Merge pull request #28 from converged-computing/allow-dynamic-command

feat: allow variadic tasks/nodes

Author: vsoch

api/v1alpha1/statemachine_types.go

@@ -165,6 +165,10 @@ type JobConfig struct {
 	// +optional
 	Nodes int32 `json:"nodes,omitempty"`
 
+	// Number of tasks
+	// +optional
+	Tasks int32 `json:"tasks,omitempty"`
+
 	// Cores per task per job
 	// 6 frontier / 3 summit / 5 on lassen (vsoch: this used to be 6 default)
 	// +kubebuilder:default=3

config/crd/bases/state-machine.converged-computing.org_statemachines.yaml

@@ -101,6 +101,10 @@ spec:
                             because we assume the initial data is bad.
                             RetryFailure
                           type: boolean
+                        tasks:
+                          description: Number of tasks
+                          format: int32
+                          type: integer
                         walltime:
                           description: Walltime (in string format) for the job
                           type: string

examples/dist/state-machine-operator-dev.yaml

@@ -109,6 +109,10 @@ spec:
                             because we assume the initial data is bad.
                             RetryFailure
                           type: boolean
+                        tasks:
+                          description: Number of tasks
+                          format: int32
+                          type: integer
                         walltime:
                           description: Walltime (in string format) for the job
                           type: string

examples/test/lammps-metrics.yaml

@@ -5,23 +5,22 @@ metadata:
 spec:
   manager:
     pullPolicy: Never
-    interactive: true
   workflow:
     completed: 10
 
     events:
       # Two studies:
-      # If lammps runtime is > 3:00 add a node (do with autoscaler)
+      # If lammps runtime is > 30 add a node (do with autoscaler)
       # what is the base case
       - metric: mean.lammps.duration
-        when: ">= 360"
+        when: ">= 30"
         action: grow
-        repetitions: 3
+        repetitions: 2
         # Require checks between before doing again
         # TODO check if this is working correctly.
         # also check the .get()
         backoff: 3
-        maxSize: 10
+        maxSize: 5
 
   cluster:
     maxSize: 2
@@ -34,16 +33,14 @@ spec:
     image: rockylinux:9
     script: echo This is a setup for lammps
 
-  # Note that this step always fails and we never make it to C
-  # This should end the workflow early
   - name: lammps
     properties:
       minicluster: "yes"
     config:
       nodes: 4
     image: ghcr.io/converged-computing/metric-lammps-cpu:zen4
     workdir: /opt/lammps/examples/reaxff/HNS/
-    script: lmp -v x 4 -v y 4 -v z 4 -in ./in.reaxff.hns -nocite
+    script: flux run -N $nodes lmp -v x 4 -v y 4 -v z 4 -in ./in.reaxff.hns -nocite
 
   - name: job_c
     config:

internal/controller/manager/jobs/jobs.go

@@ -26,6 +26,9 @@ cd -
 jobid="{{ jobid }}"
 outpath="{{ workdir }}"
 registry="{{ registry }}"
+{% if cores_per_task %}cores_per_task={{ cores_per_task }}{% endif %}
+{% if nodes %}nodes={{ nodes }}{% endif %}
+{% if tasks %}tasks={{ tasks }}{% endif %}
 {% if pull %}pull_tag={{ pull }}{% endif %}
 {% if push %}push_tag={{ push }}{% endif %}
 

internal/controller/manager/jobs/templates/components.sh

@@ -3,6 +3,7 @@ config:
   nnodes:           {{ if .Job.Config.Nodes }}{{ .Job.Config.Nodes }}{{ else }}1{{ end }}
   cores_per_task:   {{ if .Job.Config.CoresPerTask }}{{ .Job.Config.CoresPerTask }}{{ else }}6{{ end }}
   ngpus:            {{ .Job.Config.Gpus }}
+  {{ if .Job.Config.Tasks }}tasks:         {{ .Job.Config.Tasks }}{{ end }}
   {{ if .Job.Config.Walltime }}walltime:         '{{ .Job.Config.Walltime }}'{{ end }}
   # Kubernetes specific settings
   {{ if .Job.Config.GPULabel }}gpulabel:         {{ .Job.Config.GPULabel }}{{ end }}

python/state_machine_operator/manager/manager.py

@@ -472,11 +472,14 @@ def check_metrics(self, job):
 
     def trigger_grow(self, trigger, step_name, value):
         """
-        Trigger the job to grow
+        Trigger the job to grow.
+
+        Note that this is more of a static grow - subsequent jobs will be given more
+        nodes. It doesn't give currently running jobs more.
         """
         previous = self.workflow.jobs[step_name]["config"]["nnodes"]
         max_size = trigger.action.max_size
-        if max_size >= previous + 1:
+        if previous + 1 >= max_size:
             LOGGER.info(
                 f"Grow triggered: {trigger.action.metric} {trigger.when} ({value}), already >= max size {max_size}"
             )
@@ -523,9 +526,13 @@ def trigger_workflow_action(self, trigger, step_name, value):
             )
             self.complete_workflow()
 
+        # TODO: think about use case / mechanism for dynamic grow.
+        # It would likely need to be requested by the application.
+        # Static grow increases subsequent nodes for a job
         if trigger.action.name == "grow":
             self.trigger_grow(trigger, step_name, value)
 
+        # Static shrink decreases subsequent nodes for a job
         if trigger.action.name == "shrink":
             self.trigger_shrink(trigger, step_name, value)
 

python/state_machine_operator/tracker/kubernetes/tracker.py

@@ -315,6 +315,7 @@ def submit_minicluster_job(self, step, jobid):
             "workingDir": step.workdir,
             "name": container_name,
             "pullAlways": pull_always,
+            "launcher": True,
             "volumes": {
                 step.name: {
                     "configMapName": step.name,
@@ -570,6 +571,7 @@ def create_step(self, jobid):
             cores_per_task=self.ncores,
             gpus=self.ngpus,
             workdir=workdir,
+            tasks=self.tasks,
         )
 
         if "script" in self.job_desc:
@@ -584,6 +586,9 @@ def create_step(self, jobid):
                 "push": self.push_to,
                 "registry": self.registry_host,
                 "plain_http": self.registry_plain_http,
+                "nodes": step.nodes,
+                "cores_per_task": self.ncores,
+                "tasks": step.tasks,
             }
             step.script = Template(self.job_desc["script"]).render(**kwargs)
 

python/state_machine_operator/tracker/tracker.py

@@ -124,6 +124,12 @@ def type(self):
     def nnodes(self):
         return int(self.config.get("nnodes", 1))
 
+    @property
+    def tasks(self):
+        tasks = self.config.get("tasks")
+        if tasks is not None:
+            return int(tasks)
+
     @property
     def ncores(self):
         return int(self.config.get("cores_per_task", 1))

python/state_machine_operator/tracker/types.py

@@ -28,6 +28,7 @@ class JobSetup:
     nodes: int
     cores_per_task: int
     script: str = None
+    tasks: int = None
     walltime: str = None
     gpus: int = 0
     workdir: str = None