From 403d368e8ef368b7ab4c843cbc88564966fdb3ab Mon Sep 17 00:00:00 2001
From: George Hong <georgehong@meta.com>
Date: Mon, 8 Jun 2026 17:54:45 -0700
Subject: [PATCH 1/3] Update

[ghstack-poisoned]
---
 osdc/modules/nfd/deploy.sh                   | 48 +++++++++++++++
 osdc/modules/nfd/helm/values.yaml            | 57 ++++++++++++++++++
 osdc/modules/numa-scheduler/deploy.sh        | 61 ++++++++++++++++++++
 osdc/modules/numa-scheduler/helm/values.yaml | 57 ++++++++++++++++++
 4 files changed, 223 insertions(+)
 create mode 100755 osdc/modules/nfd/deploy.sh
 create mode 100644 osdc/modules/nfd/helm/values.yaml
 create mode 100755 osdc/modules/numa-scheduler/deploy.sh
 create mode 100644 osdc/modules/numa-scheduler/helm/values.yaml

diff --git a/osdc/modules/nfd/deploy.sh b/osdc/modules/nfd/deploy.sh
new file mode 100755
index 00000000..7e13d5f0
--- /dev/null
+++ b/osdc/modules/nfd/deploy.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+set -euo pipefail
+#
+# Deploy NFD topology-updater for NUMA-aware scheduling.
+# Called by: just deploy-module <cluster> nfd
+#
+# Args: $1=cluster-id  $2=cluster-name  $3=region
+#
+# Installs the Node Feature Discovery Helm chart with only the
+# topology-updater enabled. This publishes NodeResourceTopology CRDs
+# that the numa-scheduler reads to make NUMA-aware placement decisions.
+
+CLUSTER="$1"
+_CNAME="$2"  # unused but required by deploy-module interface
+_REGION="$3" # unused but required by deploy-module interface
+MODULE_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="${OSDC_ROOT:-$(cd "$MODULE_DIR/../.." && pwd)}"
+UPSTREAM_ROOT="${OSDC_UPSTREAM:-$REPO_ROOT}"
+
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/mise-activate.sh"
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/helm-upgrade.sh"
+
+CFG="$UPSTREAM_ROOT/scripts/cluster-config.py"
+
+# --- Check if enabled ---
+ENABLED=$(uv run "$CFG" "$CLUSTER" nfd.enabled "false")
+if [[ "$ENABLED" != "true" ]]; then
+  echo "NFD disabled for cluster $CLUSTER, skipping."
+  exit 0
+fi
+
+NFD_VERSION=$(uv run "$CFG" "$CLUSTER" nfd.version "0.17.1")
+NFD_UPDATE_INTERVAL=$(uv run "$CFG" "$CLUSTER" nfd.update_interval "15s")
+
+echo "Installing NFD topology-updater v${NFD_VERSION}..."
+helm_upgrade_if_changed nfd nfd \
+  --create-namespace \
+  --history-max 3 \
+  -f "$MODULE_DIR/helm/values.yaml" \
+  --set topologyUpdater.updateInterval="${NFD_UPDATE_INTERVAL}" \
+  --timeout 5m \
+  --wait \
+  oci://ghcr.io/kubernetes-sigs/charts/node-feature-discovery \
+  --version "${NFD_VERSION}"
+
+echo "NFD topology-updater deployed."
diff --git a/osdc/modules/nfd/helm/values.yaml b/osdc/modules/nfd/helm/values.yaml
new file mode 100644
index 00000000..45fe4d72
--- /dev/null
+++ b/osdc/modules/nfd/helm/values.yaml
@@ -0,0 +1,57 @@
+# Node Feature Discovery — topology-updater only
+# Chart: https://github.com/kubernetes-sigs/node-feature-discovery/tree/master/deployment/helm/node-feature-discovery
+#
+# We only need the topology-updater component, which publishes
+# NodeResourceTopology CRDs per node. The NFD worker (feature detection)
+# and master (label reconciliation) are disabled — we don't need node
+# feature labels, just per-NUMA-zone resource visibility for the
+# numa-scheduler.
+
+master:
+  enable: false
+
+worker:
+  enable: false
+
+topologyUpdater:
+  enable: true
+
+  # Poll kubelet PodResources API every 15s (default 60s). Needs to be
+  # faster than the ARC capacity provisioner interval (30s) so NRT
+  # objects reflect current GPU allocations before placeholder pods are
+  # evaluated by the NUMA-aware scheduler.
+  updateInterval: 15s
+
+  # Only run on GPU nodes — non-GPU nodes don't have NUMA topology
+  # concerns and don't need NRT objects.
+  nodeSelector:
+    nvidia.com/gpu: "true"
+
+  # Tolerate all GPU node taints so the topology-updater schedules
+  # at provisioning time, same as other GPU-node DaemonSets.
+  tolerations:
+    - key: nvidia.com/gpu
+      operator: Exists
+      effect: NoSchedule
+    - key: node-fleet
+      operator: Exists
+      effect: NoSchedule
+    - key: instance-type
+      operator: Exists
+      effect: NoSchedule
+    - key: git-cache-not-ready
+      operator: Exists
+      effect: NoSchedule
+
+  resources:
+    requests:
+      cpu: 50m
+      memory: 64Mi
+    limits:
+      cpu: 200m
+      memory: 128Mi
+
+# The topology-updater needs the NRT CRD. The NFD chart installs it
+# when topologyUpdater is enabled.
+topologyGC:
+  enable: true
diff --git a/osdc/modules/numa-scheduler/deploy.sh b/osdc/modules/numa-scheduler/deploy.sh
new file mode 100755
index 00000000..64fedb7f
--- /dev/null
+++ b/osdc/modules/numa-scheduler/deploy.sh
@@ -0,0 +1,61 @@
+#!/usr/bin/env bash
+set -euo pipefail
+#
+# Deploy NUMA-aware secondary scheduler (scheduler-plugins).
+# Called by: just deploy-module <cluster> numa-scheduler
+#
+# Args: $1=cluster-id  $2=cluster-name  $3=region
+#
+# Downloads the scheduler-plugins Helm chart from the GitHub release
+# and installs it as a secondary scheduler named "numa-scheduler" with
+# NodeResourceTopologyMatch enabled. Pods that set
+# schedulerName: numa-scheduler get NUMA-aware placement.
+
+CLUSTER="$1"
+_CNAME="$2"  # unused but required by deploy-module interface
+_REGION="$3" # unused but required by deploy-module interface
+MODULE_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="${OSDC_ROOT:-$(cd "$MODULE_DIR/../.." && pwd)}"
+UPSTREAM_ROOT="${OSDC_UPSTREAM:-$REPO_ROOT}"
+
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/mise-activate.sh"
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/helm-upgrade.sh"
+
+CFG="$UPSTREAM_ROOT/scripts/cluster-config.py"
+
+# --- Check if enabled ---
+ENABLED=$(uv run "$CFG" "$CLUSTER" numa_scheduler.enabled "false")
+if [[ "$ENABLED" != "true" ]]; then
+  echo "numa-scheduler disabled for cluster $CLUSTER, skipping."
+  exit 0
+fi
+
+CHART_VERSION=$(uv run "$CFG" "$CLUSTER" numa_scheduler.chart_version "0.34.7")
+SCHEDULER_REPLICAS=$(uv run "$CFG" "$CLUSTER" numa_scheduler.replicas "2")
+
+# --- Download chart from GitHub release ---
+CHART_TGZ="scheduler-plugins-${CHART_VERSION}.tgz"
+CHART_DIR=$(mktemp -d)
+trap 'rm -rf "$CHART_DIR"' EXIT
+
+if [[ ! -f "${CHART_DIR}/${CHART_TGZ}" ]]; then
+  echo "Downloading scheduler-plugins chart v${CHART_VERSION}..."
+  gh release download "v${CHART_VERSION}" \
+    --repo kubernetes-sigs/scheduler-plugins \
+    --pattern "${CHART_TGZ}" \
+    --dir "$CHART_DIR"
+fi
+
+echo "Installing numa-scheduler (scheduler-plugins v${CHART_VERSION})..."
+helm_upgrade_if_changed numa-scheduler numa-scheduler \
+  --create-namespace \
+  --history-max 3 \
+  -f "$MODULE_DIR/helm/values.yaml" \
+  --set scheduler.replicaCount="${SCHEDULER_REPLICAS}" \
+  --timeout 5m \
+  --wait \
+  "${CHART_DIR}/${CHART_TGZ}"
+
+echo "numa-scheduler deployed."
diff --git a/osdc/modules/numa-scheduler/helm/values.yaml b/osdc/modules/numa-scheduler/helm/values.yaml
new file mode 100644
index 00000000..c5fb3858
--- /dev/null
+++ b/osdc/modules/numa-scheduler/helm/values.yaml
@@ -0,0 +1,57 @@
+# Scheduler-plugins — NUMA-aware secondary scheduler
+# Chart source: https://github.com/kubernetes-sigs/scheduler-plugins
+#   (vendored from manifests/install/charts/as-a-second-scheduler/)
+#
+# Deploys a second kube-scheduler named "numa-scheduler" with the
+# NodeResourceTopologyMatch plugin enabled. Pods that set
+# schedulerName: numa-scheduler will be filtered/scored by per-NUMA-zone
+# resource availability (read from NodeResourceTopology CRDs published
+# by the NFD topology-updater).
+#
+# Nothing uses this scheduler until runner definitions set
+# scheduler_name: numa-scheduler — deploying it is a no-op.
+
+scheduler:
+  name: numa-scheduler
+  image: registry.k8s.io/scheduler-plugins/kube-scheduler:v0.34.7
+  replicaCount: 2
+  leaderElect: true
+
+  # Run on base-infrastructure nodes alongside other control plane components.
+  nodeSelector:
+    role: base-infrastructure
+
+  tolerations:
+    - key: CriticalAddonsOnly
+      operator: Equal
+      value: "true"
+      effect: NoSchedule
+
+  resources:
+    requests:
+      cpu: 200m
+      memory: 256Mi
+    limits:
+      cpu: 500m
+      memory: 512Mi
+
+# The controller is needed for Coscheduling/CapacityScheduling plugins.
+# We only use NodeResourceTopologyMatch, so disable it.
+controller:
+  replicaCount: 0
+
+# Only enable NodeResourceTopologyMatch — we don't need Coscheduling,
+# CapacityScheduling, or other plugins.
+plugins:
+  enabled:
+    - NodeResourceTopologyMatch
+  disabled: []
+
+pluginConfig:
+  - name: NodeResourceTopologyMatch
+    args:
+      scoringStrategy:
+        # Pack pods onto already-busy NUMA zones so other zones stay
+        # fully free for large (4-GPU) requests. Without this, small
+        # pods spread across sockets and fragment both.
+        type: MostAllocated

From 8d3917e4c6f844b5d3158d6884dcb28982c33798 Mon Sep 17 00:00:00 2001
From: George Hong <georgehong@meta.com>
Date: Mon, 8 Jun 2026 18:18:25 -0700
Subject: [PATCH 2/3] Update

[ghstack-poisoned]
---
 osdc/modules/numa-scheduler/deploy.sh | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/osdc/modules/numa-scheduler/deploy.sh b/osdc/modules/numa-scheduler/deploy.sh
index 64fedb7f..3e7bd5de 100755
--- a/osdc/modules/numa-scheduler/deploy.sh
+++ b/osdc/modules/numa-scheduler/deploy.sh
@@ -40,13 +40,11 @@ CHART_TGZ="scheduler-plugins-${CHART_VERSION}.tgz"
 CHART_DIR=$(mktemp -d)
 trap 'rm -rf "$CHART_DIR"' EXIT
 
-if [[ ! -f "${CHART_DIR}/${CHART_TGZ}" ]]; then
-  echo "Downloading scheduler-plugins chart v${CHART_VERSION}..."
-  gh release download "v${CHART_VERSION}" \
-    --repo kubernetes-sigs/scheduler-plugins \
-    --pattern "${CHART_TGZ}" \
-    --dir "$CHART_DIR"
-fi
+echo "Downloading scheduler-plugins chart v${CHART_VERSION}..."
+gh release download "v${CHART_VERSION}" \
+  --repo kubernetes-sigs/scheduler-plugins \
+  --pattern "${CHART_TGZ}" \
+  --dir "$CHART_DIR"
 
 echo "Installing numa-scheduler (scheduler-plugins v${CHART_VERSION})..."
 helm_upgrade_if_changed numa-scheduler numa-scheduler \

From 3719c31717393dd5c3d25f2411b1fce9d1136eee Mon Sep 17 00:00:00 2001
From: George Hong <georgehong@meta.com>
Date: Wed, 10 Jun 2026 15:00:59 -0700
Subject: [PATCH 3/3] Update

[ghstack-poisoned]
---
 osdc/modules/nfd/deploy.sh            | 13 ++-----------
 osdc/modules/numa-scheduler/deploy.sh | 13 ++-----------
 2 files changed, 4 insertions(+), 22 deletions(-)

diff --git a/osdc/modules/nfd/deploy.sh b/osdc/modules/nfd/deploy.sh
index 8a3fa95b..841c8c49 100755
--- a/osdc/modules/nfd/deploy.sh
+++ b/osdc/modules/nfd/deploy.sh
@@ -11,26 +11,17 @@ set -euo pipefail
 # that the numa-scheduler reads to make NUMA-aware placement decisions.
 
 CLUSTER="$1"
-_CNAME="$2"  # unused but required by deploy-module interface
-_REGION="$3" # unused but required by deploy-module interface
+export CNAME="$2"
+export REGION="$3"
 MODULE_DIR="$(cd "$(dirname "$0")" && pwd)"
 REPO_ROOT="${OSDC_ROOT:-$(cd "$MODULE_DIR/../.." && pwd)}"
 UPSTREAM_ROOT="${OSDC_UPSTREAM:-$REPO_ROOT}"
-
 # shellcheck source=/dev/null
 source "$UPSTREAM_ROOT/scripts/mise-activate.sh"
 # shellcheck source=/dev/null
 source "$UPSTREAM_ROOT/scripts/helm-upgrade.sh"
-
 CFG="$UPSTREAM_ROOT/scripts/cluster-config.py"
 
-# --- Check if enabled ---
-ENABLED=$(uv run "$CFG" "$CLUSTER" nfd.enabled "false")
-if [[ "$ENABLED" != "true" ]]; then
-  echo "NFD disabled for cluster $CLUSTER, skipping."
-  exit 0
-fi
-
 NFD_VERSION=$(uv run "$CFG" "$CLUSTER" nfd.version "0.17.1")
 NFD_UPDATE_INTERVAL=$(uv run "$CFG" "$CLUSTER" nfd.update_interval "15s")
 
diff --git a/osdc/modules/numa-scheduler/deploy.sh b/osdc/modules/numa-scheduler/deploy.sh
index a39bec01..a5ca18e1 100755
--- a/osdc/modules/numa-scheduler/deploy.sh
+++ b/osdc/modules/numa-scheduler/deploy.sh
@@ -12,26 +12,17 @@ set -euo pipefail
 # schedulerName: numa-scheduler get NUMA-aware placement.
 
 CLUSTER="$1"
-_CNAME="$2"  # unused but required by deploy-module interface
-_REGION="$3" # unused but required by deploy-module interface
+export CNAME="$2"
+export REGION="$3"
 MODULE_DIR="$(cd "$(dirname "$0")" && pwd)"
 REPO_ROOT="${OSDC_ROOT:-$(cd "$MODULE_DIR/../.." && pwd)}"
 UPSTREAM_ROOT="${OSDC_UPSTREAM:-$REPO_ROOT}"
-
 # shellcheck source=/dev/null
 source "$UPSTREAM_ROOT/scripts/mise-activate.sh"
 # shellcheck source=/dev/null
 source "$UPSTREAM_ROOT/scripts/helm-upgrade.sh"
-
 CFG="$UPSTREAM_ROOT/scripts/cluster-config.py"
 
-# --- Check if enabled ---
-ENABLED=$(uv run "$CFG" "$CLUSTER" numa_scheduler.enabled "false")
-if [[ "$ENABLED" != "true" ]]; then
-  echo "numa-scheduler disabled for cluster $CLUSTER, skipping."
-  exit 0
-fi
-
 CHART_VERSION=$(uv run "$CFG" "$CLUSTER" numa_scheduler.chart_version "0.34.7")
 SCHEDULER_REPLICAS=$(uv run "$CFG" "$CLUSTER" numa_scheduler.replicas "2")