Mailing-List: contact yarn-issues-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Date: Tue, 14 Feb 2017 23:41:43 +0000 (UTC)
From: "ASF GitHub Bot (JIRA)" <jira@apache.org>
To: yarn-issues@hadoop.apache.org
Message-ID: <JIRA.13041452.1486580325000.81618.1487115703101@Atlassian.JIRA>
In-Reply-To: <JIRA.13041452.1486580325000@Atlassian.JIRA>
References: <JIRA.13041452.1486580325000@Atlassian.JIRA> <JIRA.13041452.1486580325598@jira-lw-us.apache.org>
Subject: [jira] [Commented] (YARN-6163) FS Preemption is a trickle for
 severely starved applications
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
archived-at: Tue, 14 Feb 2017 23:41:51 -0000


    [ https://issues.apache.org/jira/browse/YARN-6163?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15866938#comment-15866938 ] 

ASF GitHub Bot commented on YARN-6163:
--------------------------------------

Github user templedf commented on a diff in the pull request:

    https://github.com/apache/hadoop/pull/192#discussion_r101171746
  
    --- Diff: hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/VisitedResourceRequestTracker.java ---
    @@ -0,0 +1,124 @@
    +/**
    + * Licensed to the Apache Software Foundation (ASF) under one
    + * or more contributor license agreements.  See the NOTICE file
    + * distributed with this work for additional information
    + * regarding copyright ownership.  The ASF licenses this file
    + * to you under the Apache License, Version 2.0 (the
    + * "License"); you may not use this file except in compliance
    + * with the License.  You may obtain a copy of the License at
    + *
    + *     http://www.apache.org/licenses/LICENSE-2.0
    + *
    + * Unless required by applicable law or agreed to in writing, software
    + * distributed under the License is distributed on an "AS IS" BASIS,
    + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    + * See the License for the specific language governing permissions and
    + * limitations under the License.
    + */
    +package org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair;
    +
    +import org.apache.hadoop.yarn.api.records.Priority;
    +import org.apache.hadoop.yarn.api.records.Resource;
    +import org.apache.hadoop.yarn.api.records.ResourceRequest;
    +import org.apache.hadoop.yarn.server.resourcemanager.scheduler.ClusterNodeTracker;
    +
    +import java.util.HashMap;
    +import java.util.HashSet;
    +import java.util.List;
    +import java.util.Map;
    +import java.util.Set;
    +
    +/**
    + * Applications place {@link ResourceRequest}s at multiple levels. This is a
    + * helper class that allows tracking if a {@link ResourceRequest} has been
    + * visited at a different locality level.
    + *
    + * This is implemented for {@link FSAppAttempt#getStarvedResourceRequests()}.
    + * The implementation is not thread-safe.
    + */
    +class VisitedResourceRequestTracker {
    +  private final Map<Priority, Map<Resource, TrackerPerPriorityResource>> map =
    +      new HashMap<>();
    +  private final ClusterNodeTracker<FSSchedulerNode> nodeTracker;
    +
    +  VisitedResourceRequestTracker(
    +      ClusterNodeTracker<FSSchedulerNode> nodeTracker) {
    +    this.nodeTracker = nodeTracker;
    +  }
    +
    +  /**
    +   * Check if the {@link ResourceRequest} is visited before, and track it.
    +   * @param rr {@link ResourceRequest} to visit
    +   * @return true if <code>rr</code> this is the first visit across all
    +   * locality levels, false otherwise
    +   */
    +  boolean visit(ResourceRequest rr) {
    +    Priority priority = rr.getPriority();
    +    Resource capability = rr.getCapability();
    +
    +    Map<Resource, TrackerPerPriorityResource> subMap = map.get(priority);
    +    if (subMap == null) {
    +      subMap = new HashMap<>();
    +      map.put(priority, subMap);
    +    }
    +
    +    TrackerPerPriorityResource tracker = subMap.get(capability);
    +    if (tracker == null) {
    +      tracker = new TrackerPerPriorityResource();
    +      subMap.put(capability, tracker);
    +    }
    +
    +    return tracker.visit(rr.getResourceName());
    +  }
    +
    +  private class TrackerPerPriorityResource {
    +    private Set<String> racksWithNodesVisited = new HashSet<>();
    +    private Set<String> racksVisted = new HashSet<>();
    +    private boolean anyVisited;
    +
    +    private boolean visitAny() {
    +      if (racksVisted.isEmpty() && racksWithNodesVisited.isEmpty()) {
    +        anyVisited = true;
    +      }
    +      return anyVisited;
    +    }
    +
    +    private boolean visitRack(String rackName) {
    +      if (anyVisited || racksWithNodesVisited.contains(rackName)) {
    +        return false;
    +      } else {
    +        racksVisted.add(rackName);
    +        return true;
    +      }
    +    }
    +
    +    private boolean visitNode(String rackName) {
    +      if (anyVisited || racksVisted.contains(rackName)) {
    +        return false;
    +      } else {
    +        racksWithNodesVisited.add(rackName);
    +        return true;
    +      }
    +    }
    +
    +    private boolean visit(String resourceName) {
    +      if (resourceName.equals(ResourceRequest.ANY)) {
    +        return visitAny();
    +      }
    +
    +      List<FSSchedulerNode> nodes =
    +          nodeTracker.getNodesByResourceName(resourceName);
    +      switch (nodes.size()) {
    +        case 0:
    +          // Log error
    +          return false;
    +        case 1:
    +          // Node
    +          return visitNode(nodes.remove(0).getRackName());
    --- End diff --
    
    Seems like you could use a second check here to make sure that this isn't actually a 1-node rack, e.g. make sure the rack name isn't the resource name.  What's worse, marking an unvisited request as visited or vice versa?


> FS Preemption is a trickle for severely starved applications
> ------------------------------------------------------------
>
>                 Key: YARN-6163
>                 URL: https://issues.apache.org/jira/browse/YARN-6163
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: fairscheduler
>    Affects Versions: 2.9.0
>            Reporter: Karthik Kambatla
>            Assignee: Karthik Kambatla
>         Attachments: yarn-6163-1.patch, yarn-6163-2.patch
>
>
> With current logic, only one RR is considered per each instance of marking an application starved. This marking happens only on the update call that runs every 500ms.  Due to this, an application that is severely starved takes forever to reach fairshare based on preemptions.


--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org