From 5982bba7e37e4167e74649cd3237756993355f7e Mon Sep 17 00:00:00 2001
From: Yanchao Lu <yanchaol@nvidia.com>
Date: Sun, 7 Sep 2025 22:26:52 +0800
Subject: [PATCH] [None][ci] Block some nodes to avoid unstable network access

Signed-off-by: Yanchao Lu <yanchaol@nvidia.com>
---
 jenkins/BuildDockerImage.groovy | 20 ++++++++++++++++++++
 jenkins/L0_Test.groovy          | 25 ++++++++++++++++++-------
 2 files changed, 38 insertions(+), 7 deletions(-)

diff --git a/jenkins/BuildDockerImage.groovy b/jenkins/BuildDockerImage.groovy
index c0e718882bd..180fcfb5f5a 100644
--- a/jenkins/BuildDockerImage.groovy
+++ b/jenkins/BuildDockerImage.groovy
@@ -94,6 +94,26 @@ def createKubernetesPodConfig(type, arch = "amd64", build_wheel = false)
         """
     }
 
+    if (arch == "amd64") {
+        // For x86_64, we block some nodes to avoid unstable network access.
+        selectors += """
+                affinity:
+                    nodeAffinity:
+                        requiredDuringSchedulingIgnoredDuringExecution:
+                            nodeSelectorTerms:
+                                - matchExpressions:
+                                    - key: "kubernetes.io/hostname"
+                                      operator: NotIn
+                                      values:
+                                        - "sc-ipp-blossom-prod-k8w-105"
+                                        - "sc-ipp-blossom-prod-k8w-114"
+                                        - "sc-ipp-blossom-prod-k8w-115"
+                                        - "sc-ipp-blossom-prod-k8w-121"
+                                        - "sc-ipp-blossom-prod-k8w-123"
+                                        - "sc-ipp-blossom-prod-k8w-124"
+        """
+    }
+
     def archSuffix = arch == "arm64" ? "arm" : "amd"
     def jnlpImage = "urm.nvidia.com/sw-ipp-blossom-sre-docker-local/lambda/custom_jnlp_images_${archSuffix}_linux:jdk17"
 
diff --git a/jenkins/L0_Test.groovy b/jenkins/L0_Test.groovy
index 65c06a888d9..572bd670ae3 100644
--- a/jenkins/L0_Test.groovy
+++ b/jenkins/L0_Test.groovy
@@ -349,13 +349,24 @@ def runLLMTestlistOnSlurm(pipeline, platform, testList, config=VANILLA_CONFIG, p
         }
 
         stage('Checking if the Node is Online') {
-            def counter = 0
-            // We submit the Slurm job with 5 hours timeout, and the K8S pod will be evicted after 22 hours.
-            // Let's use 15 hours to check if the node is online, and with 2 hours buffer.
-            while (!CloudManager.isNodeOnline(nodeName) && counter < 90) {
-                // Wait 10 minutes to check status of the node again
-                sleep(time: 10, unit: 'MINUTES')
-                counter++
+            withCredentials([usernamePassword(credentialsId: 'svc_tensorrt', usernameVariable: 'USERNAME', passwordVariable: 'PASSWORD')]) {
+                def remote = [
+                        ip           : cluster.ip,
+                        host         : cluster.host,
+                        user         : "${pipeline.USERNAME}",
+                        passwd       : "${pipeline.PASSWORD}",
+                        allowAnyHosts: true,
+                ]
+                def counter = 0
+                // We submit the Slurm job with 5 hours timeout, and the K8S pod will be evicted after 22 hours.
+                // Let's use 15 hours to check if the node is online, and with 2 hours buffer.
+                while (!CloudManager.isNodeOnline(nodeName) && counter < 90) {
+                    // Wait 10 minutes to check status of the node again
+                    sleep(time: 10, unit: 'MINUTES')
+                    // Avoid the node being stuck in the held state.
+                    Utils.exec(pipeline, Utils.sshUserCmd(remote, "\"scontrol release ${slurmJobID} || true\""))
+                    counter++
+                }
             }
 
             if (CloudManager.isNodeOnline(nodeName)) {