Support setting shared memory for training job. (#1104)

Signed-off-by: Syulin7 <735122171@qq.com>
kubeflow · Jul 11, 2024 · 5e8b6dd · 5e8b6dd
1 parent a3a348c
commit 5e8b6dd
Show file tree

Hide file tree

Showing 12 changed files with 88 additions and 17 deletions.
diff --git a/Dockerfile.install b/Dockerfile.install
@@ -8,12 +8,12 @@ COPY . .
 
 RUN make
 
-RUN wget https://get.helm.sh/helm-v2.14.1-linux-amd64.tar.gz && \
-    tar -xvf helm-v2.14.1-linux-amd64.tar.gz && \
+RUN wget https://get.helm.sh/helm-v3.13.3-linux-amd64.tar.gz && \
+    tar -xvf helm-v3.13.3-linux-amd64.tar.gz && \
     mv linux-amd64/helm /usr/local/bin/helm && \
     chmod u+x /usr/local/bin/helm
 
-ENV K8S_VERSION v1.13.6
+ENV K8S_VERSION v1.28.4
 RUN curl -o /usr/local/bin/kubectl https://storage.googleapis.com/kubernetes-release/release/${K8S_VERSION}/bin/linux/amd64/kubectl && chmod +x /usr/local/bin/kubectl
 
 

diff --git a/Dockerfile.notebook.cpu b/Dockerfile.notebook.cpu
@@ -12,12 +12,12 @@ COPY . .
 
 RUN make
 
-RUN wget https://get.helm.sh/helm-v2.14.1-linux-amd64.tar.gz && \
-    tar -xvf helm-v2.14.1-linux-amd64.tar.gz && \
+RUN wget https://get.helm.sh/helm-v3.13.3-linux-amd64.tar.gz && \
+    tar -xvf helm-v3.13.3-linux-amd64.tar.gz && \
     mv linux-amd64/helm /usr/local/bin/helm && \
     chmod u+x /usr/local/bin/helm
 
-ENV K8S_VERSION v1.13.6
+ENV K8S_VERSION v1.28.4
 RUN curl -o /usr/local/bin/kubectl https://storage.googleapis.com/kubernetes-release/release/${K8S_VERSION}/bin/linux/amd64/kubectl && chmod +x /usr/local/bin/kubectl
 
 FROM $BASE_IMAGE

diff --git a/Dockerfile.notebook.kubeflow b/Dockerfile.notebook.kubeflow
@@ -11,12 +11,12 @@ COPY . .
 
 RUN make
 
-RUN wget https://get.helm.sh/helm-v2.14.1-linux-amd64.tar.gz && \
-    tar -xvf helm-v2.14.1-linux-amd64.tar.gz && \
+RUN wget https://get.helm.sh/helm-v3.13.3-linux-amd64.tar.gz && \
+    tar -xvf helm-v3.13.3-linux-amd64.tar.gz && \
     mv linux-amd64/helm /usr/local/bin/helm && \
     chmod u+x /usr/local/bin/helm
 
-ENV K8S_VERSION v1.13.6
+ENV K8S_VERSION v1.28.4
 RUN curl -o /usr/local/bin/kubectl https://storage.googleapis.com/kubernetes-release/release/${K8S_VERSION}/bin/linux/amd64/kubectl && chmod +x /usr/local/bin/kubectl
 
 FROM $BASE_IMAGE

diff --git a/charts/pytorchjob/templates/pytorchjob.yaml b/charts/pytorchjob/templates/pytorchjob.yaml
@@ -160,11 +160,11 @@ spec:
             name: {{ .name }}
           {{- end }}
           {{- end }}
-          {{- if .Values.shmSize }}
+          {{- if .Values.shareMemory }}
           - name: dshm
             emptyDir:
               medium: Memory
-              sizeLimit: {{ .Values.shmSize }}
+              sizeLimit: {{ .Values.shareMemory }}
           {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
@@ -324,7 +324,7 @@ spec:
               mountPath: "{{ $destPath }}"
             {{- end }}
             {{- end }}
-            {{- if .Values.shmSize }}
+            {{- if .Values.shareMemory }}
             - mountPath: /dev/shm
               name: dshm
             {{- end }}
@@ -452,11 +452,11 @@ spec:
             name: {{ .name }}
           {{- end }}
           {{- end }}
-          {{- if .Values.shmSize }}
+          {{- if .Values.shareMemory }}
           - name: dshm
             emptyDir:
               medium: Memory
-              sizeLimit: {{ .Values.shmSize }}
+              sizeLimit: {{ .Values.shareMemory }}
           {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
@@ -616,7 +616,7 @@ spec:
                 mountPath: "{{ $destPath }}"
               {{- end }}
               {{- end }}
-              {{- if .Values.shmSize }}
+              {{- if .Values.shareMemory }}
               - mountPath: /dev/shm
                 name: dshm
               {{- end }}

diff --git a/charts/pytorchjob/values.yaml b/charts/pytorchjob/values.yaml
@@ -12,7 +12,6 @@ rsyncImage: registry.cn-zhangjiakou.aliyuncs.com/acs/rsync:v3.1.0-aliyun
 # git sync image
 gitImage: registry.cn-zhangjiakou.aliyuncs.com/acs/git-sync:v3.3.5
 
-shmSize: 2Gi
 privileged: false
 
 useTensorboard: false

diff --git a/charts/tfjob/templates/tfjob.yaml b/charts/tfjob/templates/tfjob.yaml
@@ -204,6 +204,12 @@ spec:
               name: {{ .name }}
             {{- end }}
             {{- end }}
+            {{- if .Values.shareMemory }}
+            - name: dshm
+              emptyDir:
+                medium: Memory
+                sizeLimit: {{ .Values.shareMemory }}
+            {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
           - name: init-code
@@ -357,6 +363,10 @@ spec:
                   mountPath: "{{ $destPath }}"
                 {{- end }}
                 {{- end }}
+                {{- if .Values.shareMemory }}
+                - mountPath: /dev/shm
+                  name: dshm
+                {{- end }}
                 {{- if $dataDirs }}
                 {{- range $dataDirs }}
                 - mountPath: {{ .containerPath }}
@@ -530,6 +540,12 @@ spec:
               name: {{ .name }}
             {{- end }}
             {{- end }}
+            {{- if .Values.shareMemory }}
+            - name: dshm
+              emptyDir:
+                medium: Memory
+                sizeLimit: {{ .Values.shareMemory }}
+            {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
           - name: init-code
@@ -709,7 +725,10 @@ spec:
                   name: {{ .name }}
                 {{- end }}
                 {{- end }}
-
+                {{- if .Values.shareMemory }}
+                - mountPath: /dev/shm
+                  name: dshm
+                {{- end }}
 {{- end }}
 {{- if .Values.chief }}
 {{ .Values.chiefName | indent  4}}:
@@ -873,6 +892,12 @@ spec:
               name: {{ .name }}
             {{- end }}
             {{- end }}
+            {{- if .Values.shareMemory }}
+            - name: dshm
+              emptyDir:
+                medium: Memory
+                sizeLimit: {{ .Values.shareMemory }}
+            {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
           - name: init-code
@@ -1050,6 +1075,10 @@ spec:
                   name: {{ .name }}
                 {{- end }}
                 {{- end }}
+                {{- if .Values.shareMemory }}
+                - mountPath: /dev/shm
+                  name: dshm
+                {{- end }}
 {{- end }}
 {{- if .Values.evaluator }}
     Evaluator:
@@ -1175,6 +1204,12 @@ spec:
               name: {{ .name }}
             {{- end }}
             {{- end }}
+            {{- if .Values.shareMemory }}
+            - name: dshm
+              emptyDir:
+                medium: Memory
+                sizeLimit: {{ .Values.shareMemory }}
+            {{- end }}
           {{- if .Values.syncMode }}
           initContainers:
           - name: init-code
@@ -1330,4 +1365,8 @@ spec:
                   name: {{ .name }}
                 {{- end }}
                 {{- end }}
+                {{- if .Values.shareMemory }}
+                - mountPath: /dev/shm
+                  name: dshm
+                {{- end }}
 {{- end }}
diff --git a/pkg/apis/training/pytorchjob_builder.go b/pkg/apis/training/pytorchjob_builder.go
@@ -296,6 +296,13 @@ func (b *PytorchJobBuilder) TTLSecondsAfterFinished(ttl int32) *PytorchJobBuilde
 	return b
 }
 
+func (b *PytorchJobBuilder) ShareMemory(shm string) *PytorchJobBuilder {
+	if shm != "" {
+		b.args.ShareMemory = shm
+	}
+	return b
+}
+
 // Build is used to build the job
 func (b *PytorchJobBuilder) Build() (*Job, error) {
 	for key, value := range b.argValues {

diff --git a/pkg/apis/training/tfjob_builder.go b/pkg/apis/training/tfjob_builder.go
@@ -481,6 +481,13 @@ func (b *TFJobBuilder) TTLSecondsAfterFinished(ttl int32) *TFJobBuilder {
 	return b
 }
 
+func (b *TFJobBuilder) ShareMemory(shm string) *TFJobBuilder {
+	if shm != "" {
+		b.args.ShareMemory = shm
+	}
+	return b
+}
+
 func (b *TFJobBuilder) Build() (*Job, error) {
 	for key, value := range b.argValues {
 		b.AddArgValue(key, value)

diff --git a/pkg/apis/types/submit_pytorchjob.go b/pkg/apis/types/submit_pytorchjob.go
@@ -42,4 +42,7 @@ type SubmitPyTorchJobArgs struct {
 
 	// TrainingOperatorCRD compatible with training-operator crd.
 	TrainingOperatorCRD bool `yaml:"trainingOperatorCRD,omitempty"`
+
+	// ShareMemory Specifies the shared memory size
+	ShareMemory string `yaml:"shareMemory"`
 }
diff --git a/pkg/apis/types/submit_tfjob.go b/pkg/apis/types/submit_tfjob.go
@@ -87,6 +87,8 @@ type SubmitTFJobArgs struct {
 	StartingDeadlineSeconds int64 `yaml:"startingDeadlineSeconds,omitempty"`
 	// Defines the TTL for cleaning up finished TFJobs. Defaults to infinite.
 	TTLSecondsAfterFinished int32 `yaml:"ttlSecondsAfterFinished,omitempty"`
+	// ShareMemory Specifies the shared memory size
+	ShareMemory string `yaml:"shareMemory"`
 	// for common args
 	CommonSubmitArgs `yaml:",inline"`
 

diff --git a/pkg/argsbuilder/submit_pytorchjob.go b/pkg/argsbuilder/submit_pytorchjob.go
@@ -82,6 +82,7 @@ func (s *SubmitPytorchJobArgsBuilder) AddCommandFlags(command *cobra.Command) {
 	command.Flags().StringVar(&s.args.Memory, "memory", "", "the memory resource to use for the training, like 1Gi.")
 	command.Flags().DurationVar(&runningTimeout, "running-timeout", runningTimeout, "Specifies the duration since startTime during which the job can remain active before it is terminated(e.g. '5s', '1m', '2h22m').")
 	command.Flags().DurationVar(&ttlAfterFinished, "ttl-after-finished", ttlAfterFinished, "Defines the TTL for cleaning up finished PytorchJobs(e.g. '5s', '1m', '2h22m'). Defaults to infinite.")
+	command.Flags().StringVar(&s.args.ShareMemory, "share-memory", "2Gi", "the shared memory of each replica to run the job, default 2Gi.")
 
 	s.AddArgValue("running-timeout", &runningTimeout).
 		AddArgValue("ttl-after-finished", &ttlAfterFinished)
@@ -163,6 +164,12 @@ func (s *SubmitPytorchJobArgsBuilder) check() error {
 	if s.args.TTLSecondsAfterFinished < 0 {
 		return fmt.Errorf("--ttl-after-finished is invalid")
 	}
+	if s.args.ShareMemory != "" {
+		_, err := resource.ParseQuantity(s.args.ShareMemory)
+		if err != nil {
+			return fmt.Errorf("--share-memory is invalid")
+		}
+	}
 	return nil
 }
 

diff --git a/pkg/argsbuilder/submit_tfjob.go b/pkg/argsbuilder/submit_tfjob.go
@@ -170,6 +170,7 @@ func (s *SubmitTFJobArgsBuilder) AddCommandFlags(command *cobra.Command) {
 	command.Flags().StringArrayVar(&evaluatorSelectors, "evaluator-selector", []string{}, `assigning jobs with "Evaluator" role to some k8s particular nodes(this option would cover --selector), usage: "--evaluator-selector=key=value"`)
 	command.Flags().StringArrayVar(&psSelectors, "ps-selector", []string{}, `assigning jobs with "PS" role to some k8s particular nodes(this option would cover --selector), usage: "--ps-selector=key=value"`)
 	command.Flags().StringVar(&roleSequence, "role-sequence", "", `specify the tfjob role sequence,like: "Worker,PS,Chief,Evaluator" or "w,p,c,e"`)
+	command.Flags().StringVar(&s.args.ShareMemory, "share-memory", "2Gi", "the shared memory of each replica to run the job, default 2Gi.")
 
 	s.AddArgValue("worker-selector", &workerSelectors).
 		AddArgValue("chief-selector", &chiefSelectors).
@@ -342,6 +343,12 @@ func (s *SubmitTFJobArgsBuilder) check() error {
 	if s.args.TTLSecondsAfterFinished < 0 {
 		return fmt.Errorf("--ttl-after-finished is invalid")
 	}
+	if s.args.ShareMemory != "" {
+		_, err := resource.ParseQuantity(s.args.ShareMemory)
+		if err != nil {
+			return fmt.Errorf("--share-memory is invalid")
+		}
+	}
 	return nil
 }