You are viewing a plain text version of this content. The canonical link for it is here.
Posted to common-commits@hadoop.apache.org by ta...@apache.org on 2021/09/13 02:54:32 UTC
[hadoop] branch trunk updated: YARN-10903. Fix the headroom check
in ParentQueue and RegularContainerAllocator for DRF (#3352)
This is an automated email from the ASF dual-hosted git repository.
taoyang pushed a commit to branch trunk
in repository https://gitbox.apache.org/repos/asf/hadoop.git
The following commit(s) were added to refs/heads/trunk by this push:
new d8026e3 YARN-10903. Fix the headroom check in ParentQueue and RegularContainerAllocator for DRF (#3352)
d8026e3 is described below
commit d8026e387e2b18a22f223488cf2504d8f7c7131f
Author: Jack <ja...@outlook.com>
AuthorDate: Mon Sep 13 10:54:11 2021 +0800
YARN-10903. Fix the headroom check in ParentQueue and RegularContainerAllocator for DRF (#3352)
Contributed by Jie Wang <ji...@hulu.com>
---
.../scheduler/capacity/ParentQueue.java | 7 ++-
.../allocator/RegularContainerAllocator.java | 20 ++++----
.../scheduler/capacity/TestLeafQueue.java | 56 ++++++++++++++++++++++
3 files changed, 67 insertions(+), 16 deletions(-)
diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/ParentQueue.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/ParentQueue.java
index e1307d7..24c3b90 100644
--- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/ParentQueue.java
+++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/ParentQueue.java
@@ -1038,10 +1038,9 @@ public class ParentQueue extends AbstractCSQueue {
// Two conditions need to meet when trying to allocate:
// 1) Node doesn't have reserved container
// 2) Node's available-resource + killable-resource should > 0
- boolean accept = node.getReservedContainer() == null && Resources
- .greaterThanOrEqual(resourceCalculator, clusterResource, Resources
- .add(node.getUnallocatedResource(),
- node.getTotalKillableResources()), minimumAllocation);
+ boolean accept = node.getReservedContainer() == null &&
+ Resources.fitsIn(resourceCalculator, minimumAllocation,
+ Resources.add(node.getUnallocatedResource(), node.getTotalKillableResources()));
if (!accept) {
ActivitiesLogger.QUEUE.recordQueueActivity(activitiesManager, node,
getParentName(), getQueuePath(), ActivityState.REJECTED,
diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/allocator/RegularContainerAllocator.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/allocator/RegularContainerAllocator.java
index cced238..b396d57 100644
--- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/allocator/RegularContainerAllocator.java
+++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/allocator/RegularContainerAllocator.java
@@ -73,10 +73,9 @@ public class RegularContainerAllocator extends AbstractContainerAllocator {
ActivitiesManager activitiesManager) {
super(application, rc, rmContext, activitiesManager);
}
-
- private boolean checkHeadroom(Resource clusterResource,
- ResourceLimits currentResourceLimits, Resource required,
- String nodePartition) {
+
+ private boolean checkHeadroom(ResourceLimits currentResourceLimits,
+ Resource required, String nodePartition) {
// If headroom + currentReservation < required, we cannot allocate this
// require
Resource resourceCouldBeUnReserved =
@@ -86,9 +85,8 @@ public class RegularContainerAllocator extends AbstractContainerAllocator {
// we won't allow to unreserve before allocation.
resourceCouldBeUnReserved = Resources.none();
}
- return Resources.greaterThanOrEqual(rc, clusterResource, Resources.add(
- currentResourceLimits.getHeadroom(), resourceCouldBeUnReserved),
- required);
+ return Resources.fitsIn(rc, required,
+ Resources.add(currentResourceLimits.getHeadroom(), resourceCouldBeUnReserved));
}
/*
@@ -97,8 +95,7 @@ public class RegularContainerAllocator extends AbstractContainerAllocator {
* We will consider stuffs like exclusivity, pending resource, node partition,
* headroom, etc.
*/
- private ContainerAllocation preCheckForNodeCandidateSet(
- Resource clusterResource, FiCaSchedulerNode node,
+ private ContainerAllocation preCheckForNodeCandidateSet(FiCaSchedulerNode node,
SchedulingMode schedulingMode, ResourceLimits resourceLimits,
SchedulerRequestKey schedulerKey) {
PendingAsk offswitchPendingAsk = application.getPendingAsk(schedulerKey,
@@ -168,8 +165,7 @@ public class RegularContainerAllocator extends AbstractContainerAllocator {
}
}
- if (!checkHeadroom(clusterResource, resourceLimits, required,
- node.getPartition())) {
+ if (!checkHeadroom(resourceLimits, required, node.getPartition())) {
LOG.debug("cannot allocate required resource={} because of headroom",
required);
ActivitiesLogger.APP.recordAppActivityWithoutAllocation(
@@ -857,7 +853,7 @@ public class RegularContainerAllocator extends AbstractContainerAllocator {
FiCaSchedulerNode node = iter.next();
if (reservedContainer == null) {
- result = preCheckForNodeCandidateSet(clusterResource, node,
+ result = preCheckForNodeCandidateSet(node,
schedulingMode, resourceLimits, schedulerKey);
if (null != result) {
continue;
diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestLeafQueue.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestLeafQueue.java
index 854eeb6..aa342af 100644
--- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestLeafQueue.java
+++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestLeafQueue.java
@@ -800,6 +800,62 @@ public class TestLeafQueue {
assertEquals((int)(a.getCapacity() * node_0.getTotalResource().getMemorySize()),
a.getMetrics().getAvailableMB());
}
+
+ @Test
+ public void testHeadroomCheckWithDRF() throws Exception {
+ CSAssignment assignment;
+ setUpWithDominantResourceCalculator();
+ // Mock the queue
+ LeafQueue b = stubLeafQueue((LeafQueue) queues.get(B));
+ // Users
+ final String user0 = "user_0";
+
+ // Submit applications
+ final ApplicationAttemptId appAttemptId0 =
+ TestUtils.getMockApplicationAttemptId(0, 0);
+ FiCaSchedulerApp app0 =
+ new FiCaSchedulerApp(appAttemptId0, user0, b,
+ b.getAbstractUsersManager(), spyRMContext);
+ b.submitApplicationAttempt(app0, user0);
+ // Setup some nodes
+ String host0 = "127.0.0.1";
+ FiCaSchedulerNode node0 =
+ TestUtils.getMockNode(host0, DEFAULT_RACK, 0, 100 * GB, 100);
+
+ int numNodes = 1;
+ Resource clusterResource =
+ Resources.createResource(numNodes * (100 * GB), numNodes * 100);
+ when(csContext.getNumClusterNodes()).thenReturn(numNodes);
+ root.updateClusterResource(clusterResource,
+ new ResourceLimits(clusterResource));
+
+ // Increase the user-limit-factor to make user_0 fully use max resources of the queue.
+ // The max resources can be used are 0.99 * [100 * GB, 100]
+ b.setUserLimitFactor(10.0f);
+
+ Map<ApplicationAttemptId, FiCaSchedulerApp> apps =
+ ImmutableMap.of(app0.getApplicationAttemptId(), app0);
+ Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node0.getNodeID(), node0);
+
+ Priority priority = TestUtils.createMockPriority(1);
+ app0.updateResourceRequests(Collections.singletonList(TestUtils
+ .createResourceRequest(ResourceRequest.ANY, 90 * GB, 10, 1, true,
+ priority, recordFactory, NO_LABEL)));
+ assignment = b.assignContainers(clusterResource, node0, new ResourceLimits(
+ clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+ applyCSAssignment(clusterResource, assignment, b, nodes, apps);
+ verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
+
+ app0.updateResourceRequests(Collections.singletonList(TestUtils
+ .createResourceRequest(ResourceRequest.ANY, 10 * GB, 10, 1, true,
+ priority, recordFactory, NO_LABEL)));
+ assignment = b.assignContainers(clusterResource, node0, new ResourceLimits(
+ clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+ // This assignment should have no containers assigned,
+ // because the used memory (90 + 10)GB will exceed the max 99GB
+ verifyNoContainerAllocated(assignment);
+ }
+
@Test
public void testDRFUsageRatioRounding() throws Exception {
CSAssignment assign;
---------------------------------------------------------------------
To unsubscribe, e-mail: common-commits-unsubscribe@hadoop.apache.org
For additional commands, e-mail: common-commits-help@hadoop.apache.org