Mpi Operator Versions Save

Changes since v0.4.0

• Features:
  • Add support for MPICH (#562, @sheevy)
  • Field runLauncherAsWorker allows to add the launcher pod into the hostfile as a worker (#612, @kuizhiqing)
  • Add PodGroup minResources calculation for volcano integration (#566, @lowang-bh)
• Bug fixes:
  • Fix panic when using PodGroups and PriorityClasses (#561, @tenzen-y)
  • Fix installation of mpijob Python module (#579, @vsoch)
  • Fix hostfile when jobs in different namespaces have the same name (#622, @kuizhiqing)
• Clean ups:
  • Upgrade k8s libraries to v1.29 (#633, @tenzen-y)
  • Fail the mpi-operator binary if access to API is denied (#619, @emsixteeen)

Acknowledgments

Thank you to all the contributors (in no particular order): @sheevy @alculquicondor @terrytangyuan @tenzen-y @kuizhiqing @lowang-bh @vsoch @emsixteeen @wang-mask @benash @yeahdongcn @xhejtman @pheianox @lianghao208

Changes since 0.3.0

• Breaking changes
  • Removed v1 operator. If you want to use MPIJob v1, you can use the training-operator.
• Support for suspending semantics. Third party controllers can leverage the suspend field to implement queuing and preemption for an MPIJob.
• Support for the coscheduling plugins of the scheduler-plugins.
• The operator supports multi-architecture (amd64, aarch64, and ppc64le).
• Bug fixes
  • Fix support for elastic Horovod.

Acknowledgements

Special thanks to @tenzen-y for multiple contributions. Thank you to all the contributors (in no particular order): @mimowo @adilhusain-s @davidLif @ArangoGutierrez @shaowei-su @ggaaooppeenngg @pugangxa @HeGaoYuan @Dimss @alculquicondor @terrytangyuan

Release v0.3.0

• Scalability improvements
  • Worker start up no longer issues requests to kube-apiserver.
  • Dropped kubectl-delivery init container, reducing stress on kube-apiserver.
• Support for Intel MPI.
• Support for runPolicy (ttlSecondsAfterFinish, activeDeadlineSeconds, backoffLimit) by using a k8s Job for the launcher.
• Samples for plain MPI applications.
• Production readiness improvements:
  • Increased coverage throughout unit, integration and E2E tests.
  • More robust API validation.
  • Revisited v2beta1 MPIJob API.
  • Using fully-qualified label names, in consistency with other kubeflow operators.

Enhancements

• Added support for RH OCP4.1 and RH OCP4.2
• Added additional installation methods
  • Using kustomize and kubeflow/manifests
  • Using Helm Chart
• Added support for Go Modules and removed vendor directories
• Added default ephemeral storage for init container
• Overwrite NVIDIA env vars to avoid using GPUs on launcher
• Added health check and callbacks around various leader election phases
• Honor user-specified worker command
• Exposed main container name as a configurable field
• Added RunPolicy to MPIJobSpec that reuses kubeflow/common spec
• Allow to specify the name of the gang scheduler and priority for pod group
• Added error log when pod spec does not have any containers
• Switched to use distroless images
• Refactored the kubectl-delivery to improve the launcher performance
• Added Prometheus metrics for job monitoring
• Added experimental version of v1 MPIJob controller and APIs
• Support Volcano as a scheduler
• Switched to use pods for launcher job and statefulset workers
• Switched to use klog for logging
• More consistent labels with other Kubeflow operators

Fixes

• Fixed nil pointer exceptions that could accidentally restart the pod
• Updated status to running only when launcher is active and all workers are ready
• Fixed the incorrect namespace for initializing informers and endpoints of leader election
• Fixed issue in v1 controller's CRD existence check

Documentation

• Added the list of adopters
• Added roadmap document
• Revamped contributing guidelines
• Added MPIJob API reference page on Kubeflow website
• Added a blog post for an introduction to MPI Operator and its industry adoption
• Added a CPU-only example
• Added licenses used by the dependencies

• Added default resource requirements for init container
• Merged multiple deployment configuration files into a single YAML file
• Switched to use JobStatus from kubeflow/common
• Launcher and workers are now created together

• Switch Docker files and examples to use v1alpha2 MPI Operator.

API Changes

• Add v1alpha2 version of the MPI Operator with more consistent API spec with other Kubeflow operators
• Support ActiveDeadlineSeconds in MPIJobSpec
• Support custom resource types other than GPUs
• Remove launcherOnMaster field

Enhancements

• Support gang scheduling
• Add StartTime and CompletionTime in job status
• Add leader election
• Switch to use pod group for gang scheduling
• Add example on Apache MXNet using v1alpha1 version of the MPI Operator

Initial release of the MPI Operator.