Summary*************************************

Submit Directory   : /usr1/cbiwer/log/H1L1V1-s6d_test-970012743-258000.9apn7X
Total jobs         :     24 (100.00%)
# jobs succeeded   :     19 (79.17%)
# jobs failed      :      5 (20.83%)
# jobs unsubmitted :      0 (0.00%)

******************************Failed jobs' details******************************

=====================ligolw_cbc_hardware_inj_page_ID000020======================

last state: POST_SCRIPT_FAILED
     site: local
submit file: ligolw_cbc_hardware_inj_page_ID000020.sub
output file: ligolw_cbc_hardware_inj_page_ID000020.out.001
error file: ligolw_cbc_hardware_inj_page_ID000020.err.001

-------------------------------Task #1 - Summary--------------------------------

site        : local
hostname    : avhe2010.sugar.phy.syr.edu
executable  : /home/cbiwer/projects/test_workflow/970012743-970270743/executables/ligolw_cbc_hardware_inj_page
arguments   : --source-xml hardware_injection_summary/H1L1V1-S6_CBC_HW_INJECTIONS-930493015-42111800.xml --outfile hardware_injection_summary/H1L1V1-HWINJ_SUMMARY_CAT_2-9
70012743-258000.html ----segments-xml-glob ../segments/*-SCIENCE_SEGMENTS-*-*.xml --v1-injections ----vetos-xml-glob ../segments/*-COMBINED_CAT_2_VETO_SEGS-*-*.xml --gps-
start-time 970012743 --segment-dir hardware_injection_summary --gps-end-time 970270743 --l1-injections --analyze-injections --cache-file full_data/H1L1V1-INSPIRAL_HIPE_FU
LL_DATA_CAT_2_VETO-970012743-258000.cache --h1-injections --cache-pattern *SIRE_FIRST*
exitcode    : 2
working dir : /home/cbiwer/projects/test_workflow/970012743-970270743

Task #1 - ligo-hwinjpagejob::ligolw_cbc_hardware_inj_page:1.0 - ID000020 - Kickstart stderr

Usage:  ligolw_cbc_hardware_inj_page [options]
Program to parse the inspiral injection log
ligolw_cbc_hardware_inj_page: error: no such option: ----segments-xml-glob

The output provides you with the stderr, the command line, and where the job was run.

If you have a subdax that failed, pegasus_analyzer will provide you with a command to recieve more information about the failed jobs in the subdax.

Reuse of data from a previous workflow

One of the features of Pegasus is reuse the data products of prior runs. This can be used to e.g. expand an analysis or recover a run with mistaken settings without duplicating work. The steps below explain how to do this.

Setting up a workflow for data reuse

The first step is to generate a new workflow that performs the analysis that you would like to do. This workflow should be generated in a new directory so that it does not overwrite data from your previous workflows. Data reuse happens at the pycbc_submit_dax step, so first run pycbc_make_offline_search_workflow to build a new workflow, following the instructions in the section Generating the workflow of this page.

Stop before you plan and submit the workflow with pycbc_submit_dax. You will pass an additional file to pycbc_submit_dax using the --cache-file option with a list of files that Pegasus can re-use from a previous run. The Pegasus Workflow Planner will reduce the workflow using this cache file. Reduction works by deleting jobs from the workflow whose output files have been found in some location in this cache file.

The key to data reuse is building the cache file passed to pycbc_submit_dax. This file maps a file created in the workflow to a URL and a site where that URL can be found. The syntax of the cache file is plain ASCII with each line in the file giving the location of a file in the format:

LOGICAL_FILE_NAME PHYSICAL_FILE_URL pool="SITE"

where LOGICAL_FILE_NAME is the name of the file as it appears in the workflow. This should include any subdirectory path used by the workflow to organize files in the case of, e.g., INSPIRAL files but it should not be the absolute path to the file. PHYSICAL_FILE_URL is a full URL where the file can be found, and SITE is the site on which that URL resides.

The URI in the PHYSICAL_FILE_URL can be any of the URIs that Pegasus recognizes. The URIs file://, https:// are likely the most useful. Pegasus will take care of adding transfer jobs for https:// URIs, if the data is not available locally.

The string SITE is a hint that tells Pegasus on which site the PHYSICAL_FILE_URL can be found. The SITE string should be one of the names used by pycbc_submit_dax to identify the cluster where jobs are run. In practice there are only two execution sites used by PyCBC workflows:

1. local which is the regular Condor pool on the local cluster where the workflow is being run from. This is typically used when re-using data that exists on the filesystem of the local cluster.

2. osg which is the Open Science Grid pool, as described in Running on the Open Science Grid below. This is only used if the data to be re-used is accessible via the /cvmfs filesystem.

If the SITE string for a file matches the site where a job will be run, then Pegasus assumes that the file can be accessed locally via the regular file open commands. If the SITE string does not match the site where a job will be run, then Pegasus adds transfer jobs to the workflow to move the file to the site where it will be needed by a job.

To tell Pegasus that the file is neither accessible via file open on the local submit host nor on the osg pool, then the SITE string can be set to remote. This tells Pegasus that the file is neither on the local or the osg site and so Pegasus must add file transfer jobs to fetch the file from some other site. This SITE attribute is needed beacuse a map between the job execution site and the location of the file might not be obvious from the hostname in the PHYSICAL_FILE_URL.

The following rule should be helpful when chosing the SITE string:

• If you are re-using a file that is available locally with a file:// URI in its PHYSICAL_FILE_URL (or has an implicit file:// URI since the PHYSICAL_FILE_URL starts with a /) then the string SITE should be set to local.

• If you are re-using a file from another cluster, e.g. you are on the Syracuse cluster and want to re-use data from AEI Atlas cluster, then the string SITE should be set to remote for that file. In this case, the URI in PHYSICAL_FILE_URL will begin with the scheme (e.g. https://) depending on how the file can be accessed.

To illustrate this, an example of a simple cache file containing four files for re-use from the local site is:

H1-VETOTIME_CAT3-1169107218-1066800.xml file://localhost/home/dbrown/projects/aligo/o2/analysis-4/o2-analysis-4/output/results/1._analysis_time/1.01_segment_data/H1-VETOTIME_CAT3-1169107218-1066800.xml pool="local"
L1-VETOTIME_CAT3-1169107218-1066800.xml file://localhost/home/dbrown/projects/aligo/o2/analysis-4/o2-analysis-4/output/results/1._analysis_time/1.01_segment_data/L1-VETOTIME_CAT3-1169107218-1066800.xml pool="local"
116912/H1-INSPIRAL_FULL_DATA_JOB0-1169120586-1662.hdf file://localhost/home/dbrown/projects/aligo/o2/analysis-4/o2-analysis-4/output/full_data/H1-INSPIRAL_FULL_DATA_JOB0-1169120586-1662.hdf pool="local"
116912/H1-INSPIRAL_FULL_DATA_JOB1-1169120586-1662.hdf file://localhost/home/dbrown/projects/aligo/o2/analysis-4/o2-analysis-4/output/full_data/H1-INSPIRAL_FULL_DATA_JOB1-1169120586-1662.hdf pool="local"

Note that the LOGICAL_FILE_NAME for the veto files is just the name of the file, but for the two inspiral files it contains the subdirectory that the workflow uses to organize the files by GPS time. In the case of this file Pegasus will delete from the workflow the jobs that create the files H1-VETOTIME_CAT3-1169107218-1066800.xml, L1-VETOTIME_CAT3-1169107218-1066800.xml, 116912/H1-INSPIRAL_FULL_DATA_JOB0-1169120586-1662.hdf, and 116912/H1-INSPIRAL_FULL_DATA_JOB1-1169120586-1662.hdf when it plans the workflow. Insted, the data will be re-used from the URLs specified in the cache. Since site="local" for these files, Pegasus expects that the files all exist on the host where the workflow is run from.

Once a cache file has been constructed, to enable data re-use, you follow the standard instructions for planning and submitting the workflow in the section Planning and Submitting the Workflow, but add the --cache-file argument that points to the cache file that you have created. For example:

pycbc_submit_dax --cache-file /path/to/prior_data.map --accounting-group ligo.dev.o1.cbc.explore.test --dax s6d_chunk3.dax

will use the URLs from the file /path/to/prior_data.map to implement data re-use and subsequent workflow reduction. If more than once cache file is provided, pass the paths as a comma separated list to pycbc_submit_dax:

pycbc_submit_dax --cache-file /path/to/prior_data.map,/path/to/other.map --accounting-group ligo.dev.o1.cbc.explore.test --dax s6d_chunk3.dax

Which file URLs should be included in the reuse cache? There is no single correct way of deciding this, as it depends on exactly what you are trying to do. The sections below explain how to do this for a few common situations.

Note

The [workflow] section of the ini configuration file contains an option file-retention-level. This is commonly set to all_files or all_triggers, in which case the data products re-used will be copied from the input locations and stored into the output location of the new workflow when the new workflow is run with data re-use. This can be wasteful of disk space, so you may want to set this option to either merged_triggers or results to store a smaller sub-set of the workflow’s data products. These setting will allow the use of data from a previous run, but not make duplicate copies of intermediate data files. See the documentation under Pycbc’s workflow module configuration file(s) and command line interface for more details of the file-retention-level configuration option.

Note

At present you cannot re-use .dax and .map files from a previous run. A workflow using data reuse must regenerate and re-run any sub-daxes from scratch. If you re-use a .map file rather than re-generating it, then the new workflow will write results files in the location of the old workflow. All of the examples below use an egrep -v '(dax|map)' to filter out these files.

Extending the GPS end time of a previous workflow

A common mode of data re-use is to extend the GPS end time of a previous workflow to generate a new result page that e.g. extends the analysis by a few days. This assumes that:

• The previous workflow completed successfully.

• There are no changes to the workflow configuration file, other than incrementing the end time of the workflow.

In this case, first re-run pycbc_make_offline_search_workflow to build the new workflow. The normal file retention level will copy a lot of reused data from the previous workflow directory into the new workflow directory. If you do not want to do this, use a --config-override to change the value of workflow:file-retention-level as described on the page Pycbc’s workflow module configuration file(s) and command line interface.

Then create a cache file in the following way:

1. Locate the PyCBC result page for the workflow that you wish to extend.

2. In the menu under Section 8: Workflow, locate the Output map section (usually Section 8.06) and open that page.

3. This page will show three output cache files that contain the URLs of the data created by the workflow. Locate the file that ends main.map and download it by clicking on the Link to file. This file contains the main intermediate and output data products of the workflow.

4. Edit this file so that it only contains the output of the pycbc_inspiral jobs, i.e. delete all of the lines that do not match the pattern *INSPIRAL*hdf. You can do this in a text editor, or with your favorite combination of UNIX grep, sed, awk, or perl commands. For example:

egrep 'INSPIRAL.*hdf' /path/to/downloaded/workflow-main.map > inspiral_files.map

will pull out all cache file lines for the outputs of pycbc_inspiral files and write them to a new cache file called inspiral_files.map.

5. If the files in the new cache file exist locally on the cluster where you are submitting the workflow, then the cache file is complete. If they do not, you will need to modify the file to change the PHYSICAL_FILE_URL to a valid gsiftp:// or http:// URL on the remote cluster, and change pool="local" to pool="remote". Again, these changes can be made with a text editor or UNIX shell tools. For example, if the file URLs begin with /home/dbrown and they are on the Syracuse cluster, to run on Atlas you would use the following sed commands to change the SITE and the URI in the cache file:

   sed 's/pool="local"/pool="remote"/g' inspiral_files.map > inspiral_files.map.tmp
   sed 's+/home/dbrown+gsiftp://sugwg-condor.phy.syr.edu/home/dbrown+g' inspiral_files.map.tmp > inspiral_files.map
   rm inspiral_files.map.tmp
   

6. Finally, copy the file inspiral_files.map to your new workflow directory and then run pycbc_submit_dax as usual, giving the path to inspiral_files.map as the --cache-file argument.

Re-running a workflow using a new veto definer file

Data reuse can be used to re-running a workflow with a new veto definer file, assuming that:

• The previous workflow completed successfully.

• No changes to the configuration file are made, other than changing the segments-veto-definer-url in the [workflow-segments] section of the workflow configration file (although the GPS end time can also be extended at the same time, if necessary).

In this case, first re-run pycbc_make_offline_search_workflow to build the new workflow. The normal file retention level will copy a lot of reused data from the previous workflow directory into the new workflow directory. If you do not want to do this, use a --config-override to change the value of workflow:file-retention-level as described on the page Pycbc’s workflow module configuration file(s) and command line interface.

Then create the cache file as follows:

1. Locate the PyCBC result page for the workflow that you wish to extend.

2. In the menu under Section 8: Workflow, locate the Output map section (usually Section 8.06) and open that page.

3. This page will show three output cache files that contain the URLs of the data created by the workflow. Locate the file that ends main.map and download it by clicking on the Link to file. This file contains the main intermediate and output data products of the workflow.

4. If only category 2 and higher vetoes have change, remove the output files that match the following strings from the output map file:

• VETOTIME to remove the files containing the old veto segments.

• LIGOLW_COMBINE_SEGMENTS to remove the files that combine the veto segments into categories.

• CUMULATIVE_CAT_12H_VETO_SEGMENTS to remove the files that contain times to veto.

• COINC to remove the output of the coincidence code.

• FIT to remove the background bin statistic results.

• STATMAP to remove the detection statistic ranking output.

• INJFIND to remove the results of software-injection tests.

• PAGE to remove the results make with the loudest events.

• FOREGROUND_CENSOR to remove the veto files used to remove events from the closed box plots.

• html to remove any output web pages genereated.

• png to remove any output plots generated.

• dax to remove any follow-up workflows generated.

This can be acomplished with the following command:

egrep -v '(VETOTIME|LIGOLW_COMBINE_SEGMENTS|CUMULATIVE_CAT_12H_VETO_SEGMENTS|COINC|FIT|STATMAP|INJFIND|PAGE|FOREGROUND_CENSOR|html|png|dax)' /path/to/main.map > /path/to/reuse_cache.map

If category 1 vetoes have changed, you must also remove files matching PSD, OPTIMAL, and MERGE to remove the PSD estimation jobs, the jobs that compute the optimal SNR of injections, and the merged single-detector inspiral trigger files which may also change if the category 1 vetoes change.

6. Copy the file reuse_cache.map to your new workflow directory and then run pycbc_submit_dax as usual, giving the path to reuse_cache.map as the --cache-file argument.

Re-running a failed workflow

Occasionally it may be necessary to use data from a partially completed workflow, e.g. if there a bug in an executable and you wish to re-run the workflow with a new version of the executable. If the workflow failed, no results web page will have been generated and the output data may not have been copied to the locations in main.map. To re-use data from a previous failed workflow, you need to create a cache file containing the completed jobs from the previous workflow.

To do this, cd into the local-site-scratch/work directory of your failed workflow. For example, if you used --output-dir output when planning the workflow, and then run the command:

cd /path/to/workflow/output/local-site-scratch/work

Once in this directory there should be a directory that ends with main_ID0000001 (e.g. my-workflow-main_ID0000001) Change into that directory.

Once in the main_ID0000001 directory, run the command:

for pfn in `find . -type f | sed 's+^./++g'` ; do echo $pfn file://`pwd`/$pfn pool=\"local\" ; done | egrep -v '(dax|map)' > /path/to/partial_workflow.map

changing /path/to to a location where you want to save the cache.

Now you can than use the partial_workflow.map cache file as the --cache-file argument to pycbc_submit_dax.

Using partial products from a previous workflow

If you are changing the configuration parameters of a workflow, then you can build a cache file from a previous main.map file or the files under local-site-scratch, but you will need to filter the cache file to remove the files for jobs that have a changed configuration. Here are a few examples:

• If you are changing the configuration of pycbc_inspiral you must regenerate almost all the files in the workflow so it easier to start from scratch.

• If you are changing the injections, but want to re-use the FULL_DATA previous analysis, you can filter the main.map to keep the veto files, template bank files, full data inspiral files, and PSD files but filtering out any plots and result pages. For example:

  egrep '(VETO|BANK|INSPIRAL_FULL_DATA|MERGE_FULL_DATA|PSD)' /path/to/main.map | egrep -v '(png|html|dax)' > /path/to/reuse.map
  

• If you are changing the configuration of the coincident code, you can reuse all the injection files and inspiral files. For example:

  egrep '(VETO|BANK|FULL_DATA|PSD)' /path/to/main.map | egrep -v '(COINC|FIT|STATMAP|INJFIND|html|png|dax)' /path/to/main.map > /path/to/reuse.map
  

Note

There is no rule for exactly which products can be reused as it depends on what you are changing in the workflow configuration. For partial reuse, it is best to consult an expert on how to build the cache file.

Running on the Open Science Grid

Prerequisites

There are a number of requirements on the machine on which the workflow will be started:

• Pegasus version 4.7.1 or later (at least 4.9.2 recommended).

• A gridftp server running on the submit machine

• Condor configured on the head node to connect to OSG as documented at:

  https://its-condor-blog.syr.edu/dokuwiki/doku.php?id=researchgroups:physics:sl7_cluster_setup
  

Configuring the workflow

These instructions are for the case where you plan to run all pycbc_inspiral jobs on the Open Science Grid, but the rest of the workflow will run on the local HTCondor pool attached to the submit machine. For many clusters, including LDG clusters, running in this fashion may not be available from every submit machine. First, check with local sysadmins or other experts if there is a particular node from which you must plan and submit your workflow if you desire to run on the OSG.

The first step in such running is to ensure that your workflow knows where to find all of the data it needs. While some of the files generated during the workflow will need to be served via gridftp from your submission head node (as detailed below), the gravitational wave frame data files are too large for this. They are available from CVMFS and other fall-back locations, but you need to make sure that your datafind server returns these locations (it may not do so on an LDG head node, for example, if you use the default datafind server). If you have LIGO.org credentials, you should execute:

export LIGO_DATAFIND_SERVER="datafind.ligo.org"

before you run pycbc_make_offline_search_workflow. Otherwise, contact your local system administrator for a valid datafind server that points to publicly available frame files.

In order to run pycbc_inspiral on OSG worker nodes, it must be available in a Singularity container served from CVMFS. Releases of PyCBC build such containers and publish them to CVMFS, but the workflow needs to be told to only run on nodes that have Singularity available, and it needs the location of the pycbc_inspiral executable inside of the Singularity image. To do this, add the following to the --config-overrides given to pycbc_make_offline_search_workflow:

"pegasus_profile-inspiral:pycbc|site:osg" \
"pegasus_profile-inspiral:hints|execution.site:osg" \
"pegasus_profile-inspiral:condor|Requirements:(HAS_SINGULARITY =?= TRUE) && (IS_GLIDEIN =?= True)" \
"executables:inspiral:/bin/pycbc_inspiral" \

These lines tell pycbc_make_offline_search_workflow that the inspiral jobs will run on the Open Science Grid, and that such jobs need to run at sites where Singularity is installed. They also tell them that the path to pycbc_inspiral inside the container is /bin/pycbc_inspiral. If you are an LVK user, and you will be accessing non-public LVK data, then you additionally must specify that you need to run at nodes where this authenticated frame data is available. To do that, change the Requirements line above to read instead:

"pegasus_profile-inspiral:condor|Requirements:(HAS_SINGULARITY =?= TRUE) && (HAS_LIGO_FRAMES =?= True) && (IS_GLIDEIN =?= True)" \

Because the data that the inspiral jobs will need in addition to the frame files must come from the submit machine, you also need a --config-overrides argument to pycbc_make_offline_search_workflow that sets the staging site for the main workflow to be the local site. To do this, add the following argument, replacing ${WORKFLOW_NAME} with the string that is given as the argument to the option --workflow-name:

"workflow-${WORKFLOW_NAME}-main:staging-site:osg=local"

Optionally, you can add a configuration that will check that your grid proxy is valid locally before submitting the job. This means that if your grid proxy expires before the workflow is complete, the failure will be on the local site before the job is actually submitted, and not on the remote site once the job has been scheduled and matched:

"pegasus_profile-inspiral:dagman|pre:/usr/bin/grid-proxy-info"

Another useful enhancement for OSG running is to add profiles to your inspiral job that will tell Condor to put it on hold if it has been running for more that 48 hours and terminate it after 5 failed attempts. To do this, add the follwing lines to your executables.ini file:

[pegasus_profile-inspiral]
condor|periodic_hold = (JobStatus == 2) && ((CurrentTime - EnteredCurrentStatus) > (2 * 86400))
condor|periodic_release = (JobStatus == 5) && (HoldReasonCode == 3) && (NumJobStarts < 5) && ((CurrentTime - EnteredCurrentStatus) > (300))
condor|periodic_remove = (NumJobStarts >= 5)

Running the workflow

Once you have planned the workflow as above, you must also modify the submission of the workflow if it is to run successfully on the OSG. Add the following additional arguments to pycbc_submit_dax:

--no-create-proxy \
--execution-sites osg \
--append-pegasus-property 'pegasus.transfer.bypass.input.staging=true' \
--local-staging-server gsiftp://`hostname -f` \
--remote-staging-server gsiftp://`hostname -f` \

hostname -f will give the correct value if there is a gsiftp server running on the submit machine. If not, change this as needed. The remote-staging-server is the intermediary computer than can pass files between the submitting computer and the computers doing the work. hostname -f returns the full name of the computer. This full name has to be one that is accessible to both the submit machine and the workers. The --no-create-proxy may be omitted if you have LIGO.org credentials and will be retrieving data from authenticated locations in CVMFS.

You will also need to specify where the code should get the data needed to generate reduced order model waveforms. To do this add the following additional arguments to pycbc_submit_dax:

--append-site-profile 'local:env|LAL_DATA_PATH:/cvmfs/software.igwn.org/pycbc/lalsuite-extra/current/share/lalsimulation' \
--append-site-profile 'osg:env|LAL_DATA_PATH:/cvmfs/software.igwn.org/pycbc/lalsuite-extra/current/share/lalsimulation' \

Here, current is a symbolic link to the latest version of the data and can be replaced with a specific release (e.g. e02dab8c) if required.

It is also through arguments to pycbc_submit_dax that the workflow is made aware of which Singularity image to use when running pycbc_inspiral. This is done by including the following argument to pycbc_submit_dax:

--append-site-profile "osg:condor|+SingularityImage:\"/cvmfs/singularity.opensciencegrid.org/pycbc/pycbc-el8:latest\"" \

The precise line above will cause pycbc_inspiral to run using the code in the latest version of PyCBC as found on the master branch. You may well prefer a specific version (for example, for a production run) and each release will also have a corresponding Singularity image published to CVMFS. For example, to use the 1.14.3 release of PyCBC, use instead the line:

--append-site-profile "osg:condor|+SingularityImage:\"/cvmfs/singularity.opensciencegrid.org/pycbc/pycbc-el8:v1.14.3\"" \

You may also direct the workflow to use a Singularity image of your own, if that has been published to CVMFS.

When running on the OSG under Singularity, by default much of the environment of the host node where the job runs will be inherited inside the container. In many cases this is desired, as some of the file-transfer tools that Pegasus requires can come from that environment. In other cases, however, that environment may interfere with what is in the container, and from release 1.14.3 onwards, the container itself includes any necessary file transfer tools. If you want to be sure that it is the tools installed inside the container that are used, then you must direct the workflow to have a clean environment inside the container, with nothing in it that you have not specified using lines of the form --append-site-profile 'osg:env|VARNAME:VALUE (there will also be present in the environment a few other variables that are needed for proper running of Pegasus and HTCondor). To specify that you need your OSG jobs to run in a clean environment, also include the following lines when invoking pycbc_submit_dax:

--append-site-profile "osg:condor|+InitializeModulesEnv:False" \
--append-site-profile "osg:condor|+SingularityCleanEnv:True" \
--append-site-profile "osg:condor|getenv:False" \

In particular, it is recommended that LVK users run with the lines above.

So far, we have described the arguments that will allow pycbc_inspiral to run on any OSG machine to which you have access. If, in addition, you would like to run on an XSEDE resource on which you have an allocation, then add the argument:

--append-site-profile 'osg:condor|+DESIRED_XSEDE_Sites:"Comet"' \

where you replace the desired site (in this example the Comet cluster) with wherever you have an allocation. If you want to run only on that XSEDE cluster, then also add:

--append-site-profile 'osg:condor|+DESIRED_SITES:"Comet"' \

Shared file systems cannot be used with the OSG, so make sure that the --enable-shared-filesystem argument is not provided to pycbc_submit_dax when running on the OSG.