Memory and vCPU considerations for AWS Batch on Amazon EKS

In AWS Batch on Amazon EKS, you can specify the resources that are made available to a container. For example, you can specify requests or limits values for vCPU and memory resources.

The following are constraints for specifying vCPU resources:

Example: job definitions

The following AWS Batch on Amazon EKS job definition configures soft vCPU shares. This lets AWS Batch on Amazon EKS use all of the vCPU capacity for the instance type. However, if there are other jobs running, the job is allocated a maximum of 2 vCPUs. Memory is limited to 2 GB.

{
    "jobDefinitionName": "MyJobOnEks_Sleep",
    "type": "container",
    "eksProperties": {
        "podProperties": {
            "containers": [
                {
                    "image": "public.ecr.aws/amazonlinux/amazonlinux:2",
                    "command": ["sleep", "60"],
                    "resources": {
                        "requests": {
                            "cpu": "2",
                            "memory": "2048Mi"
                        }
                    }
                }
            ]
        }
    }
}

The following AWS Batch on Amazon EKS job definition has a request value of 1 and allocates a maximum of 4 vCPUs to the job.

{
    "jobDefinitionName": "MyJobOnEks_Sleep",
    "type": "container",
    "eksProperties": {
        "podProperties": {
            "containers": [
                {
                    "image": "public.ecr.aws/amazonlinux/amazonlinux:2",
                    "command": ["sleep", "60"],
                    "resources": {
                        "requests": {
                            "cpu": "1"
                        },
                        "limits": {
                            "cpu": "4",
                            "memory": "2048Mi"
                        }
                    }
                }
            ]
        }
    }
}

The following AWS Batch on Amazon EKS job definition sets a vCPU limits value of 1 and a memory limits value of 1 GB.

{
    "jobDefinitionName": "MyJobOnEks_Sleep",
    "type": "container",
    "eksProperties": {
        "podProperties": {
            "containers": [
                {
                    "image": "public.ecr.aws/amazonlinux/amazonlinux:2",
                    "command": ["sleep", "60"],
                    "resources": {
                        "limits": {
                            "cpu": "1",
                            "memory": "1024Mi"
                        }
                    }
                }
            ]
        }
    }
}

When AWS Batch translates an AWS Batch on Amazon EKS job into an Amazon EKS pod, AWS Batch copies thelimits value to the requests value. This is if a requests value isn't specified. When you submit the preceding example job definition, the pod spec is as follows.

apiVersion: v1
kind: Pod
...
spec:
  ...
  containers:
    - command:
        - sleep
        - 60
      image: public.ecr.aws/amazonlinux/amazonlinux:2
      resources:
        limits:
          cpu: 1
          memory: 1024Mi
        requests:
          cpu: 1
          memory: 1024Mi
      ...

Node CPU and memory reservations

AWS Batch relies on the default logic of the bootstrap.sh file for vCPU and memory reservations. For more information about the bootstrap.sh file, see bootstrap.sh. When you size your vCPU and memory resources, consider the examples that follow.

Example: Node CPU reservation

The CPU reservation value is calculated in millicores using the total number of vCPUs that are available to the instance.

Using the preceding values, the following is true:

In this example, there are 1930 (calculated 2000-70) millicore vCPU units available to run jobs on a c5.large instance. Suppose your job requires 2 (2*1000 m) vCPU units, the job doesn't fit on a single c5.large instance. However, a job that requires 1.75 vCPU units fits.

Example: Node memory reservation

The memory reservation value is calculated in mebibytes using the following:

The formula to calculate the allocatable memory is as follows: (instance_capacity_in_MiB) - (11 * (maximum_number_of_pods)) - 255 - (HardEvictionLimit value.)).

A c5.large instance supports up to 29 pods. For an 8 GB c5.large instance with a HardEvictionLimit value of 100 MiB, the allocatable memory is 7074 MiB. This is calculated in the following way: (7748 - (11 * 29) -255 -100) = 7074 MiB. In this example, an 8,192 MiB job doesn't fit on this instance even though it's an 8 gibibyte (GiB) instance.

DaemonSets

When you use DaemonSets, consider the following: