mpi-operator/proposals/scalable-robust-operator.md

289 lines
9.9 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# A scalable and robust operator
Authors: @alculquicondor, @ahg-g
- [Motivation](#motivation)
- [Goals](#goals)
- [Background](#background)
- [Design](#design)
- [Alternatives Considered](#alternatives-considered)
- [Appendix](#appendix-prototype-objects)
## Motivation
A scalable MPI setup on Kubernetes is important for:
- Running jobs from multiple users in a single cluster.
- High-performance computing jobs that can require a big number of workers.
A robust MPI setup should be tolerant to failures, implementing retries while
keeping track of failures.
## Goals
- Allow driver-to-worker control to scale by removing kube-apiserver from
the communication channel.
- Reduce the complexity of the controller by relaying on Kubernetes workload
APIs for Pod creation and management.
## Background
The original design of the operator can be found as a
[proposal to the kubeflow community](https://github.com/kubeflow/community/blob/master/proposals/mpi-operator-proposal.md).
The latest release includes a v1alpha2 API and controller.
A v1 is under development. Something to highlight in this new version is the
replacement of the Job and StatefulSet by plain Pods, with the intent of
[tracking running Pods](https://github.com/kubeflow/mpi-operator/issues/201#issuecomment-827837831).
An MPIJob CRD describes the Job. Important fields include:
- The workers template
- The number of workers
- The launcher template, which should have a `mpirun` command.
The images are expected to have the MPI implementation binaries (such as
OpenMPI, Intel MPI or MPICH) the users MPI executable.
A controller processes the MPIJob, starting a Job with the following steps:
1. Creates ConfigMap, which contains:
- A script `kubexec.sh` that wraps `kubectl exec` and is used in replacement
of `ssh`. This script, before executing the command provided by `mpirun`,
transfers a file containing a mapping of pod names to IPs and appends it to
the workers `/etc/hosts`.
Note: The v1 controller no longer copies the pod-to-IP mapping file. The
OpenMPI implementation does the [routing](https://www.open-mpi.org/faq/?category=tcp#tcp-routability-1.3).
- The `hostfile` for `mpirun`, listing the worker pod names and number of
slots (which could be the number of CPUs or GPUs). This list is built
programmatically.
2. Creates a ServiceAccount+Role+RoleBinding for the launcher, which allow it
to:
- get/list/watch on Pods
- do pods/exec
This allows the launcher Pod to obtain details of the worker Pods and start
the process managers on them.
3. If configured, it creates a Volcano PodGroup
4. Creates a StatefulSet for workers (plain pods in v1). The Pod template
includes:
- mounts for the ConfigMap.
- `sleep` as command.
5. Creates launcher Job (plain pod in v1). The Pod template includes:
- An init container, `kubectl-delivery`, described below.
- Environment variables for:
- replacing `ssh` for `kubexec.sh`
- the `hostfile` location
- Volumes for:
- The ConfigMap
- Sharing files from `kubectl-delivery`
The launcher Job, as previously mentioned, contains an init container:
`kubectl-delivery`. This is a Kubernetes controller that watches pods. It does
the following:
1. Copy kubectl from the image into the volume shared with the main container.
2. Wait for all Pods in the hostfile to be running
3. Generates a file mapping pod name to IP, in `/etc/hosts` format.
To update the status of an MPIJob, the controller uses the status of the
launcher Job. That is, when the launcher Job fails or succeeds, its status is
copied to the MPIJob. In v1, it bases the status on the termination condition of
the Pod.
### Analysis
The above architecture for MPI Jobs puts a lot of pressure in the
`kube-apiserver`. The load increases with the number of workers in a job
and with the number of jobs in a cluster.
The reasons for this are:
- Due to the use of `kubectl exec`, every worker spawn goes through
kube-apiserver. `mpirun` starts a daemon in each worker
(like [`orted`](https://www.open-mpi.org/doc/v3.0/man1/orted.1.php)).
This process handles the worker-to-worker communication, which happens
without the intervention of kube-apiserver. However, the `exec` connection
stays up for control during the entirety of the job.
- The `kubectl-delivery` controller does a full cache sync to be able to watch
Pods. This startup penalty increases with the number of pods in the cluster
and has to be paid for every job. The API calls also cause additional
stress on the apiserver.
- The launcher role has to include a list of all the pods in the job.
Potentially, the object might not be able to accommodate jobs with immense
number of workers.
Another problem is that the v1 controller doesnt implement launcher pod
retries, although there are plans to. So the MPIJob behaves like a plain Pod in
this version.
## Design
In order to address the problems of the existing architecture, we propose the
following changes:
- **The use of `ssh` instead of `kubectl exec`.**
This would avoid any pod-to-pod communication happening through apiserver and
doesnt require giving execution permissions at the namespace level.
This can be achieved like this:
- The controller generates a single key and share it with the launcher and
worker pods through a Secret.
- The launcher and workers mount the Secret and set appropriate file
permissions.
- The workers run an SSH server instead of `sleep`.
- When encrypted communication is not a requirement, users have the choice to
use `rsh` for faster communication.
- **The use of stable hostnames and a headless Service for the workers**
- This removes the need to query Pod IPs, as the Pods can discover each other
through DNS resolution. The hostfile can be generated statically by the
controller using the stable hostnames.
- Starting with k8s 1.22, we can use
[Indexed Jobs with stable hostnames](https://git.k8s.io/enhancements/keps/sig-apps/2214-indexed-job)
to delegate the pod management to Kubernetes. Additionally, this will give
us robust failure tracking so that we can give users control over retry
limits.
In the meantime, we can continue using plain Pods.
- Caveat 1: The Job controller doesnt report the number of running Pods.
Instead, it reports active Pods, which include running and pending
(scheduling or starting). But Kubernetes SIG Apps is
[open to add a status field for running Pods](https://kubernetes.slack.com/archives/C18NZM5K9/p1619549430067400).
- Caveat 2: Horovod supports elastic workers, but the Kubernetes Job
doesnt support changes to the completions field. This can be supported
starting from 1.23. In the meantime, we can replicate the behavior by
creating a new Job and doing Pod adoption.
- For Intel MPI and MPICH, we also need a headless Service to front the launcher,
because workers communicate back to the launcher using its hostname.
- **Revert the use of the Job API for the launcher.**
- The Job controller handles retries when the launcher or any of the workers fail.
- Caveat 1 also applies: The Job controller doesnt report if the Pod is running.
We can continue watching Pods in the meantime.
- With the above changes, **the following objects can be removed**:
- The ServiceAccount+Role+RoleBinding for the launcher.
- The `kubectl-delivery` init container in the launcher, as there is no need
to obtain IPs, speeding up startup time.
## Alternatives Considered
TBD from discussions
## Appendix: Prototype objects
It uses a StatefulSet in place of Indexed Jobs, as they are still an alpha
feature in Kubernetes.
```yaml
apiVersion: v1
kind: ConfigMap
metadata:
name: mpi-config
data:
hostfile: |
mpi-workers-0.mpi-workers slots=3
mpi-workers-1.mpi-workers slots=3
mpi-workers-2.mpi-workers slots=3
```
```yaml
apiVersion: v1
kind: Secret
type: kubernetes.io/ssh-auth
data:
ssh-privatekey: PRIVATE_KEY
ssh-publickey: PUBLIC_KEY
```
```yaml
apiVersion: batch/v1
kind: Job
metadata:
name: mpi-launcher
spec:
template:
spec:
restartPolicy: OnFailure
containers:
- name: driver
image: '<USER_IMAGE>'
args:
- 'mpirun'
- '-np'
- '9'
- '<USER_EXECUTABLE>'
env:
- name: 'OMPI_MCA_orte_keep_fqdn_hostnames'
value: 'true'
- name: 'OMPI_MCA_orte_default_hostfile'
value: '/home/mpiuser/config/hostfile'
volumeMounts:
- name: ssh-auth
mountPath: /mnt/ssh
readOnly: true
- name: mpi-config
mountPath: /home/mpiuser/config
readOnly: true
volumes:
- name: mpi-config
configMap:
name: mpi-config
- name: ssh-auth
secret:
secretName: ssh-auth
items:
- key: ssh-privatekey
path: id_rsa
- key: ssh-publickey
path: id_rsa.pub
- key: ssh-publickey
path: authorized_keys
```
```yaml
apiVersion: v1
kind: Service
metadata:
name: mpi-workers
spec:
clusterIP: None
selector:
app: mpi-workers
```
```yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: mpi-workers
spec:
selector:
matchLabels:
app: mpi-workers
serviceName: mpi-workers
replicas: 3
podManagementPolicy: Parallel
template:
metadata:
labels:
app: mpi-workers
spec:
containers:
- name: worker
image: '<USER_IMAGE>'
volumeMounts:
- name: ssh-auth
mountPath: /mnt/ssh
readOnly: true
- name: mpi-config
mountPath: /home/mpiuser/config
readOnly: true
volumes:
- name: mpi-config
configMap:
name: mpi-config
- name: ssh-auth
secret:
secretName: ssh-auth
items:
- key: ssh-privatekey
path: id_rsa
- key: ssh-publickey
path: id_rsa.pub
- key: ssh-publickey
path: authorized_keys
```