tgravescs commented on code in PR #43494:
URL: https://github.com/apache/spark/pull/43494#discussion_r1372201798
##########
core/src/main/scala/org/apache/spark/resource/ResourceAllocator.scala:
##########
@@ -29,59 +65,45 @@ private[spark] trait ResourceAllocator {
protected def resourceName: String
protected def resourceAddresses: Seq[String]
- protected def slotsPerAddress: Int
/**
- * Map from an address to its availability, a value > 0 means the address is
available,
- * while value of 0 means the address is fully assigned.
- *
- * For task resources ([[org.apache.spark.scheduler.ExecutorResourceInfo]]),
this value
- * can be a multiple, such that each address can be allocated up to
[[slotsPerAddress]]
- * times.
+ * Map from an address to its availability default to RESOURCE_TOTAL_AMOUNT,
a value > 0 means
+ * the address is available, while value of 0 means the address is fully
assigned.
*/
- private lazy val addressAvailabilityMap = {
- mutable.HashMap(resourceAddresses.map(_ -> slotsPerAddress): _*)
+ protected lazy val addressAvailabilityMap = {
+ mutable.HashMap(resourceAddresses.map(address => address ->
RESOURCE_TOTAL_AMOUNT): _*)
}
/**
- * Sequence of currently available resource addresses.
- *
- * With [[slotsPerAddress]] greater than 1, [[availableAddrs]] can contain
duplicate addresses
- * e.g. with [[slotsPerAddress]] == 2, availableAddrs for addresses 0 and 1
can look like
- * Seq("0", "0", "1"), where address 0 has two assignments available, and 1
has one.
+ * Sequence of currently available resource addresses which is not fully
assigned.
Review Comment:
nit, should be: "which are not fully assigned"
##########
core/src/main/scala/org/apache/spark/deploy/master/WorkerInfo.scala:
##########
@@ -28,12 +29,24 @@ private[spark] case class WorkerResourceInfo(name: String,
addresses: Seq[String
override protected def resourceName = this.name
override protected def resourceAddresses = this.addresses
- override protected def slotsPerAddress: Int = 1
+ /**
+ * Acquire the resources.
+ * @param amount How many addresses are requesting.
+ * @return ResourceInformation
+ */
def acquire(amount: Int): ResourceInformation = {
- val allocated = availableAddrs.take(amount)
- acquire(allocated)
- new ResourceInformation(resourceName, allocated.toArray)
+
+ var count = amount
+ val allocated: mutable.HashMap[String, Double] = mutable.HashMap.empty
+ for (address <- availableAddrs if count > 0) {
+ if (addressAvailabilityMap(address) == RESOURCE_TOTAL_AMOUNT) {
Review Comment:
why do we need this check if WorkerInfo is always dealing with full
Resources/GPU's? isn't this just availableAddrs still?
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourcesAmounts.scala:
##########
@@ -0,0 +1,199 @@
+/*
+ * 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.HashMap
+
+import org.apache.spark.SparkException
+import org.apache.spark.resource.{ResourceInformation, ResourceProfile}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
+
+/**
+ * Class to hold information about a series of resources belonging to an
executor.
+ * A resource could be a GPU, FPGA, etc. And it is used as a temporary
+ * class to calculate the resources amounts when offering resources to
+ * the tasks in the [[TaskSchedulerImpl]]
+ *
+ * One example is GPUs, where the addresses would be the indices of the GPUs
+ *
+ * @param resources The executor available resources and amount. eg,
+ * Map("gpu" -> mutable.Map("0" -> 0.2, "1" -> 1.0),
+ * "fpga" -> mutable.Map("a" -> 0.3, "b" -> 0.9)
+ * )
+ */
+private[spark] class ExecutorResourcesAmounts(
+ private val resources: Map[String, Map[String, Double]]) extends
Serializable {
+
+ resources.foreach { case (_, addressMount) =>
Review Comment:
this seems like a weird assert because if asserts off I think its still
going to do the foreach. Either rewrite this or remove it.
##########
core/src/main/scala/org/apache/spark/resource/TaskResourceRequest.scala:
##########
@@ -37,8 +37,8 @@ import org.apache.spark.annotation.{Evolving, Since}
class TaskResourceRequest(val resourceName: String, val amount: Double)
extends Serializable {
- assert(amount <= 0.5 || amount % 1 == 0,
- s"The resource amount ${amount} must be either <= 0.5, or a whole number.")
+ assert(amount <= 1.0 || amount % 1 == 0,
Review Comment:
same here, this shouldn't change behavior for dynamic allocation
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourcesAmounts.scala:
##########
@@ -0,0 +1,199 @@
+/*
+ * 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.HashMap
+
+import org.apache.spark.SparkException
+import org.apache.spark.resource.{ResourceInformation, ResourceProfile}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
+
+/**
+ * Class to hold information about a series of resources belonging to an
executor.
+ * A resource could be a GPU, FPGA, etc. And it is used as a temporary
+ * class to calculate the resources amounts when offering resources to
+ * the tasks in the [[TaskSchedulerImpl]]
+ *
+ * One example is GPUs, where the addresses would be the indices of the GPUs
+ *
+ * @param resources The executor available resources and amount. eg,
+ * Map("gpu" -> mutable.Map("0" -> 0.2, "1" -> 1.0),
+ * "fpga" -> mutable.Map("a" -> 0.3, "b" -> 0.9)
+ * )
+ */
+private[spark] class ExecutorResourcesAmounts(
+ private val resources: Map[String, Map[String, Double]]) extends
Serializable {
+
+ resources.foreach { case (_, addressMount) =>
+ addressMount.foreach { case (_, amount) => assert(amount <= 1.0)}}
+
+ // multiply the RESOURCE_TOTAL_AMOUNT to avoid using double directly.
+ // and convert the addressesAmounts to be mutable.HashMap
+ private val internalResources: Map[String, HashMap[String, Long]] = {
+ resources.map { case (rName, addressAmounts) =>
+ rName -> HashMap(addressAmounts.map { case (address, amount) =>
+ address -> (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ }.toSeq: _*)
+ }
+ }
+
+ // It maps from the resource name to its amount.
+ lazy val resourceAmount: Map[String, Int] = internalResources.map { case
(rName, addressMap) =>
+ rName -> addressMap.size
+ }
+
+ // convert internal resources back to the public.
+ def availableResources: Map[String, Map[String, Double]] = {
+ internalResources.map { case (rName, addressMap) =>
+ rName -> addressMap.map { case (address, amount) =>
+ address -> amount.toDouble / RESOURCE_TOTAL_AMOUNT
+ }.toMap
+ }
+ }
+
+ // Acquire the resource and update the resource
+ def acquire(assignedResource: Map[String, Map[String, Double]]): Unit = {
+ assignedResource.foreach { case (rName, taskResAmounts) =>
+ val availableResourceAmounts = internalResources.getOrElse(rName,
+ throw new SparkException(s"Try to acquire an address from $rName that
doesn't exist"))
+ taskResAmounts.foreach { case (address, amount) =>
+ val prevInternalTotalAmount =
availableResourceAmounts.getOrElse(address,
+ throw new SparkException(s"Try to acquire an address that doesn't
exist. $rName " +
+ s"address $address doesn't exist."))
+
+ val internalTaskAmount = (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ val internalLeft = prevInternalTotalAmount - internalTaskAmount
+ val realLeft = internalLeft.toDouble / RESOURCE_TOTAL_AMOUNT
+ if (realLeft < 0) {
+ throw new SparkException(s"The total amount ${realLeft} " +
+ s"after acquiring $rName address $address should be >= 0")
+ }
+ internalResources(rName)(address) = internalLeft
+ // scalastyle:off println
+ println(s"Acquired. left ${realLeft}")
+ // scalastyle:on println
+ }
+ }
+ }
+
+ // release the resources and update the values
+ def release(assignedResource: Map[String, Map[String, Double]]): Unit = {
+ assignedResource.foreach { case (rName, taskResAmounts) =>
+ val availableResourceAmounts = internalResources.getOrElse(rName,
+ throw new SparkException(s"Try to release an address from $rName that
doesn't exist"))
+ taskResAmounts.foreach { case (address, amount) =>
+ val prevInternalTotalAmount =
availableResourceAmounts.getOrElse(address,
+ throw new SparkException(s"Try to release an address that is not
assigned. $rName " +
+ s"address $address is not assigned."))
+ val internalTaskAmount = (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ val internalTotal = prevInternalTotalAmount + internalTaskAmount
+ if (internalTotal > RESOURCE_TOTAL_AMOUNT) {
+ throw new SparkException(s"The total amount " +
+ s"${internalTotal.toDouble / RESOURCE_TOTAL_AMOUNT} " +
+ s"after releasing $rName address $address should be <= 1.0")
+ }
+ internalResources(rName)(address) = internalTotal
+ // scalastyle:off println
+ println(s"Released. amount ${internalTotal.toDouble /
RESOURCE_TOTAL_AMOUNT}")
+ // scalastyle:on println
+ }
+ }
+ }
+
+ // Try to assign the address according to the task requirement.
+ // Please note that this function will not update the values.
+ def assignResources(taskSetProf: ResourceProfile):
+ Option[(Map[String, ResourceInformation], Map[String, Map[String,
Double]])] = {
+
+ // only look at the resource other than cpus
+ val tsResources = taskSetProf.getCustomTaskResources()
+ if (tsResources.isEmpty) {
+ return Some(Map.empty, Map.empty)
+ }
+
+ val localTaskReqAssign = HashMap[String, ResourceInformation]()
+ val allocatedAddresses = HashMap[String, Map[String, Double]]()
+
+ // we go through all resources here so that we can make sure they match
and also get what the
+ // assignments are for the next task
+ for ((rName, taskReqs) <- tsResources) {
+ // if taskAmount = 1.5, we assign 2.0 gpu for user or
+ // just throw an exception in a very begging?
+ // TODO, just remove it, since we enabled the check at the very
beginning.
+ val taskAmount = if (taskReqs.amount < 1.0) taskReqs.amount else
Math.ceil(taskReqs.amount)
Review Comment:
why isn't this just using
taskSetProf.getSchedulerTaskResourceAmount(rName) and that function updated to
do the right thing ?
Also there is a comment in Resourceprofile.scala in the
getSchedulerTaskResourceAmount that assumes addresses are done old way.
I'm again worried about this changing the behavior with dynamic allocation
on.
##########
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala:
##########
@@ -389,7 +389,7 @@ private[spark] class TaskSchedulerImpl(
maxLocality: TaskLocality,
shuffledOffers: Seq[WorkerOffer],
availableCpus: Array[Int],
- availableResources: Array[Map[String, Buffer[String]]],
+ availableResources: Array[ExecutorResourcesAmounts],
Review Comment:
did the caller of this get updated? it doesn't look like it unless I'm
missing it or my ide didn't update
##########
core/src/main/scala/org/apache/spark/resource/ResourceAllocator.scala:
##########
@@ -29,59 +65,45 @@ private[spark] trait ResourceAllocator {
protected def resourceName: String
protected def resourceAddresses: Seq[String]
- protected def slotsPerAddress: Int
/**
- * Map from an address to its availability, a value > 0 means the address is
available,
- * while value of 0 means the address is fully assigned.
- *
- * For task resources ([[org.apache.spark.scheduler.ExecutorResourceInfo]]),
this value
- * can be a multiple, such that each address can be allocated up to
[[slotsPerAddress]]
- * times.
+ * Map from an address to its availability default to RESOURCE_TOTAL_AMOUNT,
a value > 0 means
+ * the address is available, while value of 0 means the address is fully
assigned.
*/
- private lazy val addressAvailabilityMap = {
- mutable.HashMap(resourceAddresses.map(_ -> slotsPerAddress): _*)
+ protected lazy val addressAvailabilityMap = {
+ mutable.HashMap(resourceAddresses.map(address => address ->
RESOURCE_TOTAL_AMOUNT): _*)
}
/**
- * Sequence of currently available resource addresses.
- *
- * With [[slotsPerAddress]] greater than 1, [[availableAddrs]] can contain
duplicate addresses
- * e.g. with [[slotsPerAddress]] == 2, availableAddrs for addresses 0 and 1
can look like
- * Seq("0", "0", "1"), where address 0 has two assignments available, and 1
has one.
+ * Sequence of currently available resource addresses which is not fully
assigned.
*/
def availableAddrs: Seq[String] = addressAvailabilityMap
- .flatMap { case (addr, available) =>
- (0 until available).map(_ => addr)
- }.toSeq.sorted
+ .filter(addresses => addresses._2 > 0).keys.toSeq.sorted
/**
* Sequence of currently assigned resource addresses.
Review Comment:
seems like this now returns something quite a bit different. I think this
goes to the UI so I assume the UI is losing details now, right? ie will just
say if one resource is used but not how much of it. I think we should find a
better way to display what is really used
##########
core/src/main/scala/org/apache/spark/resource/ResourceUtils.scala:
##########
@@ -170,16 +170,16 @@ private[spark] object ResourceUtils extends Logging {
// integer amount and the number of slots per address. For instance, if the
amount is 0.5,
// the we get (1, 2) back out. This indicates that for each 1 address, it
has 2 slots per
// address, which allows you to put 2 tasks on that address. Note if amount
is greater
- // than 1, then the number of slots per address has to be 1. This would
indicate that a
+ // than 1, then the number of parts per address has to be 1. This would
indicate that a
// would have multiple addresses assigned per task. This can be used for
calculating
// the number of tasks per executor -> (executorAmount * numParts) /
(integer amount).
// Returns tuple of (integer amount, numParts)
def calculateAmountAndPartsForFraction(doubleAmount: Double): (Int, Int) = {
- val parts = if (doubleAmount <= 0.5) {
+ val parts = if (doubleAmount <= 1.0) {
Review Comment:
so this changes this condition for all the resourceprofiles, even with
dynamic allocation, which doesn't support changing task resources within a
profile. I don't think we want that, or at least shouldn't be part of the
scope of this change.
##########
core/src/main/scala/org/apache/spark/deploy/master/WorkerInfo.scala:
##########
@@ -28,12 +29,24 @@ private[spark] case class WorkerResourceInfo(name: String,
addresses: Seq[String
override protected def resourceName = this.name
override protected def resourceAddresses = this.addresses
- override protected def slotsPerAddress: Int = 1
+ /**
+ * Acquire the resources.
+ * @param amount How many addresses are requesting.
+ * @return ResourceInformation
+ */
def acquire(amount: Int): ResourceInformation = {
- val allocated = availableAddrs.take(amount)
- acquire(allocated)
- new ResourceInformation(resourceName, allocated.toArray)
+
+ var count = amount
+ val allocated: mutable.HashMap[String, Double] = mutable.HashMap.empty
+ for (address <- availableAddrs if count > 0) {
+ if (addressAvailabilityMap(address) == RESOURCE_TOTAL_AMOUNT) {
+ allocated(address) = 1.0
Review Comment:
This code is weird to read by itself. please add a comment explaining why
doing this vs calling availableAddrs. I think its because this is worker and
need to do full GPU to an executor and I guess we just reuse the same
ResourceAllocator class
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourceInfo.scala:
##########
@@ -18,23 +18,45 @@
package org.apache.spark.scheduler
import org.apache.spark.resource.{ResourceAllocator, ResourceInformation}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
/**
* Class to hold information about a type of Resource on an Executor. This
information is managed
* by SchedulerBackend, and TaskScheduler shall schedule tasks on idle
Executors based on the
* information.
* @param name Resource name
* @param addresses Resource addresses provided by the executor
- * @param numParts Number of ways each resource is subdivided when scheduling
tasks
*/
private[spark] class ExecutorResourceInfo(
name: String,
- addresses: Seq[String],
- numParts: Int)
+ addresses: Seq[String])
extends ResourceInformation(name, addresses.toArray) with ResourceAllocator {
override protected def resourceName = this.name
+
override protected def resourceAddresses = this.addresses
- override protected def slotsPerAddress: Int = numParts
- def totalAddressAmount: Int = resourceAddresses.length * slotsPerAddress
+
+ /**
+ * Calculate how many parts the executor can offer according to the task
resource amount
+ * @param taskAmount how many resource amount the task required
+ * @return the total parts
+ */
+ def totalParts(taskAmount: Double): Int = {
+ assert(taskAmount > 0.0)
+ if (taskAmount >= 1.0) {
+ addresses.length / taskAmount.ceil.toInt
+ } else {
+ addresses.length * Math.floor(1.0 / taskAmount).toInt
+ }
+ }
+
+ /**
+ * Convert the internal address availability to the public resource format
Review Comment:
what is a public resource format?
##########
core/src/main/scala/org/apache/spark/resource/ResourceAllocator.scala:
##########
@@ -29,59 +65,45 @@ private[spark] trait ResourceAllocator {
protected def resourceName: String
protected def resourceAddresses: Seq[String]
- protected def slotsPerAddress: Int
/**
- * Map from an address to its availability, a value > 0 means the address is
available,
- * while value of 0 means the address is fully assigned.
- *
- * For task resources ([[org.apache.spark.scheduler.ExecutorResourceInfo]]),
this value
- * can be a multiple, such that each address can be allocated up to
[[slotsPerAddress]]
- * times.
+ * Map from an address to its availability default to RESOURCE_TOTAL_AMOUNT,
a value > 0 means
+ * the address is available, while value of 0 means the address is fully
assigned.
*/
- private lazy val addressAvailabilityMap = {
- mutable.HashMap(resourceAddresses.map(_ -> slotsPerAddress): _*)
+ protected lazy val addressAvailabilityMap = {
Review Comment:
why open up permissions here? I see this is accessed from WorkerInfo but I
question if that is the right way to do that.
##########
core/src/main/scala/org/apache/spark/resource/ResourceAllocator.scala:
##########
@@ -29,59 +65,45 @@ private[spark] trait ResourceAllocator {
protected def resourceName: String
protected def resourceAddresses: Seq[String]
- protected def slotsPerAddress: Int
/**
- * Map from an address to its availability, a value > 0 means the address is
available,
- * while value of 0 means the address is fully assigned.
- *
- * For task resources ([[org.apache.spark.scheduler.ExecutorResourceInfo]]),
this value
- * can be a multiple, such that each address can be allocated up to
[[slotsPerAddress]]
- * times.
+ * Map from an address to its availability default to RESOURCE_TOTAL_AMOUNT,
a value > 0 means
+ * the address is available, while value of 0 means the address is fully
assigned.
*/
- private lazy val addressAvailabilityMap = {
- mutable.HashMap(resourceAddresses.map(_ -> slotsPerAddress): _*)
+ protected lazy val addressAvailabilityMap = {
+ mutable.HashMap(resourceAddresses.map(address => address ->
RESOURCE_TOTAL_AMOUNT): _*)
}
/**
- * Sequence of currently available resource addresses.
- *
- * With [[slotsPerAddress]] greater than 1, [[availableAddrs]] can contain
duplicate addresses
- * e.g. with [[slotsPerAddress]] == 2, availableAddrs for addresses 0 and 1
can look like
- * Seq("0", "0", "1"), where address 0 has two assignments available, and 1
has one.
+ * Sequence of currently available resource addresses which is not fully
assigned.
*/
def availableAddrs: Seq[String] = addressAvailabilityMap
- .flatMap { case (addr, available) =>
- (0 until available).map(_ => addr)
- }.toSeq.sorted
+ .filter(addresses => addresses._2 > 0).keys.toSeq.sorted
Review Comment:
indentation looks off here
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourcesAmounts.scala:
##########
@@ -0,0 +1,199 @@
+/*
+ * 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.HashMap
+
+import org.apache.spark.SparkException
+import org.apache.spark.resource.{ResourceInformation, ResourceProfile}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
+
+/**
+ * Class to hold information about a series of resources belonging to an
executor.
+ * A resource could be a GPU, FPGA, etc. And it is used as a temporary
+ * class to calculate the resources amounts when offering resources to
+ * the tasks in the [[TaskSchedulerImpl]]
+ *
+ * One example is GPUs, where the addresses would be the indices of the GPUs
+ *
+ * @param resources The executor available resources and amount. eg,
+ * Map("gpu" -> mutable.Map("0" -> 0.2, "1" -> 1.0),
+ * "fpga" -> mutable.Map("a" -> 0.3, "b" -> 0.9)
+ * )
+ */
+private[spark] class ExecutorResourcesAmounts(
+ private val resources: Map[String, Map[String, Double]]) extends
Serializable {
+
+ resources.foreach { case (_, addressMount) =>
+ addressMount.foreach { case (_, amount) => assert(amount <= 1.0)}}
+
+ // multiply the RESOURCE_TOTAL_AMOUNT to avoid using double directly.
+ // and convert the addressesAmounts to be mutable.HashMap
+ private val internalResources: Map[String, HashMap[String, Long]] = {
+ resources.map { case (rName, addressAmounts) =>
+ rName -> HashMap(addressAmounts.map { case (address, amount) =>
+ address -> (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ }.toSeq: _*)
+ }
+ }
+
+ // It maps from the resource name to its amount.
+ lazy val resourceAmount: Map[String, Int] = internalResources.map { case
(rName, addressMap) =>
+ rName -> addressMap.size
+ }
+
+ // convert internal resources back to the public.
Review Comment:
define what public is
##########
core/src/main/scala/org/apache/spark/scheduler/TaskDescription.scala:
##########
@@ -58,6 +58,7 @@ private[spark] class TaskDescription(
val properties: Properties,
val cpus: Int,
val resources: immutable.Map[String, ResourceInformation],
+ val resourcesAmounts: immutable.Map[String, immutable.Map[String, Double]],
Review Comment:
this is resource amounts needed by the task, assigned to the task, other?
please clarify name and description if needed. Ideally explain what the map
is in the inner part.
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourcesAmounts.scala:
##########
@@ -0,0 +1,199 @@
+/*
+ * 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.HashMap
+
+import org.apache.spark.SparkException
+import org.apache.spark.resource.{ResourceInformation, ResourceProfile}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
+
+/**
+ * Class to hold information about a series of resources belonging to an
executor.
+ * A resource could be a GPU, FPGA, etc. And it is used as a temporary
+ * class to calculate the resources amounts when offering resources to
+ * the tasks in the [[TaskSchedulerImpl]]
+ *
+ * One example is GPUs, where the addresses would be the indices of the GPUs
+ *
+ * @param resources The executor available resources and amount. eg,
+ * Map("gpu" -> mutable.Map("0" -> 0.2, "1" -> 1.0),
+ * "fpga" -> mutable.Map("a" -> 0.3, "b" -> 0.9)
+ * )
+ */
+private[spark] class ExecutorResourcesAmounts(
+ private val resources: Map[String, Map[String, Double]]) extends
Serializable {
+
+ resources.foreach { case (_, addressMount) =>
+ addressMount.foreach { case (_, amount) => assert(amount <= 1.0)}}
+
+ // multiply the RESOURCE_TOTAL_AMOUNT to avoid using double directly.
+ // and convert the addressesAmounts to be mutable.HashMap
+ private val internalResources: Map[String, HashMap[String, Long]] = {
+ resources.map { case (rName, addressAmounts) =>
+ rName -> HashMap(addressAmounts.map { case (address, amount) =>
+ address -> (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ }.toSeq: _*)
+ }
+ }
+
+ // It maps from the resource name to its amount.
+ lazy val resourceAmount: Map[String, Int] = internalResources.map { case
(rName, addressMap) =>
+ rName -> addressMap.size
+ }
+
+ // convert internal resources back to the public.
+ def availableResources: Map[String, Map[String, Double]] = {
Review Comment:
if only available for testing, state so, otherwise this should be private
##########
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala:
##########
@@ -468,33 +463,15 @@ private[spark] class TaskSchedulerImpl(
private def resourcesMeetTaskRequirements(
Review Comment:
description of this should be updated.
##########
core/src/main/scala/org/apache/spark/scheduler/ExecutorResourcesAmounts.scala:
##########
@@ -0,0 +1,199 @@
+/*
+ * 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.HashMap
+
+import org.apache.spark.SparkException
+import org.apache.spark.resource.{ResourceInformation, ResourceProfile}
+import org.apache.spark.resource.ResourceAmountUtils.RESOURCE_TOTAL_AMOUNT
+
+/**
+ * Class to hold information about a series of resources belonging to an
executor.
+ * A resource could be a GPU, FPGA, etc. And it is used as a temporary
+ * class to calculate the resources amounts when offering resources to
+ * the tasks in the [[TaskSchedulerImpl]]
+ *
+ * One example is GPUs, where the addresses would be the indices of the GPUs
+ *
+ * @param resources The executor available resources and amount. eg,
+ * Map("gpu" -> mutable.Map("0" -> 0.2, "1" -> 1.0),
+ * "fpga" -> mutable.Map("a" -> 0.3, "b" -> 0.9)
+ * )
+ */
+private[spark] class ExecutorResourcesAmounts(
+ private val resources: Map[String, Map[String, Double]]) extends
Serializable {
+
+ resources.foreach { case (_, addressMount) =>
+ addressMount.foreach { case (_, amount) => assert(amount <= 1.0)}}
+
+ // multiply the RESOURCE_TOTAL_AMOUNT to avoid using double directly.
+ // and convert the addressesAmounts to be mutable.HashMap
+ private val internalResources: Map[String, HashMap[String, Long]] = {
+ resources.map { case (rName, addressAmounts) =>
+ rName -> HashMap(addressAmounts.map { case (address, amount) =>
+ address -> (amount * RESOURCE_TOTAL_AMOUNT).toLong
+ }.toSeq: _*)
+ }
+ }
+
+ // It maps from the resource name to its amount.
Review Comment:
this might need more description, this seems to be from a resource name
("GPU") to the number of addresses that have some capacity. It looks like its
only used for testing? If so add a comment saying that
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
