diff options
author | Andrew Or <andrewor14@apache.org> | 2015-01-08 02:33:11 +0000 |
---|---|---|
committer | Andrew Or <andrewor14@apache.org> | 2015-01-08 02:33:11 +0000 |
commit | 11276bf6a6797f113f1cfc0894a53e40f366f0b5 (patch) | |
tree | ce89255069f86248e137846f97390f73a318e9b2 /site/docs/1.2.0 | |
parent | 3f48d5b4622829a11fa1be5230289306c349f914 (diff) | |
download | spark-website-11276bf6a6797f113f1cfc0894a53e40f366f0b5.tar.gz spark-website-11276bf6a6797f113f1cfc0894a53e40f366f0b5.tar.bz2 spark-website-11276bf6a6797f113f1cfc0894a53e40f366f0b5.zip |
Push dynamic allocation docs
Diffstat (limited to 'site/docs/1.2.0')
-rw-r--r-- | site/docs/1.2.0/job-scheduling.html | 152 |
1 files changed, 140 insertions, 12 deletions
diff --git a/site/docs/1.2.0/job-scheduling.html b/site/docs/1.2.0/job-scheduling.html index a08a0e5b1..f4e39a820 100644 --- a/site/docs/1.2.0/job-scheduling.html +++ b/site/docs/1.2.0/job-scheduling.html @@ -6,7 +6,7 @@ <head> <meta charset="utf-8"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"> - <title>Job Scheduling - Spark 1.2.0 Documentation</title> + <title>Job Scheduling - Spark 1.3.0 Documentation</title> <meta name="description" content=""> @@ -53,7 +53,7 @@ <div class="navbar-inner"> <div class="container"> <div class="brand"><a href="index.html"> - <img src="img/spark-logo-hd.png" style="height:50px;"/></a><span class="version">1.2.0</span> + <img src="img/spark-logo-hd.png" style="height:50px;"/></a><span class="version">1.3.0</span> </div> <ul class="nav"> <!--TODO(andyk): Add class="active" attribute to li some how.--> @@ -113,7 +113,7 @@ </ul> </li> </ul> - <!--<p class="navbar-text pull-right"><span class="version-text">v1.2.0</span></p>--> + <!--<p class="navbar-text pull-right"><span class="version-text">v1.3.0</span></p>--> </div> </div> </div> @@ -125,7 +125,19 @@ <ul id="markdown-toc"> <li><a href="#overview">Overview</a></li> - <li><a href="#scheduling-across-applications">Scheduling Across Applications</a></li> + <li><a href="#scheduling-across-applications">Scheduling Across Applications</a> <ul> + <li><a href="#dynamic-resource-allocation">Dynamic Resource Allocation</a> <ul> + <li><a href="#configuration-and-setup">Configuration and Setup</a></li> + <li><a href="#resource-allocation-policy">Resource Allocation Policy</a> <ul> + <li><a href="#request-policy">Request Policy</a></li> + <li><a href="#remove-policy">Remove Policy</a></li> + </ul> + </li> + <li><a href="#graceful-decommission-of-executors">Graceful Decommission of Executors</a></li> + </ul> + </li> + </ul> + </li> <li><a href="#scheduling-within-an-application">Scheduling Within an Application</a> <ul> <li><a href="#fair-scheduler-pools">Fair Scheduler Pools</a></li> <li><a href="#default-behavior-of-pools">Default Behavior of Pools</a></li> @@ -186,6 +198,117 @@ the same RDDs. For example, the <a href="http://shark.cs.berkeley.edu">Shark</a> queries. In future releases, in-memory storage systems such as <a href="http://tachyon-project.org">Tachyon</a> will provide another approach to share RDDs.</p> +<h2 id="dynamic-resource-allocation">Dynamic Resource Allocation</h2> + +<p>Spark 1.2 introduces the ability to dynamically scale the set of cluster resources allocated to +your application up and down based on the workload. This means that your application may give +resources back to the cluster if they are no longer used and request them again later when there +is demand. This feature is particularly useful if multiple applications share resources in your +Spark cluster. If a subset of the resources allocated to an application becomes idle, it can be +returned to the cluster’s pool of resources and acquired by other applications. In Spark, dynamic +resource allocation is performed on the granularity of the executor and can be enabled through +<code>spark.dynamicAllocation.enabled</code>.</p> + +<p>This feature is currently disabled by default and available only on <a href="running-on-yarn.html">YARN</a>. +A future release will extend this to <a href="spark-standalone.html">standalone mode</a> and +<a href="running-on-mesos.html#mesos-run-modes">Mesos coarse-grained mode</a>. Note that although Spark on +Mesos already has a similar notion of dynamic resource sharing in fine-grained mode, enabling +dynamic allocation allows your Mesos application to take advantage of coarse-grained low-latency +scheduling while sharing cluster resources efficiently.</p> + +<h3 id="configuration-and-setup">Configuration and Setup</h3> + +<p>All configurations used by this feature live under the <code>spark.dynamicAllocation.*</code> namespace. +To enable this feature, your application must set <code>spark.dynamicAllocation.enabled</code> to <code>true</code> and +provide lower and upper bounds for the number of executors through +<code>spark.dynamicAllocation.minExecutors</code> and <code>spark.dynamicAllocation.maxExecutors</code>. Other relevant +configurations are described on the <a href="configuration.html#dynamic-allocation">configurations page</a> +and in the subsequent sections in detail.</p> + +<p>Additionally, your application must use an external shuffle service. The purpose of the service is +to preserve the shuffle files written by executors so the executors can be safely removed (more +detail described <a href="job-scheduling.html#graceful-decommission-of-executors">below</a>). To enable +this service, set <code>spark.shuffle.service.enabled</code> to <code>true</code>. In YARN, this external shuffle service +is implemented in <code>org.apache.spark.yarn.network.YarnShuffleService</code> that runs in each <code>NodeManager</code> +in your cluster. To start this service, follow these steps:</p> + +<ol> + <li>Build Spark with the <a href="building-spark.html">YARN profile</a>. Skip this step if you are using a +pre-packaged distribution.</li> + <li>Locate the <code>spark-<version>-yarn-shuffle.jar</code>. This should be under +<code>$SPARK_HOME/network/yarn/target/scala-<version></code> if you are building Spark yourself, and under +<code>lib</code> if you are using a distribution.</li> + <li>Add this jar to the classpath of all <code>NodeManager</code>s in your cluster.</li> + <li>In the <code>yarn-site.xml</code> on each node, add <code>spark_shuffle</code> to <code>yarn.nodemanager.aux-services</code>, +then set <code>yarn.nodemanager.aux-services.spark_shuffle.class</code> to +<code>org.apache.spark.network.yarn.YarnShuffleService</code>. Additionally, set all relevant +<code>spark.shuffle.service.*</code> <a href="configuration.html">configurations</a>.</li> + <li>Restart all <code>NodeManager</code>s in your cluster.</li> +</ol> + +<h3 id="resource-allocation-policy">Resource Allocation Policy</h3> + +<p>At a high level, Spark should relinquish executors when they are no longer used and acquire +executors when they are needed. Since there is no definitive way to predict whether an executor +that is about to be removed will run a task in the near future, or whether a new executor that is +about to be added will actually be idle, we need a set of heuristics to determine when to remove +and request executors.</p> + +<h4 id="request-policy">Request Policy</h4> + +<p>A Spark application with dynamic allocation enabled requests additional executors when it has +pending tasks waiting to be scheduled. This condition necessarily implies that the existing set +of executors is insufficient to simultaneously saturate all tasks that have been submitted but +not yet finished.</p> + +<p>Spark requests executors in rounds. The actual request is triggered when there have been pending +tasks for <code>spark.dynamicAllocation.schedulerBacklogTimeout</code> seconds, and then triggered again +every <code>spark.dynamicAllocation.sustainedSchedulerBacklogTimeout</code> seconds thereafter if the queue +of pending tasks persists. Additionally, the number of executors requested in each round increases +exponentially from the previous round. For instance, an application will add 1 executor in the +first round, and then 2, 4, 8 and so on executors in the subsequent rounds.</p> + +<p>The motivation for an exponential increase policy is twofold. First, an application should request +executors cautiously in the beginning in case it turns out that only a few additional executors is +sufficient. This echoes the justification for TCP slow start. Second, the application should be +able to ramp up its resource usage in a timely manner in case it turns out that many executors are +actually needed.</p> + +<h4 id="remove-policy">Remove Policy</h4> + +<p>The policy for removing executors is much simpler. A Spark application removes an executor when +it has been idle for more than <code>spark.dynamicAllocation.executorIdleTimeout</code> seconds. Note that, +under most circumstances, this condition is mutually exclusive with the request condition, in that +an executor should not be idle if there are still pending tasks to be scheduled.</p> + +<h3 id="graceful-decommission-of-executors">Graceful Decommission of Executors</h3> + +<p>Before dynamic allocation, a Spark executor exits either on failure or when the associated +application has also exited. In both scenarios, all state associated with the executor is no +longer needed and can be safely discarded. With dynamic allocation, however, the application +is still running when an executor is explicitly removed. If the application attempts to access +state stored in or written by the executor, it will have to perform a recompute the state. Thus, +Spark needs a mechanism to decommission an executor gracefully by preserving its state before +removing it.</p> + +<p>This requirement is especially important for shuffles. During a shuffle, the Spark executor first +writes its own map outputs locally to disk, and then acts as the server for those files when other +executors attempt to fetch them. In the event of stragglers, which are tasks that run for much +longer than their peers, dynamic allocation may remove an executor before the shuffle completes, +in which case the shuffle files written by that executor must be recomputed unnecessarily.</p> + +<p>The solution for preserving shuffle files is to use an external shuffle service, also introduced +in Spark 1.2. This service refers to a long-running process that runs on each node of your cluster +independently of your Spark applications and their executors. If the service is enabled, Spark +executors will fetch shuffle files from the service instead of from each other. This means any +shuffle state written by an executor may continue to be served beyond the executor’s lifetime.</p> + +<p>In addition to writing shuffle files, executors also cache data either on disk or in memory. +When an executor is removed, however, all cached data will no longer be accessible. There is +currently not yet a solution for this in Spark 1.2. In future releases, the cached data may be +preserved through an off-heap storage similar in spirit to how shuffle files are preserved through +the external shuffle service.</p> + <h1 id="scheduling-within-an-application">Scheduling Within an Application</h1> <p>Inside a given Spark application (SparkContext instance), multiple parallel jobs can run simultaneously if @@ -209,9 +332,10 @@ mode is best for multi-user settings.</p> <p>To enable the fair scheduler, simply set the <code>spark.scheduler.mode</code> property to <code>FAIR</code> when configuring a SparkContext:</p> -<div class="highlight"><pre><code class="language-scala" data-lang="scala"><span class="k">val</span> <span class="n">conf</span> <span class="k">=</span> <span class="k">new</span> <span class="nc">SparkConf</span><span class="o">().</span><span class="n">setMaster</span><span class="o">(...).</span><span class="n">setAppName</span><span class="o">(...)</span> +<div class="highlight"><pre><code class="scala"><span class="k">val</span> <span class="n">conf</span> <span class="k">=</span> <span class="k">new</span> <span class="nc">SparkConf</span><span class="o">().</span><span class="n">setMaster</span><span class="o">(...).</span><span class="n">setAppName</span><span class="o">(...)</span> <span class="n">conf</span><span class="o">.</span><span class="n">set</span><span class="o">(</span><span class="s">"spark.scheduler.mode"</span><span class="o">,</span> <span class="s">"FAIR"</span><span class="o">)</span> -<span class="k">val</span> <span class="n">sc</span> <span class="k">=</span> <span class="k">new</span> <span class="nc">SparkContext</span><span class="o">(</span><span class="n">conf</span><span class="o">)</span></code></pre></div> +<span class="k">val</span> <span class="n">sc</span> <span class="k">=</span> <span class="k">new</span> <span class="nc">SparkContext</span><span class="o">(</span><span class="n">conf</span><span class="o">)</span> +</code></pre></div> <h2 id="fair-scheduler-pools">Fair Scheduler Pools</h2> @@ -225,15 +349,17 @@ many concurrent jobs they have instead of giving <em>jobs</em> equal shares. Thi adding the <code>spark.scheduler.pool</code> “local property” to the SparkContext in the thread that’s submitting them. This is done as follows:</p> -<div class="highlight"><pre><code class="language-scala" data-lang="scala"><span class="c1">// Assuming sc is your SparkContext variable</span> -<span class="n">sc</span><span class="o">.</span><span class="n">setLocalProperty</span><span class="o">(</span><span class="s">"spark.scheduler.pool"</span><span class="o">,</span> <span class="s">"pool1"</span><span class="o">)</span></code></pre></div> +<div class="highlight"><pre><code class="scala"><span class="c1">// Assuming sc is your SparkContext variable</span> +<span class="n">sc</span><span class="o">.</span><span class="n">setLocalProperty</span><span class="o">(</span><span class="s">"spark.scheduler.pool"</span><span class="o">,</span> <span class="s">"pool1"</span><span class="o">)</span> +</code></pre></div> <p>After setting this local property, <em>all</em> jobs submitted within this thread (by calls in this thread to <code>RDD.save</code>, <code>count</code>, <code>collect</code>, etc) will use this pool name. The setting is per-thread to make it easy to have a thread run multiple jobs on behalf of the same user. If you’d like to clear the pool that a thread is associated with, simply call:</p> -<div class="highlight"><pre><code class="language-scala" data-lang="scala"><span class="n">sc</span><span class="o">.</span><span class="n">setLocalProperty</span><span class="o">(</span><span class="s">"spark.scheduler.pool"</span><span class="o">,</span> <span class="kc">null</span><span class="o">)</span></code></pre></div> +<div class="highlight"><pre><code class="scala"><span class="n">sc</span><span class="o">.</span><span class="n">setLocalProperty</span><span class="o">(</span><span class="s">"spark.scheduler.pool"</span><span class="o">,</span> <span class="kc">null</span><span class="o">)</span> +</code></pre></div> <h2 id="default-behavior-of-pools">Default Behavior of Pools</h2> @@ -267,12 +393,13 @@ of the cluster. By default, each pool’s <code>minShare</code> is 0.</li> and setting a <code>spark.scheduler.allocation.file</code> property in your <a href="configuration.html#spark-properties">SparkConf</a>.</p> -<div class="highlight"><pre><code class="language-scala" data-lang="scala"><span class="n">conf</span><span class="o">.</span><span class="n">set</span><span class="o">(</span><span class="s">"spark.scheduler.allocation.file"</span><span class="o">,</span> <span class="s">"/path/to/file"</span><span class="o">)</span></code></pre></div> +<div class="highlight"><pre><code class="scala"><span class="n">conf</span><span class="o">.</span><span class="n">set</span><span class="o">(</span><span class="s">"spark.scheduler.allocation.file"</span><span class="o">,</span> <span class="s">"/path/to/file"</span><span class="o">)</span> +</code></pre></div> <p>The format of the XML file is simply a <code><pool></code> element for each pool, with different elements within it for the various settings. For example:</p> -<div class="highlight"><pre><code class="language-xml" data-lang="xml"><span class="cp"><?xml version="1.0"?></span> +<div class="highlight"><pre><code class="xml"><span class="cp"><?xml version="1.0"?></span> <span class="nt"><allocations></span> <span class="nt"><pool</span> <span class="na">name=</span><span class="s">"production"</span><span class="nt">></span> <span class="nt"><schedulingMode></span>FAIR<span class="nt"></schedulingMode></span> @@ -284,7 +411,8 @@ within it for the various settings. For example:</p> <span class="nt"><weight></span>2<span class="nt"></weight></span> <span class="nt"><minShare></span>3<span class="nt"></minShare></span> <span class="nt"></pool></span> -<span class="nt"></allocations></span></code></pre></div> +<span class="nt"></allocations></span> +</code></pre></div> <p>A full example is also available in <code>conf/fairscheduler.xml.template</code>. Note that any pools not configured in the XML file will simply get default values for all settings (scheduling mode FIFO, |