Name: ecs-refarch-mixed-mode
Owner: AWS Samples
Description: Provides reference architecture and CloudFormation templates to power ECS using a mix of spot instances and ondemand instances with independent autoscaling
Created: 2017-11-21 05:03:17.0
Updated: 2017-12-07 18:05:24.0
Pushed: 2017-11-24 03:04:08.0
Homepage: null
Size: 142
Language: Shell
GitHub Committers
| User | Most Recent Commit | # Commits |
|---|
Other Committers
| User | Most Recent Commit | # Commits |
|---|
AWS Elastic Compute Cloud (EC2) Spot instances provide you with access to unused Amazon EC2 capacity at steep discounts relative to On-Demand prices. The low cost of Spot makes it an attractive option for running containerized workloads. The downside is that if the spot price for a particular instance type exceeds your set bid, Amazon EC2 will terminate the instance, stopping any containers running on it.
In order to overcome this problem, in this solution, we created a design pattern and codebase, in which a mix of spot instances and on demand instances can be used to serve critical workloads on Amazon EC2 Container Service (ECS) in conjunction with AWS AutoScaling. With this architecture, if spot instances are terminated due to change in market prices, the critical workload gets serves automatically by on-demand instances.
Architecture
In this solution, two ECS clusters are used, one made up of spot instances and another of on-demand instances. Separating the two clusters in this way gives an opportunity to harness cluster reservation metrics. If the cluster CPUReservation and/or MemoryReservation AWS CloudWatch metrics exceed 70%, autoScaling adds one instance to the respective cluster, thereby ensuring sufficient capacity to run the tasks.
Both the clusters run separate ECS Services which allows to maintain desired numbers of instances of a specific ECS Task definition. Autoscaling adds one more instance of the task definition by increasing desiredCount by one, if the service CPUUtilization and/or MemoryUtilization metrics exceeds 70% .
Both the ECS Services run behind the same Application Load Balancer For the cluster, which runs spot instances, AutoScaling places the bid for spot instances and if it gets fulfilled, the instance is added to the cluster and tasks will be placed to run on the instances using "spread" ECS Task placement strategy. If the market price exceeds bid price , the spot instances are terminated, resulting in increase of Memory Utilization/CPU Utilization Metrics due to increased load on tasks running on on-demand instances. Once Memory Utilization/CPU Utilization Metrics breaches the threshold (set to 70%) , ECS Service adds one more instance of the task definition at on-demand instances, which then increases the cluster CPUReservation and/or MemoryReservation. Finally, once cluster CPUReservation and/or MemoryReservation metrics breaches the threshold (set to 70%), AutoScaling adds one more instance (on-demand)to the cluster running .
In the same way, AutoScaling removes one instance from the respective clusters, if the cluster CPUReservation and/or MemoryReservation metrics dips below 20%, thereby removing excess capacity. A shell script (created at the userdata during the instance bootstrap) listens to the termination notices sent to the spot instances in order set the ECS container instance in DRAINING state. ECS Service AutoScaling decreases desiredCount by one, if the service CPUUtilization and/or MemoryUtilization metrics dips below 20%.
This example uses AWS Command Line Interface. Please follow instructions if you haven't installed AWS CLI. Your CLI configuration need PowerUserAccess and IAMFullAccess IAM policies associated with your credentials
--version
Output from above must yield AWS CLI version >= 1.11.37
clone https://github.com/aws-samples/ecs-refarch-mixed-mode.git
deploy
Here are the inputs required to launch CloudFormation templates:
* S3 Bucket : Enter S3 Bucket for storing your CloudFormation templates and scripts. This bucket must be in the same region where you wish to launch all the AWS resources created by this example.
* CloudFormation Stack Name : Enter CloudFormation Stack Name to create stacks
Sit back and relax until all the resources are created for you. After the templates are created, you can open ELB DNS URL to see the ECS Sample App
AWS Cloudwatch ECS Cluster CPUReservation and/or MemoryReservation metrics can be used to dynamically increase/decrease the instances running in an AutoScaling group. Separating the two ECS clusters based on the pricing options gives an opportunity to harness cluster CPUReservation and/or MemoryReservation metrics and thus increase or decrease the number of spot and on-demand instances independently. Setting up higher number of minimum instances of the task definition that the ECS Service will place at spot instance will optimize cost.
Count | AWS resources | — | — | 4 | AWS CloudFormation templates 1 | Amazon VPC (10.215.0.0/16) 2 | Amazon ECS Cluster 2 | Amazon ECS Service 1 | t2.small EC2 on-demand instance 1 | c3.large EC2 spot instance 1 | Application Load Balancer 1 | Application Load Balancer Target Groups
AWS CloudFormation is a free service; however, you are charged for the AWS resources you include in your stacks at the current rates for each. For more information about AWS pricing, go to the detail page for each product on http://aws.amazon.com.
Comments, feedback, and pull requests are always welcome.
This project is licensed under the Amazon Software License - see the LICENSE file for details.