Geant4 validation on Azure batch

Introduction

A wrapper around batch-shipyard specific for validation jobs fo Geant4 See:

1. Validation Docker Image
2. Docker Repository

The main goal of this small bash application is to allow for an easy generation of jobs.json configuraiton file for batch-shipyard based on specific Geant4 needs.
It provides interfaces to the most common batch-shipyard commands auto-generating the configuration files.

Prerequisites

1. jq Ver > 1.3 (works with 1.5)
2. batch-shipyard at least version 2.5.1 is required

Usage

Type az-batch -h for help. The general structure of the command is az-batch <command> [subcommand] [options] <summaryfile>. For example to submit jobs according to a recipe (see later), you will:

1. Edit the json configuration files
2. Initialize a batch pool on Azure: az-batch init summary.json. You can skip this step if add_pool property is set to true in the summary json file.
3. Submit jobs: az-batch submit -r recipename summary.json
4. Optional: Monitor the status of jobs: az-batch status -r recipename summary.json
5. Optional: Retrieve stdout/stderr from a job: az-batch getfile stdout.txt -t taskID -r reccipename summary.json For all files skip the file name, for all tasks, skip the -t option
6. Deallocate pool once finished: az-batch terminate -r recipename summary.json

Note: You cannot add jobs to an existing pool unless all jobs use the same docker image (the image is pulled when pool is created). This means if you want to run two set of jobs that use different images you can create a separate pool for each set of jobs or wait for the first set to end and restart the pool.

Additional scripts

The following scripts make some opration automatic:

• submitmany.sh [job1 [job2 [...]]]: submit the list of jobs monitoring the status of the queue, if the job limit is hit, wait for a job to finish before submitting another one (job limit is set in the script)
• autoresize.sh: monitor number of running job and auto-resize (scaling down) pool to reduce the number of jobs

The two scripts can be used in a chain. For example to initialize pool, submit all jobs defined in summary, scale down pool when jobs finish and finally to de-allocate pool:

az login
jobs=`jq -r '.recipes | keys | .[]' summary.json`
az-batch init summary.json && submitmany.sh $jobs && \
    autoresize.sh && az-batch terminate summary.json

Configuration

As for batch-shipyard configuration is provided via a set of json configuration files. A top-level configuration file summary.json specifies the shipyard specific files and general configuration.
File content:

{
  "add_pool": false,
  "registry": null,
  "docker_image":"andreadotti/geant4-val:latest",
  "azure_output_container":null,
  "configurations": {
    "credentials": "credentials.json",
    "config": "config.json",
    "pool": "pool.json",
    "jobs": null 
  },  
  "recipes": {
    "crosssections": " ... "
  }
}

• If add_pool is set to true, initialize pool automatically when submitting jobs.
• Specify in registry the server name for the registry (e.g. gitlab-registry.cern.ch) If images are on github no need to specify. If a registry is specified, the credentials.json file must contain an entry in docker_registry with the registry name as property for username and pwd
• docker_image is the Geant4 validation docker image to use.
• azure_output_container is the name of the Azure storage container blob where to store output data. It should already exist. If null, container name is the recipe name being used (if no recipe is used a default name "outpucontainer" is used). The container is created on the storage account if not present.
• configurations an object specifying the names of the batch-shipyard configuration files. They must all be present with the expcetion of jobs, this can be null, in such a case it will be auto generated following one of the provided recipes.
• recipes is an object containing properties (the recipe names) and one or more bash commands that are used to generate a valid batch-shipyard jobs.json configuration file.

Note that the az-batch script may modify some of the batch-shipyard configuration files based on the content of the summary.json file (e.g. the docker image to be used).
Recipes are strings of shell commands to generate jobs json file. The auto-generated file should be called .gen-jobs.json and should be placed in the same directory where all other json files are located.

A tarball containing these generated files will be created in $PWD.

TODO?: move all configuration files inline in summary.json (script will then split the files into the shipyard ones)

Batch-shipyard configuration files and additional informaiton

Note on accounts being used

In the various json configuration files storage and batch accounts are referenced via aliases. The Azure identifier (name) of the accounts should be changed in the credentials.json file and not in the other files (e.g. config.json).

Credentials

Copy the credentials-template.json file and add the passwords from azure accunts and docker registries.

General configuraions and pools

global.conf and pool.json configure general parameters and the pool to be used.

Example for jobs configuration

The file jobs-example.json is a standalone example with a single task using the andreadotti/geant4-val:latest docker image that should contain the ProcessTest application.

Under the directory ProcessTest real configuration files can be found.
Since the corresponding jobs json configuration file is quite long and tedious to write, scripts to genrate them are provided. Refer to the instructions contained in that directory.

Note: for some functionalities, this script requires that tasks have an explicit id, thus is mandatory, differently from original shipyard.

Manual usage of batch-shipyard

Some notes for reference.

Start tasks

1. Add pool: shipyard pool add --credentials credentials.json --config global.json --pool pool.json
2. Generate jobs file if needed: cd ProcessTest && python gen-jobs-crosssections.py crosssections-list.json
3. Add jobs: shipyard jobs add --credentials credentials.json --config global.json --pool pool.json --jobs ProcessTest/gen-jobs.json

Stop tasks

To remove tasks meta-data: shipyard jobs del --credentials credentials.json --config global.json --pool pool.json --jobs jobs-example.json To remove pool: shipyard pool del --credentials credentials.json --config global.json --pool pool.json --jobs jobs-example.json

Docker

A docker container is available. The Dockerfile is the one used to create the image on dockerhub: andreadotti/geant4-azure-tools.
If something in the application changes these files must be re-generated.
How to use. On the host setup the credentials file, then run:

docker run -t -i -v "$PWD:/myjobs:rw" andreadotti/geant4-azure-rools /bin/bash
$ cd /myjobs
$ az-batch init summary-docker.json
$ az-batch submit -r FTFP_BERT_Compton summary-docker.json

If you need to manipulate database files, add another -v option mounting the volume with the DBs in the image.

Advanced Azure CLI examples

Status of nodes:

az login
export AZURE_BATCH_POOL_ID=`jq -r '.pool_specification.id' pool.json`
export AZURE_BATCH_ACCOUNT=`jq -r '.credentials.batch.account' \
                            credentials.json`
export AZURE_BATCH_ACCESS_KEY=`jq -r '.credentials.batch.account_key' \
                            credentials.json `
export AZURE_BATCH_ENDPOINT=`jq -r '.credentials.batch.account_service_url' \
                            credentials.json`
az batch node list --pool-id $AZIRE_BATCH_POOL_ID \
                   --query '[*].[id,state]' --out table

Print number of nodes in running state:

az batch node list --pool-id $AZURE_BATCH_POOL_ID --query '[].id' \
                   --filter "state eq 'running'" --out tsv | wc -l

For example to perform a resize:

newnum=`az batch node list --pool-id $AZURE_BATCH_POOL_ID \
        --query '[].id' --filter "state eq 'running'" --out tsv | wc -l`
mv pool.json pool.original.json
jq --argjson numjobs ${newnum} \
    '.pool_specification.vm_count.dedicated=$numjobs' \
        pool.original.json > pool.json
az-batch resize summary.json