Github user pwendell commented on a diff in the pull request:
https://github.com/apache/spark/pull/2746#discussion_r19195340
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/ExecutorAllocationManager.scala
---
@@ -0,0 +1,345 @@
+/*
+ * 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.spark.scheduler
+
+import scala.collection.mutable
+
+import org.apache.spark.{Logging, SparkException}
+import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend
+
+/**
+ * An agent that dynamically allocates and removes executors based on the
workload.
+ *
+ * The add policy depends on the number of pending tasks. If the queue of
pending tasks is not
+ * drained in N seconds, then new executors are added. If the queue
persists for another M
+ * seconds, then more executors are added and so on. The number added in
each round increases
+ * exponentially from the previous round until an upper bound on the
number of executors has
+ * been reached.
+ *
+ * The rationale for the exponential increase is twofold: (1) Executors
should be added slowly
+ * in the beginning in case the number of extra executors needed turns out
to be small. Otherwise,
+ * we may add more executors than we need just to remove them later. (2)
Executors should be added
+ * quickly over time in case the maximum number of executors is very high.
Otherwise, it will take
+ * a long time to ramp up under heavy workloads.
+ *
+ * The remove policy is simpler: If an executor has been idle for K
seconds (meaning it has not
+ * been scheduled to run any tasks), then it is removed. This requires
starting a timer on each
+ * executor instead of just starting a global one as in the add case.
+ *
+ * There is no retry logic in either case. Because the requests to the
cluster manager are
+ * asynchronous, this class does not know whether a request has been
granted until later. For
+ * this reason, both add and remove are treated as best-effort only.
+ *
+ * The relevant Spark properties include the following:
+ *
+ * spark.dynamicAllocation.enabled - Whether this feature is enabled
+ * spark.dynamicAllocation.minExecutors - Lower bound on the number of
executors
+ * spark.dynamicAllocation.maxExecutors - Upper bound on the number of
executors
+ *
+ * spark.dynamicAllocation.addExecutorThresholdSeconds - How long before
new executors are added
+ * spark.dynamicAllocation.addExecutorIntervalSeconds - How often to add
new executors
+ * spark.dynamicAllocation.removeExecutorThresholdSeconds - How long
before an executor is removed
+ *
+ * Synchronization: Because the schedulers in Spark are single-threaded,
contention should only
+ * arise when new executors register or when existing executors are
removed, both of which are
+ * relatively rare events with respect to task scheduling. Thus,
synchronizing each method on the
+ * same lock should not be expensive assuming biased locking is enabled in
the JVM (on by default
+ * for Java 6+). This may not be true, however, if the application itself
runs multiple jobs
+ * concurrently.
+ *
+ * Note: This is part of a larger implementation (SPARK-3174) and
currently does not actually
+ * request to add or remove executors. The mechanism to actually do this
will be added separately,
+ * e.g. in SPARK-3822 for Yarn.
+ */
+private[scheduler] class ExecutorAllocationManager(scheduler:
TaskSchedulerImpl) extends Logging {
+ import ExecutorAllocationManager._
+
+ private val conf = scheduler.conf
+
+ // Lower and upper bounds on the number of executors. These are required.
+ private val minNumExecutors =
conf.getInt("spark.dynamicAllocation.minExecutors", -1)
+ private val maxNumExecutors =
conf.getInt("spark.dynamicAllocation.maxExecutors", -1)
+ if (minNumExecutors < 0 || maxNumExecutors < 0) {
+ throw new SparkException("spark.dynamicAllocation.{min/max}Executors
must be set!")
+ }
+
+ // How frequently to add and remove executors (seconds)
+ private val addThresholdSeconds =
+ conf.getLong("spark.dynamicAllocation.addExecutorThresholdSeconds", 60)
+ private val addIntervalSeconds =
+ conf.getLong("spark.dynamicAllocation.addExecutorIntervalSeconds",
addThresholdSeconds)
+ private val removeThresholdSeconds =
+ conf.getLong("spark.dynamicAllocation.removeExecutorThresholdSeconds",
600)
+
+ // Number of executors to add in the next round
+ private var numExecutorsToAdd = 1
+
+ // Number of executors that have been requested but have not registered
yet
+ private var numExecutorsPendingToAdd = 0
+
+ // Executors that have been requested to be removed but have not been
killed yet
+ private val executorsPendingToRemove = new mutable.HashSet[String]
+
+ // Keep track of all executors here to decouple us from the logic in
TaskSchedulerImpl
+ private val executorIds = new mutable.HashSet[String]
+
+ // A timestamp of when the add timer should be triggered, or NOT_STARTED
if the timer is not
+ // started. This timer is started when there are pending tasks built up,
and canceled when
+ // there are no more pending tasks.
+ private var addTime = NOT_STARTED
+
+ // A timestamp for each executor of when the remove timer for that
executor should be triggered.
+ // Each remove timer is started when the executor first registers or
when the executor finishes
+ // running a task, and canceled when the executor is scheduled to run a
new task.
+ private val removeTimes = new mutable.HashMap[String, Long]
+
+ // A timestamp of when all pending add requests should expire
+ private var pendingAddExpirationTime = NOT_STARTED
+
+ // A timestamp for each executor of when the pending remove request for
the executor should expire
+ private val pendingRemoveExpirationTimes = new mutable.HashMap[String,
Long]
+
+ // How long before expiring pending requests to add or remove executors
(seconds)
+ private val pendingAddTimeoutSeconds = 300 // 5 min
+ private val pendingRemoveTimeoutSeconds = 300
+
+ // Polling loop interval (ms)
+ private val intervalMillis = 100
+
+ // Scheduler backend through which requests to add/remove executors are
made
+ // Note that this assumes the backend has already initialized when this
is first used
+ // Otherwise, an appropriate exception is thrown
+ private lazy val backend = scheduler.backend match {
+ case b: CoarseGrainedSchedulerBackend => b
+ case null =>
+ throw new SparkException("Scheduler backend not initialized yet!")
+ case _ =>
+ throw new SparkException(
+ "Dynamic allocation of executors is not applicable to fine-grained
schedulers. " +
+ "Please set spark.dynamicAllocation.enabled to false.")
+ }
+
+ initialize()
+
+ /**
+ * Start the main polling thread that keeps track of when to add and
remove executors.
+ * During each loop interval, this thread checks if any of the timers
have timed out, and,
+ * if so, triggers the relevant timer actions.
+ */
+ def initialize(): Unit = {
+ val thread = new Thread {
+ override def run(): Unit = {
+ while (true) {
+ ExecutorAllocationManager.this.synchronized {
+ val now = System.currentTimeMillis
+ try {
+ // If the add timer has timed out, add executors and refresh
the timer
+ if (addTime != NOT_STARTED && now >= addTime) {
+ addExecutors()
+ logDebug(s"Restarting add executor timer " +
+ s"(to be triggered in $addIntervalSeconds seconds)")
+ addTime += addIntervalSeconds * 1000
+ }
+
+ // If a remove timer has timed out, remove the executor and
cancel the timer
+ removeTimes.foreach { case (executorId, triggerTime) =>
+ if (now > triggerTime) {
+ removeExecutor(executorId)
+ cancelRemoveTimer(executorId)
+ }
+ }
+
+ // Expire any outstanding pending add requests that have
timed out
+ if (pendingAddExpirationTime != NOT_STARTED && now >=
pendingAddExpirationTime) {
+ logDebug(s"Expiring all pending add requests because they
have " +
+ s"not been fulfilled after $pendingAddTimeoutSeconds
seconds")
+ numExecutorsPendingToAdd = 0
+ pendingAddExpirationTime = NOT_STARTED
+ }
+
+ // Expire any outstanding pending remove requests that have
timed out
+ pendingRemoveExpirationTimes.foreach { case (executorId,
expirationTime) =>
+ if (now > expirationTime) {
+ logDebug(s"Expiring pending request to remove executor
$executorId because " +
+ s"it has not been fulfilled after
$pendingRemoveTimeoutSeconds seconds")
+ executorsPendingToRemove.remove(executorId)
+ pendingRemoveExpirationTimes.remove(executorId)
+ }
+ }
+ } catch {
+ case e: Exception => logError("Exception in dynamic executor
allocation thread!", e)
+ }
+ }
+ Thread.sleep(intervalMillis)
+ }
+ }
+ }
+ thread.setName("spark-dynamic-executor-allocation")
+ thread.setDaemon(true)
+ thread.start()
+ }
+
+ /**
+ * Request a number of executors from the scheduler backend.
+ * If the cap on the number of executors is reached, give up and reset
the
+ * number of executors to add next round instead of continuing to double
it.
+ */
+ private def addExecutors(): Unit = synchronized {
+ // Do not request more executors if we have already reached the upper
bound
+ val numExistingExecutors = executorIds.size + numExecutorsPendingToAdd
+ if (numExistingExecutors >= maxNumExecutors) {
+ logDebug(s"Not adding executors because there are already " +
+ s"$maxNumExecutors executor(s), which is the limit")
+ numExecutorsToAdd = 1
+ return
+ }
+
+ // Request executors with respect to the upper bound
+ val actualNumExecutorsToAdd =
+ math.min(numExistingExecutors + numExecutorsToAdd, maxNumExecutors)
- numExistingExecutors
--- End diff --
This expression is quite complicated, so I'd break it up a bit:
```
// Number to add if continuing exponential increase
val targetNumExecutors = executorIds.size + numExecutorsPending +
numExecutorsToAdd
// Take into account max
val adjustedTargetNumExecutors = math.min(targetNumExecutors,
maxNumExecutors)
// Compute delta
val adjustedNumExecutorsToAdd = adjustedTargetNumExecutors -
numExistingExecutors
```
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
