droot · Oct 15, 2026 · Oct 15, 2026 · Oct 17, 2026 · Oct 17, 2026 · Oct 17, 2026
diff --git a/k8s/deploy/base/api-server.yaml b/k8s/deploy/base/api-server.yaml
@@ -53,6 +53,11 @@ spec:
               key: BASE_MODEL
         - name: OPEN_RL_TMP_DIR
           value: "/mnt/shared/open-rl"
+        # vLLM's compile cache, shared by every sampler in the fleet: the first
+        # sampler of a model on a GPU type compiles (about a minute), the rest
+        # load the result in a second. Unset it to keep a per-node hostPath instead.
+        - name: OPEN_RL_VLLM_CACHE_DIR
+          value: "/mnt/shared/open-rl/vllm-cache"
         # vLLM sampling enabled on cluster.
         - name: SAMPLING_BACKEND
           value: "vllm"

diff --git a/k8s/deploy/base/kustomization.yaml b/k8s/deploy/base/kustomization.yaml
@@ -9,6 +9,7 @@ resources:
   - rbac.yaml
   - redis.yaml
   - api-server.yaml
+  - worker-prepull.yaml
 
 generatorOptions:
   disableNameSuffixHash: true

diff --git a/k8s/deploy/base/worker-prepull.yaml b/k8s/deploy/base/worker-prepull.yaml
@@ -0,0 +1,44 @@
+# Pulls the worker image onto every node the scheduler may use, at deploy time
+# instead of inside the first job. Measured on GKE (runs/2026-10-15_cold_start):
+# a worker pull is 92 s on an H100 node with local SSD and 169 s on an L4 node
+# with a pd-balanced boot disk, and a cold job pays it twice, trainer and
+# sampler. The container only sleeps; it holds no GPU. The image reference is
+# rewritten by the kustomize images transformer like every other OpenRL image,
+# so a tag bump in the bundle re-pulls on every node as the DaemonSet rolls.
+apiVersion: apps/v1
+kind: DaemonSet
+metadata:
+  name: open-rl-worker-prepull
+  labels:
+    app: open-rl-worker-prepull
+spec:
+  selector:
+    matchLabels:
+      app: open-rl-worker-prepull
+  updateStrategy:
+    type: RollingUpdate
+    rollingUpdate:
+      # Every node pulls at once; there is nothing to protect by rolling slowly.
+      maxUnavailable: 100%
+  template:
+    metadata:
+      labels:
+        app: open-rl-worker-prepull
+    spec:
+      nodeSelector:
+        openrl.io/enabled: "true"
+      tolerations:
+      - key: "nvidia.com/gpu"
+        operator: "Exists"
+        effect: "NoSchedule"
+      containers:
+      - name: prepull
+        image: ghcr.io/gke-labs/open-rl/server:latest
+        imagePullPolicy: IfNotPresent
+        command: ["sleep", "infinity"]
+        resources:
+          requests:
+            cpu: "10m"
+            memory: "16Mi"
+          limits:
+            memory: "64Mi"
diff --git a/runs/2026-10-15_cold_start/A-0.5b-l4-cold/monitoring.md b/runs/2026-10-15_cold_start/A-0.5b-l4-cold/monitoring.md
@@ -0,0 +1,12 @@
+| minute ending (UTC) | read MiB/s | read ops/s | write MiB/s |
+|---|---|---|---|
+| 2026-10-15T03:04:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:05:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:06:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:07:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:08:00Z | 0.0 | 0 | 1.7 |
+| 2026-10-15T03:09:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:10:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:11:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:12:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:13:00Z | 0.0 | 0 | 0.0 |
diff --git a/runs/2026-10-15_cold_start/A-0.5b-l4-cold/timeline.md b/runs/2026-10-15_cold_start/A-0.5b-l4-cold/timeline.md
@@ -0,0 +1,111 @@
+# Cold start timeline: A-0.5b-l4-cold
+
+t=0 is the client's `create_model` call at 2026-10-15T03:04:15.892+00:00 UTC. Offsets are seconds after t=0; a negative offset is before it.
+
+## trainer: `orw-lora-qwen-qwen2-5-0-5b-0-trainer` on `gke-open-rl-dra-l4-trainer-2x-ed601d8b-lvw8`
+
+| Mark | Clock (UTC) | t+ s | Self-reported s | Source |
+|---|---|---|---|---|
+| trainer.workload_created | 03:04:16.000 | 0.1 |  | Workload creationTimestamp (apiserver) |
+| trainer.pod_created | 03:04:16.000 | 0.1 |  | pod creationTimestamp (apiserver) |
+| trainer.workload_first_seen_by_poller | 03:04:16.578 | 0.7 |  | workloads.tsv poller (laptop clock, 1 s) |
+| trainer.workload_reports_pod | 03:04:16.578 | 0.7 |  | workloads.tsv poller: status.podName set |
+| trainer.pod_Initialized | 03:04:17.000 | 1.1 |  | pod condition Initialized |
+| trainer.pod_PodScheduled | 03:04:17.000 | 1.1 |  | pod condition PodScheduled |
+| trainer.event_Scheduled | 03:04:17.000 | 1.1 |  | event: Successfully assigned openrl-system/orw-lora-qwen-qwen2-5-0-5b-0-trainer to gke-open-rl-dra-l4-trainer-2x-ed601d8b-lvw8 |
+| trainer.workload_reports_node | 03:04:18.224 | 2.3 |  | workloads.tsv poller: status.nodeName set |
+| trainer.event_Pulling | 03:04:21.000 | 5.1 |  | event: Pulling image "us-central1-docker.pkg.dev/cdrollouts-sunilarora/open-rl/open-rl-server:agent3-9f8dc13" |
+| trainer.event_Pulled | 03:07:09.000 | 173.1 | 168.59 | event: Successfully pulled image "us-central1-docker.pkg.dev/cdrollouts-sunilarora/open-rl/open-rl-server:agent3-9f8dc13" in 2m48.59s (2m48.59s inc |
+| trainer.event_Created | 03:07:09.000 | 173.1 |  | event: Container created |
+| trainer.container_started | 03:07:10.000 | 174.1 |  | containerStatuses.state.running.startedAt |
+| trainer.event_Started | 03:07:10.000 | 174.1 |  | event: Container started |
+| trainer.condition_Placed_True | 03:07:11.000 | 175.1 |  | Workload condition Placed |
+| trainer.pod_Ready | 03:07:11.000 | 175.1 |  | pod condition Ready |
+| trainer.pod_ContainersReady | 03:07:11.000 | 175.1 |  | pod condition ContainersReady |
+| trainer.process_first_log | 03:07:17.253 | 181.4 |  | container log, first line |
+| trainer.load_start | 03:07:17.410 | 181.5 |  | container log: Loading base model Qwen/Qwen2.5-0.5B to cuda:0... |
+| trainer.ready_key | 03:07:21.000 | 185.1 |  | redis poller: open_rl:trainer_ready:Qwen/Qwen2.5-0.5B |
+| trainer.load_done | 03:07:21.686 | 185.8 |  | container log: Successfully loaded. |
+| trainer.processor_started | 03:07:21.689 | 185.8 |  | container log: [WORKER] Training requests processor started (model=Qwen/Qwen2.5-0.5B active_set=Qwen/Qwen2.5-0.5B-1). |
+
+## sampler: `orw-lora-qwen-qwen2-5-0-5b-0-sampler` on `gke-open-rl-dra-l4-sampler-2x-ab5fdc9d-2208`
+
+| Mark | Clock (UTC) | t+ s | Self-reported s | Source |
+|---|---|---|---|---|
+| sampler.workload_first_seen_by_poller | 03:07:28.976 | 193.1 |  | workloads.tsv poller (laptop clock, 1 s) |
+| sampler.workload_created | 03:07:29.000 | 193.1 |  | Workload creationTimestamp (apiserver) |
+| sampler.pod_created | 03:07:29.000 | 193.1 |  | pod creationTimestamp (apiserver) |
+| sampler.pod_Initialized | 03:07:30.000 | 194.1 |  | pod condition Initialized |
+| sampler.pod_PodScheduled | 03:07:30.000 | 194.1 |  | pod condition PodScheduled |
+| sampler.event_Scheduled | 03:07:30.000 | 194.1 |  | event: Successfully assigned openrl-system/orw-lora-qwen-qwen2-5-0-5b-0-sampler to gke-open-rl-dra-l4-sampler-2x-ab5fdc9d-2208 |
+| sampler.workload_reports_pod | 03:07:30.816 | 194.9 |  | workloads.tsv poller: status.podName set |
+| sampler.workload_reports_node | 03:07:30.816 | 194.9 |  | workloads.tsv poller: status.nodeName set |
+| sampler.event_Pulling | 03:07:33.000 | 197.1 |  | event: Pulling image "us-central1-docker.pkg.dev/cdrollouts-sunilarora/open-rl/open-rl-server:agent3-9f8dc13" |
+| sampler.container_started | 03:10:22.000 | 366.1 |  | containerStatuses.state.running.startedAt |
+| sampler.event_Pulled | 03:10:22.000 | 366.1 | 168.566 | event: Successfully pulled image "us-central1-docker.pkg.dev/cdrollouts-sunilarora/open-rl/open-rl-server:agent3-9f8dc13" in 2m48.566s (2m48.566s i |
+| sampler.event_Created | 03:10:22.000 | 366.1 |  | event: Container created |
+| sampler.event_Started | 03:10:22.000 | 366.1 |  | event: Container started |
+| sampler.condition_Placed_True | 03:10:23.000 | 367.1 |  | Workload condition Placed |
+| sampler.pod_Ready | 03:10:23.000 | 367.1 |  | pod condition Ready |
+| sampler.pod_ContainersReady | 03:10:23.000 | 367.1 |  | pod condition ContainersReady |
+| sampler.process_first_log | 03:10:33.926 | 378.0 |  | container log, first line |
+| sampler.engine_init_start | 03:11:01.668 | 405.8 |  | container log: (EngineCore pid=161) INFO 10-15 03:11:01 [core.py:123] Initializing a V1 LLM engine (v0.30.0) with config: model='Qwen/Q |
+| sampler.load_start | 03:11:03.371 | 407.5 |  | container log: (EngineCore pid=161) INFO 10-15 03:11:03 [model_runner.py:396] Loading model from scratch... |
+| sampler.prefetch_start | 03:11:04.731 | 408.8 |  | container log: (EngineCore pid=161) INFO 10-15 03:11:04 [weight_utils.py:857] Prefetching checkpoint files into page cache started (in  |
+| sampler.weights_loaded | 03:11:06.644 | 410.8 | 1.92 | container log: (EngineCore pid=161) INFO 10-15 03:11:06 [default_loader.py:430] Loading weights took 1.92 seconds |
+| sampler.model_loaded | 03:11:07.445 | 411.6 | 4.387871 | container log: (EngineCore pid=161) INFO 10-15 03:11:07 [model_runner.py:428] Model loading took 1.47 GiB memory and 4.387871 seconds |
+| sampler.compile_done | 03:11:38.154 | 442.3 | 28.48 | container log: (EngineCore pid=161) INFO 10-15 03:11:38 [monitor.py:53] torch.compile took 28.48 s in total |
+| sampler.graphs_captured | 03:11:53.539 | 457.6 | 11.0 | container log: (EngineCore pid=161) INFO 10-15 03:11:53 [model_runner.py:1066] Graph capturing finished in 11 secs, took 0.23 GiB |
+| sampler.kv_cache_sized | 03:11:54.078 | 458.2 |  | container log: (EngineCore pid=161) INFO 10-15 03:11:54 [gpu_worker.py:640] Available KV cache memory: 17.33 GiB |
+| sampler.engine_init_done | 03:12:04.781 | 468.9 | 57.33 | container log: (EngineCore pid=161) INFO 10-15 03:12:04 [core.py:372] init engine (profile, create kv cache, warmup model) took 57.33 s |
+| sampler.ready_key | 03:12:06.000 | 470.1 |  | redis poller: open_rl:sampler_ready:Qwen/Qwen2.5-0.5B |
+
+## client
+
+| Phase | Clock (UTC) | t+ s |
+|---|---|---|
+| client.start | 03:04:15.891 | -0.0 |
+| client.create_model.start | 03:04:15.892 | 0.0 |
+| client.create_model.done | 03:07:22.617 | 186.7 |
+| client.forward_backward.start | 03:07:25.417 | 189.5 |
+| client.forward_backward.done | 03:07:28.837 | 192.9 |
+| client.optim_step.done | 03:07:29.600 | 193.7 |
+| client.save_weights_and_get_sampling_client.start | 03:07:29.600 | 193.7 |
+| client.save_weights_and_get_sampling_client.done | 03:07:30.173 | 194.3 |
+| client.sample.start | 03:07:30.174 | 194.3 |
+| client.sample.done | 03:12:07.151 | 471.3 |
+| client.end | 03:12:07.154 | 471.3 |
+
+## Critical path (derived from the marks above)
+
+### trainer
+
+| Phase | From | To | Seconds |
+|---|---|---|---|
+| placement: Workload to pod | client.create_model.start | trainer.pod_created | 0.1 |
+| pod scheduled | trainer.pod_created | trainer.pod_PodScheduled | 1.0 |
+| image pull (0 if cached) | trainer.pod_PodScheduled | trainer.event_Pulled | 172.0 |
+| container create | trainer.event_Pulled | trainer.container_started | 1.0 |
+| process start to load start (imports) | trainer.container_started | trainer.load_start | 7.4 |
+| weights load | trainer.load_start | trainer.load_done | 4.3 |
+| ready key | trainer.load_done | trainer.ready_key | -0.7 |
+| client: create_model blocked | client.create_model.start | client.create_model.done | 186.7 |
+| client: first forward_backward | client.forward_backward.start | client.forward_backward.done | 3.4 |
+
+### sampler
+
+| Phase | From | To | Seconds |
+|---|---|---|---|
+| placement: Workload to pod | client.save_weights_and_get_sampling_client.start | sampler.pod_created | -0.6 |
+| pod scheduled | sampler.pod_created | sampler.pod_PodScheduled | 1.0 |
+| image pull (0 if cached) | sampler.pod_PodScheduled | sampler.event_Pulled | 172.0 |
+| container create | sampler.event_Pulled | sampler.container_started | 0.0 |
+| process start to first log | sampler.container_started | sampler.process_first_log | 11.9 |
+| imports until engine init | sampler.process_first_log | sampler.engine_init_start | 27.7 |
+| engine init until load start | sampler.engine_init_start | sampler.load_start | 1.7 |
+| weights load (vLLM prefetch into page cache + load) | sampler.load_start | sampler.weights_loaded | 3.3 |
+| torch.compile | sampler.weights_loaded | sampler.compile_done | 31.5 |
+| profiling, KV cache, graph capture | sampler.compile_done | sampler.engine_init_done | 26.6 |
+| ready key | sampler.engine_init_done | sampler.ready_key | 1.2 |
+| client: first sample blocked | client.sample.start | client.sample.done | 277.0 |
+
diff --git a/runs/2026-10-15_cold_start/B-0.5b-l4-cached/monitoring.md b/runs/2026-10-15_cold_start/B-0.5b-l4-cached/monitoring.md
@@ -0,0 +1,7 @@
+| minute ending (UTC) | read MiB/s | read ops/s | write MiB/s |
+|---|---|---|---|
+| 2026-10-15T02:59:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:00:00Z | 0.0 | 0 | 1.3 |
+| 2026-10-15T03:01:00Z | 0.0 | 0 | 0.4 |
+| 2026-10-15T03:02:00Z | 0.0 | 0 | 0.0 |
+| 2026-10-15T03:03:00Z | 0.0 | 0 | 0.0 |