feat: Restore Tuned CPU isolation, add Nova pinning and docs

Recommend to prefer Tuned over isolcpu kernel args

Co-authored-by: aider (gemini/gemini-2.5-pro) <[email protected]>

Author: bogdando

dt/nova/nova04delta/edpm/nodeset/kustomization.yaml

@@ -39,6 +39,18 @@ replacements:
           create: true
 
   # Nova compute PCI passthrough customization
+  - source:
+      kind: ConfigMap
+      name: edpm-nodeset-values
+      fieldPath: data.nova.compute.conf
+    targets:
+      - select:
+          kind: ConfigMap
+          name: cpu-pinning-nova
+        fieldPaths:
+          - data.25-cpu-pinning-nova\.conf
+        options:
+          create: true
   - source:
       kind: ConfigMap
       name: edpm-nodeset-values

dt/nova/nova04delta/edpm/nodeset/nova_gpu.yaml

@@ -1,5 +1,12 @@
 apiVersion: v1
 kind: ConfigMap
+metadata:
+  name: cpu-pinning-nova
+data:
+  25-cpu-pinning-nova.conf: _replaced_
+---
+apiVersion: v1
+kind: ConfigMap
 metadata:
   name: gpu-nova
 data:
@@ -13,6 +20,8 @@ spec:
   label: dataplane-deployment-nova-custom-gpu
   edpmServiceType: nova
   dataSources:
+    - configMapRef:
+        name: cpu-pinning-nova
     - configMapRef:
         name: gpu-nova
     - secretRef:

dt/nova/nova04delta/edpm/nodeset2/kustomization.yaml

@@ -39,6 +39,18 @@ replacements:
           create: true
 
   # Nova compute PCI passthrough customization
+  - source:
+      kind: ConfigMap
+      name: edpm-nodeset-values
+      fieldPath: data.nova.compute.conf
+    targets:
+      - select:
+          kind: ConfigMap
+          name: cpu-pinning-nova-2
+        fieldPaths:
+          - data.25-cpu-pinning-nova\.conf
+        options:
+          create: true
   - source:
       kind: ConfigMap
       name: edpm-nodeset-values

dt/nova/nova04delta/edpm/nodeset2/nova_gpu.yaml

@@ -1,5 +1,12 @@
 apiVersion: v1
 kind: ConfigMap
+metadata:
+  name: cpu-pinning-nova-2
+data:
+  25-cpu-pinning-nova.conf: _replaced_
+---
+apiVersion: v1
+kind: ConfigMap
 metadata:
   name: gpu-nova-2
 data:
@@ -13,6 +20,8 @@ spec:
   label: dataplane-deployment-nova-custom-gpu
   edpmServiceType: nova
   dataSources:
+    - configMapRef:
+        name: cpu-pinning-nova-2
     - configMapRef:
         name: gpu-nova-2
     - secretRef:

examples/dt/nova/nova04delta/README.md

@@ -21,6 +21,10 @@ The following parameters are crucial for host-level configuration:
     *   `vfio-pci.ids=10de:20f1`: Instructs the `vfio-pci` driver to claim the specified GPU(s) by their vendor and product IDs at boot time. The example IDs `10de:20f1` are for an NVIDIA A100 GPU.
     *   `rd.driver.pre=vfio-pci`: Avoids race conditions during boot by loading vfio-pci kernel module early.
 
+*   `edpm_tuned_profile` and `edpm_tuned_isolated_cores`: These parameters configure the `tuned` service.
+    *   `edpm_tuned_profile` is set to `cpu-partitioning-powersave` to enable CPU isolation features.
+    *   `edpm_tuned_isolated_cores` specifies the cores to be isolated. For CPU isolation we strongly recommend using the Tuned approach rather than `isolcpus` kernel argument.
+
 *   **VFIO-PCI Binding Service**: The `vfio-pci-bind` service in `dt/nova/nova04delta/edpm/nodeset/nova_gpu.yaml` blacklists the `nouveau` and `nvidia` kernel modules to ensure they do not interfere with the `vfio-pci` driver. The service also regenerates the initramfs and grub configuration to apply these changes. A reboot is required for these changes to take effect.
 
 ## Nova Configuration
@@ -63,6 +67,8 @@ That is a contrary to the legacy mode where PCI devices used to be requested thr
 In addition to PCI device configuration, the `nova.compute.conf` section includes parameters for resource management on the compute node:
 
 *   `[DEFAULT]reserved_host_memory_mb`: Specifies the amount of memory (in megabytes) to reserve for the host operating system and other non-OpenStack services. This memory will not be available for allocation to virtual machines.
+*   `[compute]cpu_shared_set`: A list of physical CPUs that are available for host processes and for virtual machines that do not have dedicated CPUs (i.e., unpinned VMs). These should be the CPUs that are **not** isolated by `edpm_tuned_isolated_cores`.
+*   `[compute]cpu_dedicated_set`: A list of physical CPUs that are exclusively reserved for virtual machines with dedicated CPU pinning policies. To ensure performance isolation, this list should correspond directly to the CPUs isolated using `edpm_tuned_isolated_cores` parameter.
 *   `[DEFAULT]reserved_huge_pages`: Defines the number and size of huge pages to reserve for the host, making them unavailable for guest VMs. This configuration works in conjunction with the `hugepages` and `hugepagesz` kernel arguments, which define the total pool of huge pages on the host.
 
 **Note**: In a full device passthrough scenario, the `[devices]enabled_vgpu_types` option in Nova's configuration is not used. This option is specific to mediated device (mdev) configurations.

examples/dt/nova/nova04delta/edpm/nodeset/values.yaml

@@ -41,6 +41,8 @@ data:
         # Here, 10de represents NVIDIA Corporation, and 20f1 corresponds to the GA100 [A100 PCIe 40GB] GPU, according to Device Hunt
         # https://devicehunt.com/view/type/pci/vendor/10DE/device/20F1
         edpm_kernel_args: "default_hugepagesz=1GB hugepagesz=1G hugepages=16 intel_iommu=on iommu=pt vfio-pci.ids=10de:20f1 rd.driver.pre=vfio-pci"
+        edpm_tuned_profile: "cpu-partitioning-powersave"
+        edpm_tuned_isolated_cores: "4-23,28-47"
         # edpm_network_config
         # These vars are edpm_network_config role vars
         edpm_network_config_hide_sensitive_logs: false
@@ -138,6 +140,9 @@ data:
         reserved_host_memory_mb = 4096
         reserved_huge_pages = node:0,size:4,count:524160
         reserved_huge_pages = node:1,size:4,count:524160
+        [compute]
+        cpu_shared_set = 0-3,24-27
+        cpu_dedicated_set = 8-23,32-47
     migration:
       ssh_keys:
         private: CHANGEME4

examples/dt/nova/nova04delta/edpm/nodeset2/values.yaml

@@ -41,6 +41,8 @@ data:
         # Here, 10de represents NVIDIA Corporation, and 20f1 corresponds to the GA100 [A100 PCIe 40GB] GPU, according to Device Hunt
         # https://devicehunt.com/view/type/pci/vendor/10DE/device/20F1
         edpm_kernel_args: "default_hugepagesz=1GB hugepagesz=1G hugepages=16 intel_iommu=on iommu=pt vfio-pci.ids=10de:20f1 rd.driver.pre=vfio-pci"
+        edpm_tuned_profile: "cpu-partitioning-powersave"
+        edpm_tuned_isolated_cores: "4-23,28-47"
         # edpm_network_config
         # These vars are edpm_network_config role vars
         edpm_network_config_hide_sensitive_logs: false
@@ -138,6 +140,9 @@ data:
         reserved_host_memory_mb = 4096
         reserved_huge_pages = node:0,size:4,count:524160
         reserved_huge_pages = node:1,size:4,count:524160
+        [compute]
+        cpu_shared_set = 0-3,24-27
+        cpu_dedicated_set = 8-23,32-47
     migration:
       ssh_keys:
         private: CHANGEME4

examples/va/nvidia-vfio-passthrough/edpm/nodeset/values.yaml

@@ -41,6 +41,8 @@ data:
         # Here, 10de represents NVIDIA Corporation, and 20f1 corresponds to the GA100 [A100 PCIe 40GB] GPU, according to Device Hunt
         # https://devicehunt.com/view/type/pci/vendor/10DE/device/20F1
         edpm_kernel_args: "default_hugepagesz=1GB hugepagesz=1G hugepages=16 intel_iommu=on iommu=pt vfio-pci.ids=10de:20f1 rd.driver.pre=vfio-pci"
+        edpm_tuned_profile: "cpu-partitioning-powersave"
+        edpm_tuned_isolated_cores: "4-23,28-47"
         # edpm_network_config
         # These vars are edpm_network_config role vars
         edpm_network_config_hide_sensitive_logs: false
@@ -138,6 +140,9 @@ data:
         reserved_host_memory_mb = 4096
         reserved_huge_pages = node:0,size:4,count:524160
         reserved_huge_pages = node:1,size:4,count:524160
+        [compute]
+        cpu_shared_set = 0-3,24-27
+        cpu_dedicated_set = 8-23,32-47
     migration:
       ssh_keys:
         private: CHANGEME4

va/nvidia-vfio-passthrough/edpm/nodeset/kustomization.yaml

@@ -39,6 +39,18 @@ replacements:
           create: true
 
   # Nova compute PCI passthrough customization
+  - source:
+      kind: ConfigMap
+      name: edpm-nodeset-values
+      fieldPath: data.nova.compute.conf
+    targets:
+      - select:
+          kind: ConfigMap
+          name: cpu-pinning-nova
+        fieldPaths:
+          - data.25-cpu-pinning-nova\.conf
+        options:
+          create: true
   - source:
       kind: ConfigMap
       name: edpm-nodeset-values

va/nvidia-vfio-passthrough/edpm/nodeset/nova_gpu.yaml

@@ -1,5 +1,12 @@
 apiVersion: v1
 kind: ConfigMap
+metadata:
+  name: cpu-pinning-nova
+data:
+  25-cpu-pinning-nova.conf: _replaced_
+---
+apiVersion: v1
+kind: ConfigMap
 metadata:
   name: gpu-nova
 data:
@@ -13,6 +20,8 @@ spec:
   label: dataplane-deployment-nova-custom-gpu
   edpmServiceType: nova
   dataSources:
+    - configMapRef:
+        name: cpu-pinning-nova
     - configMapRef:
         name: gpu-nova
     - secretRef: