MPICH support

Makefile

@@ -23,6 +23,7 @@ BASE_IMAGE_SSH_PORT?=2222
 IMG_BUILDER=docker
 PLATFORMS ?= linux/amd64
 INTEL_PLATFORMS ?= linux/amd64
+MPICH_PLATFORMS ?= linux/amd64
 LD_FLAGS_V2=" \
     -X '${REPO_PATH}/pkg/version.GitSHA=${GitSHA}' \
     -X '${REPO_PATH}/pkg/version.Built=${Date}'   \
@@ -71,6 +72,7 @@ test: bin/envtest scheduler-plugins-crd
 test_e2e: export TEST_MPI_OPERATOR_IMAGE=${IMAGE_NAME}:${RELEASE_VERSION}
 test_e2e: export TEST_OPENMPI_IMAGE=mpioperator/mpi-pi:${RELEASE_VERSION}-openmpi
 test_e2e: export TEST_INTELMPI_IMAGE=mpioperator/mpi-pi:${RELEASE_VERSION}-intel
+test_e2e: export TEST_MPICH_IMAGE=mpioperator/mpi-pi:${RELEASE_VERSION}-mpich
 test_e2e: bin/kubectl kind helm images test_images dev_manifest scheduler-plugins-chart
 	go test -v ./test/e2e/...
 
@@ -108,6 +110,9 @@ test_images:
 	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(INTEL_PLATFORMS) --build-arg BASE_LABEL=${RELEASE_VERSION} -t mpioperator/intel:${RELEASE_VERSION} build/base -f build/base/intel.Dockerfile
 	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(INTEL_PLATFORMS) -t mpioperator/intel-builder:${RELEASE_VERSION} build/base -f build/base/intel-builder.Dockerfile
 	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(INTEL_PLATFORMS) --build-arg BASE_LABEL=${RELEASE_VERSION} -t mpioperator/mpi-pi:${RELEASE_VERSION}-intel examples/v2beta1/pi -f examples/v2beta1/pi/intel.Dockerfile
+	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(MPICH_PLATFORMS) --build-arg BASE_LABEL=${RELEASE_VERSION} -t mpioperator/mpich:${RELEASE_VERSION} build/base -f build/base/mpich.Dockerfile
+	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(MPICH_PLATFORMS) -t mpioperator/mpich-builder:${RELEASE_VERSION} build/base -f build/base/mpich-builder.Dockerfile
+	${IMG_BUILDER} build $(BUILD_ARGS) --platform $(MPICH_PLATFORMS) --build-arg BASE_LABEL=${RELEASE_VERSION} -t mpioperator/mpi-pi:${RELEASE_VERSION}-mpich examples/v2beta1/pi -f examples/v2beta1/pi/mpich.Dockerfile
 
 .PHONY: tidy
 tidy:

README.md

@@ -218,6 +218,12 @@ For a sample that uses Intel MPI, see:
 cat examples/pi/pi-intel.yaml
 ```
 
+For a sample that uses MPICH, see:
+
+```bash
+cat examples/pi/pi-mpich.yaml
+```
+
 ## Exposed Metrics
 
 | Metric name | Metric type | Description | Labels |

build/base/intel.Dockerfile

@@ -22,5 +22,5 @@ RUN apt update \
         intel-oneapi-mpi \
     && rm -rf /var/lib/apt/lists/*
 
-COPY intel-entrypoint.sh /entrypoint.sh
+COPY entrypoint.sh /entrypoint.sh
 ENTRYPOINT ["/entrypoint.sh"]

build/base/mpich-builder.Dockerfile

@@ -0,0 +1,7 @@
+FROM debian:bullseye as builder
+
+RUN apt update \
+    && apt install -y --no-install-recommends \
+        g++ \
+        libmpich-dev \
+    && rm -rf /var/lib/apt/lists/*

build/base/mpich.Dockerfile

@@ -0,0 +1,12 @@
+ARG BASE_LABEL
+
+FROM mpioperator/base:${BASE_LABEL}
+
+RUN apt update \
+    && apt install -y --no-install-recommends \
+        dnsutils \
+        mpich \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY entrypoint.sh /entrypoint.sh
+ENTRYPOINT ["/entrypoint.sh"]

deploy/v2beta1/mpi-operator.yaml

@@ -58,10 +58,11 @@ spec:
               mpiImplementation:
                 default: OpenMPI
                 description: MPIImplementation is the MPI implementation. Options
-                  are "OpenMPI" (default) and "Intel".
+                  are "OpenMPI" (default), "Intel" and "MPICH".
                 enum:
                 - OpenMPI
                 - Intel
+                - MPICH
                 type: string
               mpiReplicaSpecs:
                 additionalProperties:

examples/v2beta1/pi/Dockerfile

@@ -5,7 +5,6 @@ FROM mpioperator/openmpi-builder:${BASE_LABEL} as builder
 COPY pi.cc /src/pi.cc
 RUN mpic++ /src/pi.cc -o /pi
 
-
 FROM mpioperator/openmpi:${BASE_LABEL}
 
-COPY --from=builder /pi /home/mpiuser/pi
+COPY --from=builder /pi /home/mpiuser/pi

examples/v2beta1/pi/README.md

@@ -19,9 +19,15 @@ For Intel MPI:
 docker build -t mpi-pi . -f intel.Dockerfile
 ```
 
+For MPICH:
+
+```bash
+docker build -t mpi-pi . -f mpich.Dockerfile
+```
+
 ## Create MPIJob
 
-Modify `pi.yaml` (for OpenMPI) or `pi-intel.yaml` (for Intel MPI) to set up the
+Modify `pi.yaml` (for OpenMPI), `pi-intel.yaml` (for Intel MPI) or `pi-mpich.yaml` (for MPICH) to set up the
 image name from your own registry.
 
 Then, run:

examples/v2beta1/pi/intel.Dockerfile

@@ -7,4 +7,4 @@ RUN bash -c "source /opt/intel/oneapi/setvars.sh && mpicxx /src/pi.cc -o /pi"
 
 FROM mpioperator/intel:${BASE_LABEL}
 
-COPY --from=builder /pi /home/mpiuser/pi
+COPY --from=builder /pi /home/mpiuser/pi

examples/v2beta1/pi/mpich.Dockerfile

@@ -0,0 +1,10 @@
+ARG BASE_LABEL
+
+FROM mpioperator/mpich-builder:${BASE_LABEL} as builder
+
+COPY pi.cc /src/pi.cc
+RUN mpic++ /src/pi.cc -o /pi
+
+FROM mpioperator/mpich:${BASE_LABEL}
+
+COPY --from=builder /pi /home/mpiuser/pi

examples/v2beta1/pi/pi-mpich.yaml

@@ -0,0 +1,54 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: pi
+spec:
+  slotsPerWorker: 1
+  runPolicy:
+    cleanPodPolicy: Running
+  sshAuthMountPath: /home/mpiuser/.ssh
+  mpiImplementation: MPICH
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          containers:
+          - image: mpioperator/mpi-pi:mpich
+            imagePullPolicy: Always
+            name: mpi-launcher
+            securityContext:
+              runAsUser: 1000
+            args:
+            - mpirun
+            - -n
+            - "2"
+            - /home/mpiuser/pi
+            resources:
+              limits:
+                cpu: 1
+                memory: 1Gi
+    Worker:
+      replicas: 2
+      template:
+        spec:
+          containers:
+          - image: mpioperator/mpi-pi:mpich
+            imagePullPolicy: Always
+            name: mpi-worker
+            securityContext:
+              runAsUser: 1000
+            command:
+            args:
+            - /usr/sbin/sshd
+            - -De
+            - -f
+            - /home/mpiuser/.sshd_config
+            readinessProbe:
+              tcpSocket:
+                port: 2222
+              initialDelaySeconds: 2
+            resources:
+              limits:
+                cpu: 1
+                memory: 1Gi

manifests/base/kubeflow.org_mpijobs.yaml

@@ -35,10 +35,11 @@ spec:
               mpiImplementation:
                 default: OpenMPI
                 description: MPIImplementation is the MPI implementation. Options
-                  are "OpenMPI" (default) and "Intel".
+                  are "OpenMPI" (default), "Intel" and "MPICH".
                 enum:
                 - OpenMPI
                 - Intel
+                - MPICH
                 type: string
               mpiReplicaSpecs:
                 additionalProperties:

pkg/apis/kubeflow/v2beta1/default_test.go

@@ -38,7 +38,7 @@ func TestSetDefaults_MPIJob(t *testing.T) {
 				},
 			},
 		},
-		"base defaults overridden": {
+		"base defaults overridden (intel)": {
 			job: MPIJob{
 				Spec: MPIJobSpec{
 					SlotsPerWorker: newInt32(10),
@@ -66,6 +66,34 @@ func TestSetDefaults_MPIJob(t *testing.T) {
 				},
 			},
 		},
+		"base defaults overridden (mpich)": {
+			job: MPIJob{
+				Spec: MPIJobSpec{
+					SlotsPerWorker: newInt32(10),
+					RunPolicy: RunPolicy{
+						CleanPodPolicy:          NewCleanPodPolicy(CleanPodPolicyRunning),
+						TTLSecondsAfterFinished: newInt32(2),
+						ActiveDeadlineSeconds:   newInt64(3),
+						BackoffLimit:            newInt32(4),
+					},
+					SSHAuthMountPath:  "/home/mpiuser/.ssh",
+					MPIImplementation: MPIImplementationMPICH,
+				},
+			},
+			want: MPIJob{
+				Spec: MPIJobSpec{
+					SlotsPerWorker: newInt32(10),
+					RunPolicy: RunPolicy{
+						CleanPodPolicy:          NewCleanPodPolicy(CleanPodPolicyRunning),
+						TTLSecondsAfterFinished: newInt32(2),
+						ActiveDeadlineSeconds:   newInt64(3),
+						BackoffLimit:            newInt32(4),
+					},
+					SSHAuthMountPath:  "/home/mpiuser/.ssh",
+					MPIImplementation: MPIImplementationMPICH,
+				},
+			},
+		},
 		"launcher defaults": {
 			job: MPIJob{
 				Spec: MPIJobSpec{

pkg/apis/kubeflow/v2beta1/swagger.json

@@ -322,7 +322,7 @@
       ],
       "properties": {
         "mpiImplementation": {
-          "description": "MPIImplementation is the MPI implementation. Options are \"OpenMPI\" (default) and \"Intel\".",
+          "description": "MPIImplementation is the MPI implementation. Options are \"OpenMPI\" (default), \"Intel\" and \"MPICH\".",
           "type": "string"
         },
         "mpiReplicaSpecs": {

pkg/apis/kubeflow/v2beta1/types.go

@@ -155,8 +155,8 @@ type MPIJobSpec struct {
 	SSHAuthMountPath string `json:"sshAuthMountPath,omitempty"`
 
 	// MPIImplementation is the MPI implementation.
-	// Options are "OpenMPI" (default) and "Intel".
-	// +kubebuilder:validation:Enum:=OpenMPI;Intel
+	// Options are "OpenMPI" (default), "Intel" and "MPICH".
+	// +kubebuilder:validation:Enum:=OpenMPI;Intel;MPICH
 	// +kubebuilder:default:=OpenMPI
 	MPIImplementation MPIImplementation `json:"mpiImplementation,omitempty"`
 }
@@ -177,6 +177,7 @@ type MPIImplementation string
 const (
 	MPIImplementationOpenMPI MPIImplementation = "OpenMPI"
 	MPIImplementationIntel   MPIImplementation = "Intel"
+	MPIImplementationMPICH   MPIImplementation = "MPICH"
 )
 
 // JobStatus represents the current observed state of the training Job.

pkg/apis/kubeflow/validation/validation.go

@@ -35,7 +35,8 @@ var (
 
 	validMPIImplementations = sets.NewString(
 		string(kubeflow.MPIImplementationOpenMPI),
-		string(kubeflow.MPIImplementationIntel))
+		string(kubeflow.MPIImplementationIntel),
+		string(kubeflow.MPIImplementationMPICH))
 
 	validRestartPolicies = sets.NewString(
 		string(common.RestartPolicyNever),

pkg/controller/mpi_job_controller.go

@@ -199,6 +199,16 @@ var (
 			Value: "-o ConnectionAttempts=10",
 		},
 	}
+	mpichEnvVars = []corev1.EnvVar{
+		{
+			Name:  "HYDRA_HOST_FILE",
+			Value: fmt.Sprintf("%s/%s", configMountPath, hostfileName),
+		},
+		{
+			Name:  "HYDRA_LAUNCH_EXTRA_ARGS",
+			Value: "-o ConnectionAttempts=10",
+		},
+	}
 	nvidiaDisableEnvVars = []corev1.EnvVar{
 		{Name: "NVIDIA_VISIBLE_DEVICES"},
 		{Name: "NVIDIA_DRIVER_CAPABILITIES"},
@@ -588,8 +598,9 @@ func (c *MPIJobController) syncHandler(key string) error {
 				return err
 			}
 		}
-		if mpiJob.Spec.MPIImplementation == kubeflow.MPIImplementationIntel {
-			// The Intel implementation requires workers to communicate with the
+		if mpiJob.Spec.MPIImplementation == kubeflow.MPIImplementationIntel ||
+			mpiJob.Spec.MPIImplementation == kubeflow.MPIImplementationMPICH {
+			// The Intel and MPICH implementations require workers to communicate with the
 			// launcher through its hostname. For that, we create a Service which
 			// has the same name as the launcher's hostname.
 			_, err := c.getOrCreateService(mpiJob, newLauncherService(mpiJob))
@@ -1201,7 +1212,7 @@ func newConfigMap(mpiJob *kubeflow.MPIJob, workerReplicas int32) *corev1.ConfigM
 		switch mpiJob.Spec.MPIImplementation {
 		case kubeflow.MPIImplementationOpenMPI:
 			buffer.WriteString(fmt.Sprintf("%s%s-%d.%s.%s.svc slots=%d\n", mpiJob.Name, workerSuffix, i, workersService, mpiJob.Namespace, slots))
-		case kubeflow.MPIImplementationIntel:
+		case kubeflow.MPIImplementationIntel, kubeflow.MPIImplementationMPICH:
 			buffer.WriteString(fmt.Sprintf("%s%s-%d.%s.%s.svc:%d\n", mpiJob.Name, workerSuffix, i, workersService, mpiJob.Namespace, slots))
 		}
 	}
@@ -1429,6 +1440,8 @@ func (c *MPIJobController) newLauncherPodTemplate(mpiJob *kubeflow.MPIJob) corev
 			Name:  intelMPISlotsEnv,
 			Value: slotsStr,
 		})
+	case kubeflow.MPIImplementationMPICH:
+		container.Env = append(container.Env, mpichEnvVars...)
 	}
 
 	container.Env = append(container.Env,