YARN-9509: Added a configuration for admins to be able to capped per-container cpu usage based on a multiplier

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-api/src/main/java/org/apache/hadoop/yarn/conf/YarnConfiguration.java

@@ -2190,6 +2190,15 @@ public static boolean isAclEnabled(Configuration conf) {
   public static final boolean DEFAULT_NM_LINUX_CONTAINER_CGROUPS_STRICT_RESOURCE_USAGE =
       false;
 
+  /**
+   * In case of strict resource usage provide capability to capped cpu resource
+   * usage by <code>DEFAULT_NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER</code>
+   * times the requested one
+   */
+  public static final String NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER =
+      NM_PREFIX + "linux-container-executor.cgroups.capped-multiplier";
+  public static final float DEFAULT_NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER =
+      1.0f;
 
   // Configurations for applicaiton life time monitor feature
   public static final String RM_APPLICATION_MONITOR_INTERVAL_MS =

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-common/src/main/resources/yarn-default.xml

@@ -1730,6 +1730,16 @@
     <value>false</value>
   </property>
 
+  <property>
+    <description>If yarn.nodemanager.linux-container-executor.cgroups.strict-resource-usage
+    is set to true this flag determines how many times their share of CPU they can use. For
+    example setting it to 1.0f will restrict apps to use only their share of CPU but setting it to 2.0f
+    will authorize apps to use up to 2 times their share of CPU.
+    </description>
+    <name>yarn.nodemanager.linux-container-executor.cgroups.capped-multiplier</name>
+    <value>1.0f</value>
+  </property>
+
   <property>
     <description>Comma separated list of runtimes that are allowed when using
     LinuxContainerExecutor. The allowed values are default, docker, and

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-nodemanager/src/main/java/org/apache/hadoop/yarn/server/nodemanager/util/CgroupsLCEResourcesHandler.java

@@ -94,6 +94,7 @@ public class CgroupsLCEResourcesHandler implements LCEResourcesHandler {
 
   private float yarnProcessors;
   private int nodeVCores;
+  private float cappedMultiplier;
 
   public CgroupsLCEResourcesHandler() {
     this.controllerPaths = new HashMap<String, String>();
@@ -138,13 +139,16 @@ void initConfig() throws IOException {
                 .NM_LINUX_CONTAINER_CGROUPS_STRICT_RESOURCE_USAGE,
             YarnConfiguration
                 .DEFAULT_NM_LINUX_CONTAINER_CGROUPS_STRICT_RESOURCE_USAGE);
+    this.cappedMultiplier = conf.getFloat(
+        YarnConfiguration.NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER,
+        YarnConfiguration.DEFAULT_NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER);
 
     int len = cgroupPrefix.length();
     if (cgroupPrefix.charAt(len - 1) == '/') {
       cgroupPrefix = cgroupPrefix.substring(0, len - 1);
     }
   }
-  
+
   public void init(LinuxContainerExecutor lce) throws IOException {
     this.init(lce,
         ResourceCalculatorPlugin.getResourceCalculatorPlugin(null, conf));
@@ -334,8 +338,8 @@ private void setupLimits(ContainerId containerId,
           String.valueOf(cpuShares));
       if (strictResourceUsageMode) {
         if (nodeVCores != containerVCores) {
-          float containerCPU =
-              (containerVCores * yarnProcessors) / (float) nodeVCores;
+          float containerVCoresCapped = Math.min(containerVCores * cappedMultiplier, nodeVCores);
+          float containerCPU = (containerVCoresCapped * yarnProcessors) / (float) nodeVCores;
           int[] limits = getOverallLimits(containerCPU);
           updateCgroup(CONTROLLER_CPU, containerName, CPU_PERIOD_US,
               String.valueOf(limits[0]));

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-nodemanager/src/test/java/org/apache/hadoop/yarn/server/nodemanager/util/TestCgroupsLCEResourcesHandler.java

@@ -341,8 +341,29 @@ public void testContainerLimits() throws IOException {
     Assert.assertEquals(500 * 1000, readIntFromFile(periodFile));
     Assert.assertEquals(1000 * 1000, readIntFromFile(quotaFile));
 
+    // 1/8 of CPU with multiplier set at 2
+    FileUtils.deleteQuietly(containerCpuDir);
+    conf.setBoolean(
+            YarnConfiguration.NM_LINUX_CONTAINER_CGROUPS_STRICT_RESOURCE_USAGE,
+        true);
+    conf.setFloat(
+            YarnConfiguration.NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER,
+            2.0f);
+    handler.initConfig();
+    handler.preExecute(id,
+            Resource.newInstance(1024, YarnConfiguration.DEFAULT_NM_VCORES / 8));
+    Assert.assertTrue(containerCpuDir.exists());
+    Assert.assertTrue(containerCpuDir.isDirectory());
+    periodFile = new File(containerCpuDir, "cpu.cfs_period_us");
+    quotaFile = new File(containerCpuDir, "cpu.cfs_quota_us");
+    Assert.assertTrue(periodFile.exists());
+    Assert.assertTrue(quotaFile.exists());
+    Assert.assertEquals(1000 * 1000, readIntFromFile(periodFile));
+    Assert.assertEquals(1000 * 1000, readIntFromFile(quotaFile));
+
     // CGroups set to 50% of CPU, container set to 50% of YARN CPU
     FileUtils.deleteQuietly(containerCpuDir);
+    conf.unset(YarnConfiguration.NM_LINUX_CONTAINER_CGROUPS_CAPPED_MULTIPLIER);
     conf.setBoolean(
         YarnConfiguration.NM_LINUX_CONTAINER_CGROUPS_STRICT_RESOURCE_USAGE,
         true);