Skip to content

Latest commit

 

History

History
174 lines (119 loc) · 8.91 KB

README.md

File metadata and controls

174 lines (119 loc) · 8.91 KB
Raw

Running VariantSpark on AWS EMR

This page contains 'working' instructions for running variant-spark on Amazon EMR.

NOTE: It is also possible to run VariantSpark on AWS EC2 with Hadoop/Spark libraries installed, if your team has expertise in managing EC2 and Hadoop clusters. We do not provide instructions for this scenario on this page.

ALTERNATIVE: If you'd prefer to quickly try a working sample of the variant-spark importance analysis, then you can use our example on the Databricks platform. See this HipsterIndex Blog Post for detail. You can sign up for the community edition and use it for one hour at a time for free. See this link for the 6 quick steps to run this example.

Process and Goal

You will set up both a client and a server.

For the client you'll use/set up the following:

  • your terminal
  • awscli - the AWS client scripting utility
  • vs-emr - a utility wrapper on aws cli that simplifies creation of EMR clusters for variant-spark and submission of variant-spark job steps
  • server configuration file - config.yaml

For the server you'll use/set up the following:

  • several S3 buckets
  • at least one AWS EMR cluster with one master and one or more worker nodes w/Apache Spark
  • an IAM role

We've included a diagram for your reference:

AWS EMR Architecture

Client Setup

  • verify system requirements include python2.7+ and associated version of pip
  • install the AWS command line utility aws cli as described in Installing the AWS Command Line Interface.
  • configure your aws cli providing your default region, access keys, etc using the aws configure command. More info on the configuration process and using the aws cli can be found at AWS CLI User Guide
  • run pip install --user vs-emr to install the the vs-emr utility
  • verify the installation and see the available commands by running this command: vs-emr --help
  • configure vs-emr by running:
    • vs-emr configure

NOTES on vs-emr install:

  • If you use some form of python virtual environment, or want to install vs-emr system-wide, you should skip the --user option.

Server Setup - First Part (for S3)

  • choose S3 storage for variant-spark output files and for the EMR cluster logs, use existing S3 bucket, or create a new bucket. We will use:
  • use s3://<your-bucket-name>/variant-spark/output/ as the output location
  • use s3://<your-bucket-name>/variant-spark/logs/ as the cluster log location.

NOTES on S3:

  • Please note the trailing slashes! They are important to designate folders in S3.

  • The sample data for hipsterIndex demo are available at s3://variant-spark/datasets/hipsterIndex/.

    Server Setup - Second Part (for the EMR cluster)

  • create a small cluster with two m4.large worker instances (as defined in ~/.vs_emr/config.yaml) by running this command:vs-emr start-cluster --cluster-id-file /tmp/cluster-id

NOTES on EMR:

  • This command will start an EMR cluster configured to run variant-spark and save its id to /tmp/cluster-id.

  • The cluster will auto-terminate after the last step has been completed so we need to submit the steps before the cluster setup has finished (which usually takes between 5 to 8 minutes).

    Run the Analysis

  • run the variant-spark importance on hipster index data run (use your S3 output bucket name):

    vs-emr submit-cmd --cluster-id-file /tmp/cluster-id importance -if s3://variant-spark/datasets/hipsterIndex/hipster.vcf.bz2 -ff s3://variant-spark/datasets/hipsterIndex/hipster_labels.txt -fc label -v -rn 1000 -rbs 100 -on 100 -of s3://<your-bucket-name>/variant-spark/output/hipster-importance.csv

NOTES on Analysis:

  • Run the command above on a single line in your terminal
  • This should complete in about 5 minutes (after the cluster setup has finished).
  • This should save the top 100 important variables to s3://<your-bucket-name>/variant-spark/output/hipster-importance-1.csv.
  • You can now download the result file from S3 using the AWS console or aws cli

Alternative Cluster Setup Options

In this example we use on-demand instances for both the master and the workers. If you prefer to use spot instances at the max price of say $0.1 you can add the --conf "bidPrice=0.1" option to start-cluster e.g.

vs-emr start-cluster --cluster-id-file /tmp/cluster-id --conf "bidPrice=0.1"

You can examine the cluster configuration and log files using AWS console or aws cli while it's running as well as up to seven days after its termination.

Distributions

variant-spark distributions are available on S3 (local to ap-southeast-2 region)

  • development builds in s3://variant-spark/unstable/<version>
  • release builds in: s3://variant-spark/stable/<version>

where <version> is:

  • <maj-ver>.<min-ver>/<maj-ver>.<min-ver>-<git-hash> for unstable builds
  • <maj-ver>.<min-ver>/<maj-ver>.<min-ver>.<rev> for stable builds

For example the distribution for release 0.0.2 is available at: s3://variant-spark/stable/0.0/0.0.2

The basic of running on EMR

To run variant-spark you need emr-5.7+ with Spark 2.1+ installed. To install variant spark in EMR cluster add the use the following bootstrap step:

s3://variant-spark/stable/<version>/bootstrap/install-variant-spark.sh --release-url s3://variant-spark/stable/<version>

This bootstrap action will deploy the desired version of variant-spark's assembly jar to the master node at /mnt/variant-spark/variant-spark_2.11-all.jar. This allows to run variant-spark in the client deploy mode.

To add an ApacheSpark step for variant spark use:

Type: Spark application  
Deploy mode: Client
Spark-submit options: --class au.csiro.variantspark.cli.VariantSparkApp
Jar location: /mnt/variant-spark/variant-spark_2.11-all.jar
Arguments: <variant-spark arguments, e.g: importance -if s3://au.csiro.pbdava.test/variant-spark/data/chr22_1000.vcf ...>

To add a custom EMR step for variant spark use:

Jar location: command-runner.jar
Argument: spark-submit --deploy-mode client --class au.csiro.variantspark.cli.VariantSparkApp /mnt/variant-spark/variant-spark_2.11-all.jar <variant-spark arguments>

Configuration

The default location for vs-emr configuration is .vs_emr/config.yaml. A different file can be passed with '--config' options.

The configuration file is in YAML format and consists of two sections:

  • default - with the default values for configuration parameters
  • profiles with a list of profiles that can be used to override the default values.

For example:

default:
  variantSparkReleaseUrl: "s3://variant-spark/unstable/0.1/0.1-134ed4e"
  instanceType: "m4.large"
  autoTerminate: true
  releaseLabel: "emr-5.8.0"
  bidPrice: 0.07
  ec2Attributes:
    AdditionalMasterSecurityGroups: ["sg-32323232", "sg-32323222"]
    KeyName: "mykey"
    SubnetId: "sub-323232"
    InstanceProfile: "EMR_EC2_DefaultRole"
  worker:
    instanceCount: 2
  conf:
    logBucketUri: "s3://my-bucket/logs/"
profiles:
  db-small:   #profile name "db-small" 
    bidPrice: null
    instanceType: "r4.2xlarge"
    worker:
      instanceCount: 2
  db-large:
    bidPrice: 0.6
    instanceType: "r4.8xlarge"
    worker: 
      instanceCount: 4

Profiles are activated with the --profile option:

vs-emr start-cluster --profile db-large

Multiple profiles can be activated together (when it makes senses) with repeated '--profile' options.

Specific values can be overridden with the --conf option (multiple entries are allowed) e.g.:

vs-emr start-cluster —-profile --profile db-large --conf bidPrice=1.0

Some most common options are also exposed as explicit parameters (e.g.: --worker-instances):

vs-emr start-cluster —-profile large --conf bidPrice=1.0 --worker-instances 20

For the list of available configuration parameters and how they are translated to aws emr create-cluster call please check the python/vs_emr/templates/spot-cluster.yaml file. It's a mustache template and hopefully it's self explanatory.

Running on sample datasets

As as a test you can can variant-spark on a small subset of 1000 Genome Project chromosome 22. The results (top 20 important variables are produced to stdout).

vs-emr start-cluster 
vs-emr submit-cmd --cluster-id <YOUR_CLUSTE_ID>  importance -if s3://variant-spark/datasets/examples/chr22_1000.vcf -ff s3://variant-spark/datasets/examples/chr22-labels.csv -fc 22_16051249 -v -rn 500 -rbs 20