Why does ECS binpack strategy on memory always scale up a new EC2 instance despite available resources?

Question

I'm using ECS with an Auto Scaling group as a capacity provider to deploy tasks on EC2 (t3.micro, 2 vCPU, 1 GB memory). I've set the task placement strategy to binpack based on memory. However, I've noticed that ECS always scales up a new EC2 instance to deploy a new task, even if an existing instance has enough CPU and memory available. Consequently, there is only one task per EC2 instance. I expect that all tasks should be placed on a single EC2 instance if it has sufficient memory.

Here are some actions I've already checked:

1. Port conflict: The network mode is set to awsvpc in the ECS task definition, so each task gets its own ENI, which prevents port conflicts.
2. EC2 storage: Each EC2 instance has a storage size of 30GB (EBS GP3). My container is a nginx-based web app with 1 MB of static files, so the storage is more than sufficient for running multiple containers.

The following configurations might be related to this issue,

Capacity provider configurations
```yaml
capacity provider: autoscaling group
base: 0
weight: 100
target capacity: 100%
managed instance draining: true
managed instance scaling: true
scale in protection: false
```

ECS service configurations
```yaml
desired count: 1
placement strategy:
  type: binpack
  field: memory
scheduling strategy: REPLICA
service connect:
  enabled: true
  namespace: my_namespace
  services:
    - port name: web
      discovery name: hello
      client aliases:
        - port: 80
          dns name: hello
  
```

ECS task definition
```yaml
network mode: awsvpc
CPU: 256
Memory: 128
container definitions:
  - name: web
    image: nginx
    port mappings:
      - name: web
        container port: 80
        protocol: tcp
```

Any insights or suggestions?

**Additional information** \
I changed the instance type from t3.micro to t3.small (2 vCPUs, 2 GB memory) and deployed 4 ECS tasks. The ECS cluster autoscaled up 2 EC2 instances, placing 2 tasks on each instance.

Syd · Answer

Hi,

I would assume it is perhaps due to the baseline constraints for T series instances which can burst upto 100% but normally operates at baseline. (Cannot corroborate this with AWS docs unfortunately)

You might want to check on a small non T type instance.

Update: I tried launching 8 tasks on c7g.medium instance. It launched 4 tasks on single node before scaling another node for the other 4. A slight improvement as compared to T.x instances but definitely not as expected for binpack. Maybe some more underlying factors determining binpack behaviour

--Syd

Why does ECS binpack strategy on memory always scale up a new EC2 instance despite available resources?

相關內容