schedulers: set /dev/shm size for docker based schedulers to bypass 64M default

torchx/schedulers/aws_batch_scheduler.py

@@ -85,10 +85,10 @@ def _role_to_node_properties(idx: int, role: Role) -> Dict[str, object]:
         cpu = 1
     reqs.append({"type": "VCPU", "value": str(cpu)})
 
-    mem = resource.memMB
-    if mem <= 0:
-        mem = 1000
-    reqs.append({"type": "MEMORY", "value": str(mem)})
+    memMB = resource.memMB
+    if memMB <= 0:
+        memMB = 1000
+    reqs.append({"type": "MEMORY", "value": str(memMB)})
 
     if resource.gpu > 0:
         reqs.append({"type": "GPU", "value": str(resource.gpu)})
@@ -130,6 +130,11 @@ def _role_to_node_properties(idx: int, role: Role) -> Dict[str, object]:
         "image": role.image,
         "environment": [{"name": k, "value": v} for k, v in role.env.items()],
         "resourceRequirements": reqs,
+        "linuxParameters": {
+            # To support PyTorch dataloaders we need to set /dev/shm to larger
+            # than the 64M default.
+            "sharedMemorySize": memMB,
+        },
         "logConfiguration": {
             "logDriver": "awslogs",
         },

torchx/schedulers/docker_scheduler.py

@@ -269,7 +269,11 @@ def _submit_dryrun(
                     }
                 resource = replica_role.resource
                 if resource.memMB >= 0:
-                    c.kwargs["mem_limit"] = f"{int(resource.memMB)}m"
+                    # To support PyTorch dataloaders we need to set /dev/shm to
+                    # larger than the 64M default.
+                    c.kwargs["mem_limit"] = c.kwargs[
+                        "shm_size"
+                    ] = f"{int(resource.memMB)}m"
                 if resource.cpu >= 0:
                     c.kwargs["nano_cpus"] = int(resource.cpu * 1e9)
                 if resource.gpu > 0:

torchx/schedulers/kubernetes_scheduler.py

@@ -172,6 +172,7 @@ def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod
         V1Volume,
         V1HostPathVolumeSource,
         V1PersistentVolumeClaimVolumeSource,
+        V1EmptyDirVolumeSource,
     )
 
     requests = {}
@@ -189,8 +190,21 @@ def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod
         requests=requests,
     )
 
-    volumes = []
-    volume_mounts = []
+    # To support PyTorch dataloaders we need to set /dev/shm to larger than the
+    # 64M default so we mount an unlimited sized tmpfs directory on it.
+    SHM_VOL = "dshm"
+    volumes = [
+        V1Volume(
+            name=SHM_VOL,
+            empty_dir=V1EmptyDirVolumeSource(
+                medium="Memory",
+            ),
+        ),
+    ]
+    volume_mounts = [
+        V1VolumeMount(name=SHM_VOL, mount_path="/dev/shm"),
+    ]
+
     for i, mount in enumerate(role.mounts):
         mount_name = f"mount-{i}"
         if isinstance(mount, BindMount):

torchx/schedulers/test/aws_batch_scheduler_test.py

@@ -112,6 +112,9 @@ def test_submit_dryrun(self) -> None:
                                     {"type": "MEMORY", "value": "3000"},
                                     {"type": "GPU", "value": "4"},
                                 ],
+                                "linuxParameters": {
+                                    "sharedMemorySize": 3000,
+                                },
                                 "logConfiguration": {"logDriver": "awslogs"},
                                 "mountPoints": [
                                     {
@@ -154,6 +157,9 @@ def test_submit_dryrun(self) -> None:
                                     {"type": "MEMORY", "value": "3000"},
                                     {"type": "GPU", "value": "4"},
                                 ],
+                                "linuxParameters": {
+                                    "sharedMemorySize": 3000,
+                                },
                                 "logConfiguration": {"logDriver": "awslogs"},
                                 "mountPoints": [
                                     {

torchx/schedulers/test/docker_scheduler_test.py

@@ -100,6 +100,7 @@ def test_submit_dryrun(self) -> None:
                             "torchx.pytorch.org/version": "0.1.2dev0",
                         },
                         "mem_limit": "3000m",
+                        "shm_size": "3000m",
                         "name": "app_name_42-trainer-0",
                         "hostname": "app_name_42-trainer-0",
                         "nano_cpus": int(2e9),

torchx/schedulers/test/kubernetes_scheduler_test.py

@@ -118,6 +118,7 @@ def test_role_to_pod(self) -> None:
             V1Volume,
             V1VolumeMount,
             V1HostPathVolumeSource,
+            V1EmptyDirVolumeSource,
         )
 
         app = _test_app()
@@ -148,11 +149,15 @@ def test_role_to_pod(self) -> None:
             resources=resources,
             ports=[V1ContainerPort(name="foo", container_port=1234)],
             volume_mounts=[
+                V1VolumeMount(
+                    name="dshm",
+                    mount_path="/dev/shm",
+                ),
                 V1VolumeMount(
                     name="mount-0",
                     mount_path="/dst",
                     read_only=True,
-                )
+                ),
             ],
         )
         want = V1Pod(
@@ -161,6 +166,12 @@ def test_role_to_pod(self) -> None:
                 restart_policy="Never",
                 service_account_name="srvacc",
                 volumes=[
+                    V1Volume(
+                        name="dshm",
+                        empty_dir=V1EmptyDirVolumeSource(
+                            medium="Memory",
+                        ),
+                    ),
                     V1Volume(
                         name="mount-0",
                         host_path=V1HostPathVolumeSource(
@@ -272,11 +283,16 @@ def test_submit_dryrun(self) -> None:
               memory: 3000M
               nvidia.com/gpu: '4'
           volumeMounts:
+          - mountPath: /dev/shm
+            name: dshm
           - mountPath: /dst
             name: mount-0
             readOnly: true
         restartPolicy: Never
         volumes:
+        - emptyDir:
+            medium: Memory
+          name: dshm
         - hostPath:
             path: /src
           name: mount-0
@@ -289,6 +305,7 @@ def test_volume_mounts(self) -> None:
             V1Volume,
             V1VolumeMount,
             V1PersistentVolumeClaimVolumeSource,
+            V1EmptyDirVolumeSource,
         )
 
         role = specs.Role(
@@ -302,6 +319,12 @@ def test_volume_mounts(self) -> None:
         self.assertEqual(
             pod.spec.volumes,
             [
+                V1Volume(
+                    name="dshm",
+                    empty_dir=V1EmptyDirVolumeSource(
+                        medium="Memory",
+                    ),
+                ),
                 V1Volume(
                     name="mount-0",
                     persistent_volume_claim=V1PersistentVolumeClaimVolumeSource(
@@ -313,11 +336,15 @@ def test_volume_mounts(self) -> None:
         self.assertEqual(
             pod.spec.containers[0].volume_mounts,
             [
+                V1VolumeMount(
+                    name="dshm",
+                    mount_path="/dev/shm",
+                ),
                 V1VolumeMount(
                     name="mount-0",
                     mount_path="/dst",
                     read_only=True,
-                )
+                ),
             ],
         )