GitHub - IBM/kube-mpi: High performance computing (HPC) depends on the distributed parallel programming API called the Message Passing Interface (MPI). Cloud native technologies depend on fault tolerance and resiliency normally associated with stateless applications. HPC applications are generally stateful and hence supporting programming models such as MPI have not been made available in public or private clouds that are enabled with Docker and/or Kubernetes. Kube-mpi is a prototype that provides high performance computing developers of simulation, distributed deep learning, and analytics applications a mechanism for building and deploying their applications in container orchestrations services such as kubernetes and IBM's IBM Cloud private.

IBM / kube-mpi Public

High performance computing (HPC) depends on the distributed parallel programming API called the Message Passing Interface (MPI). Cloud native technologies depend on fault tolerance and resiliency normally associated with stateless applications. HPC applications are generally stateful and hence supporting programming models such as MPI have not b…

6 stars 11 forks Activity

Star

Notifications

1 branch 0 tags

Code

6 commits

Failed to load latest commit information.

README.md

These directories provide examples of a kube-mpi controller for OpenMPI. They have been tested on OpenShift but should be portable to any kubernetes cluster. The examples support hostNetwork: True | False. In both cases, the mpirun command substitutes ssh with kssh (alias of kubectl) to start the ORTED in pods across the cluster. Each application tasks waits until all tasks are scheduled by the default kubernetes scheduler. At that time, an mpi controller executes mpirun.