Apache
/
hadoop
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

YARN-9173. FairShare calculation broken for large values after YARN-8833. Contributed Wilfred Spiegelenburg.

This commit is contained in:
Weiwei Yang 2019-01-08 13:56:21 +08:00
parent 3570713a69
commit 91e9c9f96e
3 changed files with 125 additions and 54 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/policies/ComputeFairShares.java
View File

@ -24,6 +24,8 @@ import org.apache.hadoop.yarn.api.records.Resource;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSQueue;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.Schedulable;
import static java.lang.Math.addExact;
/**
* Contains logic for computing the fair shares. A {@link Schedulable}'s fair
* share is {@link Resource} it is entitled to, independent of the current
@ -31,10 +33,13 @@ import org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.Schedulable;
* consumption lies at or below its fair share will never have its containers
* preempted.
*/
public class ComputeFairShares {
public final class ComputeFairShares {
private static final int COMPUTE_FAIR_SHARES_ITERATIONS = 25;
private ComputeFairShares() {
}
/**
* Compute fair share of the given schedulables.Fair share is an allocation of
* shares considering only active schedulables ie schedulables which have
@ -100,19 +105,20 @@ public class ComputeFairShares {
* all Schedulables are only given their minShare) and an upper bound computed
* to be large enough that too many slots are given (by doubling R until we
* use more than totalResources resources). The helper method
* resourceUsedWithWeightToResourceRatio computes the total resources used with a
* given value of R.
* resourceUsedWithWeightToResourceRatio computes the total resources used
* with a given value of R.
* <p>
* The running time of this algorithm is linear in the number of Schedulables,
* because resourceUsedWithWeightToResourceRatio is linear-time and the number of
* iterations of binary search is a constant (dependent on desired precision).
* because resourceUsedWithWeightToResourceRatio is linear-time and the
* number of iterations of binary search is a constant (dependent on desired
* precision).
*/
private static void computeSharesInternal(
Collection<? extends Schedulable> allSchedulables,
Resource totalResources, String type, boolean isSteadyShare) {
Collection<Schedulable> schedulables = new ArrayList<>();
int takenResources = handleFixedFairShares(
long takenResources = handleFixedFairShares(
allSchedulables, schedulables, isSteadyShare, type);
if (schedulables.isEmpty()) {
@ -121,12 +127,11 @@ public class ComputeFairShares {
// Find an upper bound on R that we can use in our binary search. We start
// at R = 1 and double it until we have either used all the resources or we
// have met all Schedulables' max shares.
int totalMaxShare = 0;
long totalMaxShare = 0;
for (Schedulable sched : schedulables) {
long maxShare = sched.getMaxShare().getResourceValue(type);
totalMaxShare = (int) Math.min(maxShare + (long)totalMaxShare,
Integer.MAX_VALUE);
if (totalMaxShare == Integer.MAX_VALUE) {
totalMaxShare = safeAdd(maxShare, totalMaxShare);
if (totalMaxShare == Long.MAX_VALUE) {
break;
}
}
@ -166,23 +171,24 @@ public class ComputeFairShares {
target = sched.getFairShare();
}
target.setResourceValue(type, (long)computeShare(sched, right, type));
target.setResourceValue(type, computeShare(sched, right, type));
}
}
/**
* Compute the resources that would be used given a weight-to-resource ratio
* w2rRatio, for use in the computeFairShares algorithm as described in #
* w2rRatio, for use in the computeFairShares algorithm as described in
* {@link #computeSharesInternal}.
*/
private static long resourceUsedWithWeightToResourceRatio(double w2rRatio,
Collection<? extends Schedulable> schedulables, String type) {
long resourcesTaken = 0;
for (Schedulable sched : schedulables) {
long share = computeShare(sched, w2rRatio, type);
if (Long.MAX_VALUE - resourcesTaken < share) {
return Long.MAX_VALUE;
resourcesTaken = safeAdd(resourcesTaken, share);
if (resourcesTaken == Long.MAX_VALUE) {
break;
}
resourcesTaken += share;
}
return resourcesTaken;
}
@ -204,11 +210,11 @@ public class ComputeFairShares {
* Returns the resources taken by fixed fairshare schedulables,
* and adds the remaining to the passed nonFixedSchedulables.
*/
private static int handleFixedFairShares(
private static long handleFixedFairShares(
Collection<? extends Schedulable> schedulables,
Collection<Schedulable> nonFixedSchedulables,
boolean isSteadyShare, String type) {
int totalResource = 0;
long totalResource = 0;
for (Schedulable sched : schedulables) {
long fixedShare = getFairShareIfFixed(sched, isSteadyShare, type);
@ -224,15 +230,15 @@ public class ComputeFairShares {
}
target.setResourceValue(type, fixedShare);
totalResource = (int) Math.min((long)totalResource + (long)fixedShare,
Integer.MAX_VALUE);
totalResource = safeAdd(totalResource, fixedShare);
}
}
return totalResource;
}
/**
* Get the fairshare for the {@link Schedulable} if it is fixed, -1 otherwise.
* Get the fairshare for the {@link Schedulable} if it is fixed,
* -1 otherwise.
*
* The fairshare is fixed if either the maxShare is 0, weight is 0,
* or the Schedulable is not active for instantaneous fairshare.
@ -259,4 +265,20 @@ public class ComputeFairShares {
return -1;
}
/**
* Safely add two long values. The result will always be a valid long value.
* If the addition caused an overflow the return value will be set to
* <code>Long.MAX_VALUE</code>.
* @param a first long to add
* @param b second long to add
* @return result of the addition
*/
private static long safeAdd(long a, long b) {
try {
return addExact(a, b);
} catch (ArithmeticException ae) {
return Long.MAX_VALUE;
}
}
}

11
hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FakeSchedulable.java
View File

@ -70,6 +70,17 @@ public class FakeSchedulable implements Schedulable {
Resources.createResource(0, 0), 0);
}
public FakeSchedulable(long minShare, long maxShare) {
this(minShare, maxShare, 1L);
}
public FakeSchedulable(long minShare, long maxShare, float weights) {
this(Resources.createResource(minShare, 0),
Resources.createResource(maxShare, 0),
weights, Resources.createResource(0, 0),
Resources.createResource(0, 0), 0);
}
public FakeSchedulable(Resource minShare, Resource maxShare,
float weight, Resource fairShare, Resource usage, long startTime) {
this.minShare = minShare;

102
hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/TestComputeFairShares.java
View File

@ -19,10 +19,8 @@
package org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair;
import java.util.ArrayList;
import java.util.Collection;
import java.util.List;
import org.apache.hadoop.yarn.api.records.Resource;
import org.apache.hadoop.yarn.api.records.ResourceInformation;
import org.junit.Assert;
@ -40,7 +38,7 @@ public class TestComputeFairShares {
@Before
public void setUp() throws Exception {
scheds = new ArrayList<Schedulable>();
scheds = new ArrayList<>();
}
/**
@ -149,20 +147,70 @@ public class TestComputeFairShares {
}
/**
* Test that shares are computed accurately even when the number of slots is
* very large.
* Test that shares are computed accurately even when the number of
* resources is very large.
* Test adapted to accommodate long values for resources.
*/
@Test
public void testLargeShares() {
int million = 1000 * 1000;
scheds.add(new FakeSchedulable());
scheds.add(new FakeSchedulable());
scheds.add(new FakeSchedulable());
scheds.add(new FakeSchedulable());
long giga = 1000L * 1000L * 1000L * 4L;
scheds.add(new FakeSchedulable(0L, giga));
scheds.add(new FakeSchedulable(0L, giga));
scheds.add(new FakeSchedulable(0L, giga));
scheds.add(new FakeSchedulable(0L, giga));
ComputeFairShares.computeShares(scheds,
Resources.createResource(40 * million),
Resources.createResource(4 * giga),
ResourceInformation.MEMORY_MB.getName());
verifyMemoryShares(10 * million, 10 * million, 10 * million, 10 * million);
verifyMemoryShares(giga, giga, giga, giga);
}
/**
* Test overflow in the resources taken and upper bound.
*/
@Test
public void testLargeMinimums() {
long giga = 1000L * 1000L * 1000L * 4L;
scheds.add(new FakeSchedulable(Long.MAX_VALUE, Long.MAX_VALUE));
scheds.add(new FakeSchedulable(giga, giga));
ComputeFairShares.computeShares(scheds,
Resources.createResource(4 * giga),
ResourceInformation.MEMORY_MB.getName());
verifyMemoryShares(Long.MAX_VALUE, giga);
}
/**
* Test overflow in the upper bound calculation for the binary search.
*/
@Test
public void testOverflowMaxShare() {
long giga = 1000L * 1000L * 1000L;
scheds.add(new FakeSchedulable(0L, giga));
scheds.add(new FakeSchedulable(0L, Long.MAX_VALUE));
ComputeFairShares.computeShares(scheds,
Resources.createResource(2 * giga),
ResourceInformation.MEMORY_MB.getName());
verifyMemoryShares(giga, giga);
}
/**
* Test overflow in the fixed share calculations. The 3th schedulable should
* not get any share as all resources are taken by the handleFixedShare()
* call.
* With the overflow it looked like there were more resources available then
* there really are.
* The values in the test might not be "real" but they show the overflow.
*/
@Test
public void testOverflowFixedShare() {
long giga = 1000L * 1000L * 1000L;
long minValue = Long.MAX_VALUE - 1L;
scheds.add(new FakeSchedulable(giga, giga, 0));
scheds.add(new FakeSchedulable(minValue, Long.MAX_VALUE, 0));
scheds.add(new FakeSchedulable(0L, giga));
ComputeFairShares.computeShares(scheds,
Resources.createResource(1000L),
ResourceInformation.MEMORY_MB.getName());
verifyMemoryShares(giga, minValue, 0);
}
/**
@ -176,7 +224,7 @@ public class TestComputeFairShares {
}
/**
* Test that CPU works as well as memory
* Test that CPU works as well as memory.
*/
@Test
public void testCPU() {
@ -192,10 +240,12 @@ public class TestComputeFairShares {
/**
* Check that a given list of shares have been assigned to this.scheds.
*/
private void verifyMemoryShares(int... shares) {
Assert.assertEquals(scheds.size(), shares.length);
private void verifyMemoryShares(long... shares) {
Assert.assertEquals("Number of shares and schedulables are not consistent",
scheds.size(), shares.length);
for (int i = 0; i < shares.length; i++) {
Assert.assertEquals(shares[i], scheds.get(i).getFairShare().getMemorySize());
Assert.assertEquals("Expected share number " + i + " in list wrong",
shares[i], scheds.get(i).getFairShare().getMemorySize());
}
}
@ -203,23 +253,11 @@ public class TestComputeFairShares {
* Check that a given list of shares have been assigned to this.scheds.
*/
private void verifyCPUShares(int... shares) {
Assert.assertEquals(scheds.size(), shares.length);
Assert.assertEquals("Number of shares and schedulables are not consistent",
scheds.size(), shares.length);
for (int i = 0; i < shares.length; i++) {
Assert.assertEquals(shares[i], scheds.get(i).getFairShare().getVirtualCores());
Assert.assertEquals("Expected share number " + i + " in list wrong",
shares[i], scheds.get(i).getFairShare().getVirtualCores());
}
}
/**
* Test computeShares will not enter into infinite loop.
*/
@Test(timeout = 10000)
public void testResourceUsedWithWeightToResourceRatio() {
Collection<Schedulable> schedulables = new ArrayList<>();
schedulables.add(new FakeSchedulable(Integer.MAX_VALUE));
schedulables.add(new FakeSchedulable(Integer.MAX_VALUE));
Resource totalResource = Resource.newInstance(Integer.MAX_VALUE, 0);
ComputeFairShares.computeShares(
schedulables, totalResource, ResourceInformation.MEMORY_URI);
}
}