This Sample CDK App builds an Amazon ECS GPU Cluster on AWS using Spot instances, and to run Folding@Home. Folding@Home is deployed via the raykrueger/FoldingAtHomeContainer container. The container is configured to consume the GPU on each instance in the cluster.
Folding@home, in their own words...
"Folding@home (FAH or F@h) is a distributed computing project for simulating protein dynamics, including the process of protein folding and the movements of proteins implicated in a variety of diseases. It brings together citizen scientists who volunteer to run simulations of protein dynamics on their personal computers. Insights from this data are helping scientists to better understand biology, and providing new opportunities for developing therapeutics."
COVID-19 is ravaging the world right now. The folks at NVIDIA put out a call-to-arms for gamers to fire up their GPUs and put them to work on the new GPU COVID-19 workloads that are ready to run on the Folding@home network.
This inspired me to put this to work on the cloud. If you, or your employer, are able to spare some compute dollars for this critical need, it is easy to fire this up and contribute. The cluster uses Spot nodes for incredible savings over on-demand costs.
For me, personally, I also needed an excuse to force myself to do something in CDK. That being said, this is the first thing I've built with CDK. Which means I'm definitely open to feedback on how it could be done better. For example, I'm beginning to realize that CDK "apps" don't share well.
This Sample Application is based on two depdencies namely the Folding on ECS AWS CDK Library, which is hosted on npm, and the Folding at Home Container which is hosted on Docker Hub.
This sample application is meant to be run as-is, or cloned and modified to your needs.
git clone https://github.com/raykrueger/FoldingOnECS.git
cd FoldingOnECS
npm install
npm run build
cdk deploy
To clean up, you can simply delete the stack in your cloudformation console, or you can have CDK do it.
cdk destroy
This runs three g4dn.xlarge instances as spot requests by default. The spot request is for the base price of that instance in us-east-2, which is $0.52 an hour. Adjust for your region, and budget as needed.
In us-east-2 spot pricing, at the time of this writing, for g4dn.xlarge is around $0.16 on average. Which means a single g4dn.xlarge will cost about $117 a month. The default configuration of three nodes will cost around $351 a month. There would be additional costs for CloudWatch logs and network egress as well. I have not yet calculated an estimate for that.
To calculate the avage spot price for the past 30 days in a region the following command will work on a mac.
aws ec2 describe-spot-price-history --region us-east-2 --instance-type g4dn.xlarge --start-time $(date -v-30d "+%Y-%m-%d") --product-descriptions Linux/UNIX | jq '[.SpotPriceHistory[].SpotPrice | tonumber] | length as $l | add / $l'
For linux the date command needs to be changed.
aws ec2 describe-spot-price-history --region us-east-2 --instance-type g4dn.xlarge --start-time $(date -d "30 days ago" "+%Y-%m-%d") --product-descriptions Linux/UNIX | jq '[.SpotPriceHistory[].SpotPrice | tonumber] | length as $l | add / $l'
npm run buildcompile typescript to jsnpm run watchwatch for changes and compilenpm run testperform the jest unit testscdk deploydeploy this stack to your default AWS account/regioncdk diffcompare deployed stack with current statecdk synthemits the synthesized CloudFormation template