Amazon Web Services¶
Amazon Web Services (AWS) provides a flexible cloud based environment for running analyses. Cloud approaches offer the ability to perform analyses at scale with no investment in local hardware. They also offer full programmatic control over the environment, allowing bcbio to automate the entire setup, run and teardown process.
bcbio-vm provides a wrapper
around bcbio-nextgen that automates interaction with AWS and Docker.
bcbio_vm.py also cleans up the command line
usage to make it more intuitive and provides a superset of functionality
bcbio_nextgen.py. bcbio-vm uses Elasticluster to build a cluster on AWS with
an encrypted NFS mounted drive and an optional Lustre shared filesystem.
We’re phasing out this approach to cloud support in bcbio. Instead of building this directly into bcbio, we’ll now be using Common Workflow Language (CWL) support. Runners like Cromwell can directly interface and run with cloud services and we’ll document these approaches as they’re tested and available. To use cloud integration immediately, we’d suggest using the simplified ansible based approach which spins up single multicore machines for running.
bcbio_vm.py provides the automation to start up and administer remote bcbio
runs on AWS. This only requires a local installation of the python wrapper code,
not any of the Docker containers or biological data, which will all get
installed on AWS. The easier way to install is using conda with an isolated
wget http://repo.continuum.io/miniconda/Miniconda2-latest-Linux-x86_64.sh bash Miniconda2-latest-Linux-x86_64.sh -b -p ~/install/bcbio-vm/anaconda ~/install/bcbio-vm/anaconda/bin/conda install --yes -c conda-forge -c bioconda bcbio-nextgen-vm ln -s ~/install/bcbio-vm/anaconda/bin/bcbio_vm.py /usr/local/bin/bcbio_vm.py
We support both Linux and Mac OSX as clients for running remote AWS bcbio clusters.
The easiest way to organize AWS projects is using an analysis folder inside an
S3 bucket. Create a bucket and folder for your analysis and
upload fastq, BAM and, optionally, a region BED file. Bucket names should
include only lowercase letters, numbers and hyphens (
-) to conform to
S3 bucket naming restrictions
and avoid issues with resolution of SSL keys. You can create buckets and upload
files using the
AWS S3 web console,
the AWS cli client or specialized tools
You will also need a template file describing the type of run to do and a CSV file mapping samples in the bucket to names and any other metadata. See the Automated sample configuration docs for more details about these files. Also upload both of these files to S3.
With that in place, prepare and upload the final configuration to S3 with:
bcbio_vm.py template s3://your-project/your-analysis/template.yaml s3://your-project/your-analysis/name.csv
This will find the input files in the
s3://your-project/your-analysis bucket, associate
fastq and BAM files with the right samples, and add a found BED files as
variant_regions in the configuration. It will then upload the final
configuration back to S3 as
s3://your-project/your-analysis/name.yaml, which you can run
directly from a bcbio cluster on AWS. By default, bcbio will use the us-east S3
region, but you can specify a different region in the s3 path to the
We currently support human analysis with both the GRCh37 and hg19 genomes. We can also add additional genomes as needed by the community and generally welcome feedback and comments on reference data support.
We’re not able to automatically install some useful tools in pre-built docker
containers due to licensing restrictions. Variant calling with GATK requires a
manual download from the GATK download site for academic users. Commercial
users need a license for GATK and for somatic calling with muTect. To make these jars available,
upload them to the S3 bucket in a
jars directory. bcbio will automatically
include the correct GATK and muTect directives during your run. Alternatively,
you can also manually specify the path to the jars using a global
resources section of your input sample YAML file:
resources: gatk: jar: s3://bcbio-syn3-eval/jars/GenomeAnalysisTK.jar
As with sample YAML scripts, specify a different region with an
@ in the
The first time running bcbio on AWS you’ll need to setup permissions, VPCs and local configuration files. We provide commands to automate all these steps and once finished, they can be re-used for subsequent runs. To start you’ll need to have an account at Amazon and your Access Key ID and Secret Key ID from the AWS security credentials page. These can be IAM credentials instead of root credentials as long as they have administrator privileges. Make them available to bcbio using the standard environmental variables:
export AWS_ACCESS_KEY_ID=your_access_key export AWS_SECRET_ACCESS_KEY=your_secret_key
With this in place, two commands setup your elasticluster and AWS environment to
run a bcbio cluster. The first creates public/private keys, a bcbio IAM user,
and sets up an elasticluster config in
bcbio_vm.py aws iam --region=us-east-1
The second configures a VPC to host bcbio:
bcbio_vm.py aws vpc --region=us-east-1
aws vpc command is idempotent and can run multiple times if you change or
remove parts of the infrastructure. You can also rerun the
aws iam command,
but if you’d like to generate a new elasticluster configuration file
~/.bcbio/elasticluster/config) add the recreate flag:
bcbio_vm.py aws iam
--recreate. This generates a new set of IAM credentials and public/private
keys. These are only stored in the
~/.bcbio directory so you need to fully
recreate them if you delete the old ones.
Running a cluster¶
Following this setup, you’re ready to run a bcbio cluster on AWS. We start from a standard Ubuntu AMI, installing all software for bcbio and the cluster as part of the boot process.
To configure your cluster run:
bcbio_vm.py aws config edit
This dialog allows you to define the cluster size and machine resources you’d like to use. The defaults only have small instances to prevent accidentally starting an expensive run. If you’re planning a run with less than 32 cores, do not use a cluster and instead run directly on a single machine using one of the large r3 or c3 instances.
This script also sets the size of the encrypted NFS-mounted drive, which you can use to store processing data when running across a distributed cluster. At scale, you can replace this with a Lustre shared filesystem. See below for details on launching and attaching a Lustre filesystem to a cluster.
To ensure everything is correctly configured, run:
bcbio_vm.py aws info
When happy with your setup, start the cluster with:
bcbio_vm.py aws cluster start
The cluster will take five to ten minutes to start and be provisioned. If you encounter any
intermittent failures, you can rerun the cluster configuration step with
bcbio_vm.py aws cluster setup or the bcbio-specific installation with
bcbio_vm.py aws cluster bootstrap.
Elasticluster mounts the
/encrypted directory as a NFS share available
across all of the worker machines. You can use this as a processing directory
for smaller runs but for larger runs may need a scalable distributed file
system. bcbio supports using
Intel Cloud Edition for Lustre (ICEL)
to set up a Lustre scratch filesystem on AWS.
Subscribe to ICEL in the Amazon Marketplace.
By default, the Lustre filesystem will be 2TB and will be accessible to all hosts in the VPC. Creation takes about ten minutes and can happen in parallel while elasticluster sets up the cluster. Start the stack:
bcbio_vm.py aws icel create
If you encounter any intermittent failures when installing collectl plugin, that means lustre server is created successfully, you can rerun the lustre configuration step with
bcbio_vm.py aws icel create --setup. If you had any failure creating the lustre server before the collectl plugin installation, you should stop it, and try again.
Once the ICEL stack and elasticluster cluster are both running, mount the filesystem on the cluster:
bcbio_vm.py aws icel mount
The cluster instances will reboot with the Lustre filesystem mounted.
Running an analysis¶
To run the analysis, connect to the head node with:
bcbio_vm.py aws cluster ssh
Create your project directory and link the global bcbio configuration file in there with:
NFS file system (no Lustre):
mkdir /encrypted/your-project cd !$ && mkdir work && cd work
Lustre file system:
sudo mkdir /scratch/cancer-dream-syn3-exome sudo chown ubuntu !$ cd !$ && mkdir work && cd work
If you started a single machine, run with:
bcbio_vm.py run -n 8 s3://your-project/your-analysis/name.yaml
-n argument should be the number of cores on the machine.
To run on a full cluster:
bcbio_vm.py ipythonprep s3://your-project/your-analysis/name.yaml slurm cloud -n 60 sbatch bcbio_submit.sh
Where 60 is the total number of cores to use across all the worker nodes. Of
your total machine cores, allocate 2 for the base bcbio_vm script and IPython
controller instances. The SLURM workload manager
distributes jobs across your cluster on a queue called
slurm-PID.out file in the work directory contains the current status of the
sacct_std provides the status of jobs on the cluster. If you are
new to SLURM, here is a summary of useful
On successful completion, bcbio uploads the results of the analysis back into your s3
bucket and folder as
s3://your-project/your-analysis/final. You can now cleanup the cluster and
Graphing resource usage¶
AWS runs include automatic monitoring of resource usage with
collectl. bcbio_vm uses collectl statistics
to plot CPU, memory, disk and network usage during each step of a run. To
prepare resource usage plots after finishing an analysis, first copy the
bcbio-nextgen.log file to your local computer. Either use
bcbio_vm.py elasticluster sftp bcbio to copy from the work directory on AWS
/encrypted/your-project/work/log/bcbio-nextgen.log) or transfer it from the
output S3 bucket (
If your run worked cleanly you can use the log input file directly. If you had
failures and restarts, or would only like to graph part of the run, you can edit
the timing steps. Run
grep Timing bcbio-nextgen.log > your-run.txt to get
the timing steps only, then edit as desired.
Retrieve the collectl statistics from the AWS cluster and prepare the resource usage graphs with:
bcbio_vm.py graph bcbio-nextgen.log
By default the collectl stats will be in
monitoring/collectl and plots in
monitoring/graphs based on the above log timeframe. If you need to re-run
plots later after shutting the cluster down, you can use the none cluster flag
bcbio_vm.py graph bcbio-nextgen.log --cluster none.
If you’d like to run graphing from a local non-AWS run, such as a local HPC cluster,
bcbio_vm.py graph bcbio-nextgen.log --cluster local instead.
For convenience, there’s a “serialize” flag (‘-s’) that saves the dataframe used for plotting. In order to explore the data and extract specific datapoints or zoom, one could just deserialize the ouput like a python pickle file:
import cPickle as pickle with gzip.open(“./monitoring/collectl_info.pickle.gz”, “rb”) as decomp:collectl_info = pickle.load(decomp) data, hardware, steps = collectl_info, collectl_info, collectl_info
And plot, slice, zoom it in an jupyter notebook using matplotlib, [highcharts](https://github.com/arnoutaertgeerts/python-highcharts).
In addition to plots, the summarize_timing.py utility script prepares a summary table of run times per step.
The bcbio Elasticluster and Lustre integration can spin up a lot of AWS resources. You’ll be paying for these by the hour so you want to clean them up when you finish running your analysis. To stop the cluster:
bcbio_vm.py aws cluster stop
To remove the Lustre stack:
bcbio_vm.py aws icel stop
Double check that all instances have been properly stopped by looking in the AWS console.
Experienced elasticluster users
can edit the configuration files themselves. bcbio provides a small wrapper
that automatically reads and writes these configurations to avoid users needing
to understand elasticluster internals, but all functionality is fully available.
~/.bcbio/elasticluster/config file to change parameters. You can
also see the latest example configuration.
in the bcbio-vm GitHub repository for more details on the other available options.