[SPARK-23285][K8S] Allow fractional values for spark.executor.cores by liyinan926 · Pull Request #20460 · apache/spark

liyinan926 · 2018-01-31T21:34:29Z

What changes were proposed in this pull request?

K8s treats CPU as a "compressible resource" and can actually assign millicpus to individual containers, e.g., 0.1 (100 millicpus). In https://github.com/apache/spark/blob/master/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala#L94, we already parse spark.executor.cores as a double value. This PR enables use of fractional values for spark.executor.cores. It also allows fractional values for spark.task.cpus correspondingly. The various places where spark.executor.cores and spark.task.cpus are used are updated to do a conversion from double to integer values. The conversion is needed to not introduce a significant behavioral change to the way the other backends deal with executor cores. The maximum tasks per executor is now calculated as spark.executor.cores / spark.task.cpus rounded to an integer using the FLOOR mode.

How was this patch tested?

Manual tests.

@foxish @vanzin

K8s treats CPU as a "compressible resource" and can actually assign millicpus to individual containers, e.g., 0.1 or 100m. In https://github.com/apache/spark/blob/master/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala#L94, we already parse `spark.executor.cores` as a double value. This PR simply bypasses the check for integral values for the property in Kubernetes mode.

foxish · 2018-01-31T21:38:51Z

+ // the Kubernetes mode.
+ if (!master.startsWith("k8s")
+ && executorCores != null
+ && Try(executorCores.toInt).getOrElse(-1) <= 0) {


should this be nested if..() statements?
That's more readable IMO but not sure what's considered more idiomatic here.

If the intent here is just to check that it's positive, then just try to get it as a double and check it. This doesn't require making the logic dependent on a particular resource manager. Of course it relaxes the condition from what it is now, which would reject input like "1.5". But it would reject it not with a graceful exit but an exception, so I kind of think it's not what this code intends to catch anyway right now. (It would cause an error pretty quickly in YARN et al anyway)

Makes sense, done.

foxish

lgtm modulo one question

SparkQA · 2018-01-31T21:46:07Z

Kubernetes integration test starting
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/442/

SparkQA · 2018-01-31T21:52:22Z

Kubernetes integration test status success
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/442/

srowen · 2018-01-31T23:36:43Z

 In standalone and Mesos coarse-grained modes, for more detail, see
- <a href="spark-standalone.html#Executors Scheduling">this description</a>.
+ <a href="spark-standalone.html#Executors Scheduling">this description</a>. In Kubernetes mode,
+ a fractional value can be used, e.g., 0.1 or 100m.


100m isn't fractional though?

Oh, right. Fixed.

SparkQA · 2018-01-31T23:57:10Z

Test build #86894 has finished for PR 20460 at commit 712a5ff.

This patch fails to build.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-02-01T00:05:44Z

Kubernetes integration test starting
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/446/

SparkQA · 2018-02-01T00:11:43Z

Kubernetes integration test status success
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/446/

SparkQA · 2018-02-01T01:17:21Z

Test build #86890 has finished for PR 20460 at commit a971870.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-02-01T02:58:58Z

Test build #86895 has finished for PR 20460 at commit d9805c3.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

felixcheung · 2018-02-01T05:59:33Z

Jenkins, retest this please

SparkQA · 2018-02-01T06:10:58Z

Kubernetes integration test starting
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/462/

SparkQA · 2018-02-01T06:16:44Z

Kubernetes integration test status success
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/462/

SparkQA · 2018-02-01T09:13:30Z

Test build #86915 has finished for PR 20460 at commit d9805c3.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

srowen · 2018-02-01T13:50:23Z

 SparkSubmit.printErrorAndExit("Executor Memory cores must be a positive number")
 }
- if (executorCores != null && Try(executorCores.toInt).getOrElse(-1) <= 0) {
+ if (executorCores != null && Try(executorCores.toDouble).getOrElse(-1d) <= 0d) {


Although I think "0.0" is more readable than "0d", don't bother changing it here. Just an aside

SparkQA · 2018-02-01T17:40:02Z

Test build #4088 has finished for PR 20460 at commit d9805c3.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

jiangxb1987 · 2018-02-01T19:05:51Z

Do we also want to update the comment of SPARK_EXECUTOR_CORES in spark-env.sh ?

jiangxb1987 · 2018-02-01T19:16:29Z

Actually I think this may fail some check (though may not throw exceptions) for instance this one:

spark/core/src/main/scala/org/apache/spark/SparkConf.scala

Lines 553 to 562 in 9678941

if (contains("spark.cores.max") && contains("spark.executor.cores")) { 
valtotalCores= getInt("spark.cores.max", 1) 
valexecutorCores= getInt("spark.executor.cores", 1) 
valleftCores= totalCores % executorCores 
if (leftCores !=0) { 
 logWarning(s"Total executor cores: ${totalCores} is not "+
s"divisible by cores per executor: ${executorCores}, "+
s"the left cores: ${leftCores} will not be allocated") 
 } 
 } 

liyinan926 · 2018-02-01T22:27:55Z

@jiangxb1987 it seems spark.cores.max only applies to standalone and Mesos coarse-grained modes. In k8s mode, spark.cores.max is not applicable. I think the check really should not be here, but rather pushed to specific modes where this applies.

jiangxb1987 · 2018-02-01T22:31:35Z

Yea, we should either move the check to elsewhere or modify it, to avoid potential failure.

liyinan926 · 2018-02-01T22:45:00Z

@jiangxb1987 fixed the check in 44e489e.

SparkQA · 2018-02-01T22:55:48Z

Kubernetes integration test starting
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/488/

SparkQA · 2018-02-01T23:02:02Z

Kubernetes integration test status success
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/488/

jerryshao · 2018-02-02T02:16:52Z

I think here (

spark/core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala

Line 117 in 032c11b

privatevaltasksPerExecutor=

) should also be fixed.

Besides, I think "spark.task.cpus" should also be fixed, right? BTW, looks like dynamic allocation can not be well supported if cores are fractional.

SparkQA · 2018-02-02T02:20:46Z

Test build #86949 has finished for PR 20460 at commit 44e489e.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

jerryshao · 2018-02-02T02:20:54Z

I'd target this 2.3 & master. Waiting for tests

@felixcheung is it too risky to target to 2.3, this is a fundamental behavior change. We should make sure k8s could well use fractional cores, in the mean time we should also guard other cluster manager to use it. I think we need more tests on it.

jiangxb1987 · 2018-02-02T02:28:16Z

This may require more changes than it appears to be, agree we may need more tests to ensure it don't break anything, so +1 on target it to 2.4 instead of 2.3

liyinan926 · 2018-02-02T04:06:51Z

Agreed. It seems much more complicated than expected, considering the interaction between spark.executor.cores and spark.task.cpus and the implication for each scheduler backend. Will do a more thorough investigation.

felixcheung · 2018-02-02T06:51:13Z

ah, sounds like this is more impactful than we thought

liyinan926 · 2018-02-02T17:11:12Z

@jiangxb1987 @felixcheung @jerryshao @srowen this PR has been updated to allow both spark.executor.cores and spark.task.cpus to have fractional values. Please take a look at the updated PR description. Thanks!

SparkQA · 2018-02-02T17:15:51Z

Kubernetes integration test starting
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/527/

SparkQA · 2018-02-02T17:17:13Z

Test build #86992 has finished for PR 20460 at commit dc7483a.

This patch fails to build.
This patch merges cleanly.
This patch adds the following public classes (experimental):
case class WorkerOffer(executorId: String, host: String, cores: Double)

SparkQA · 2018-02-02T17:21:55Z

Kubernetes integration test status success
URL: https://amplab.cs.berkeley.edu/jenkins/job/testing-k8s-prb-spark-integration/527/

SparkQA · 2018-02-02T17:57:25Z

Test build #86993 has finished for PR 20460 at commit 2d61782.

This patch fails to build.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-02-02T18:03:10Z

Test build #86995 has finished for PR 20460 at commit 476221d.

This patch fails to build.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-02-02T18:52:06Z

Test build #86998 has finished for PR 20460 at commit b21e864.

This patch fails to build.
This patch merges cleanly.
This patch adds no public classes.

liyinan926 · 2018-02-02T19:07:42Z

Just realized that the update I did still won't work when dynamic resource allocation is enabled. So please ignore the update. This is definitely much more impactful than I thought.

jerryshao · 2018-02-03T03:50:52Z

I would suggest to bring out a discussion or even a design on dev mail list before doing such ground changing. This may affect not only dynamic allocation, but also scheduler. It is better to collect all the feedbacks (especially those who works on scheduler side).

liyinan926 · 2018-02-03T07:46:56Z

Agreed. This is a fundamental change to the way Spark handles task scheduling, task parallelism, and dynamic resource allocation, etc., and it impacts every scheduler backends. I'm closing this PR for now. Thanks for reviewing and giving feedbacks!

zenglian · 2018-09-07T15:20:42Z

I suggest to introduce a new property spark.kubernetes.coreUnit = cpu | mcpu (cpu by default), which would affect k8s only and the change should be minimized.

foxish reviewed Jan 31, 2018
View reviewed changes

srowen reviewed Jan 31, 2018
View reviewed changes

Addressed comments
d9805c3

srowen reviewed Feb 1, 2018
View reviewed changes

Fixed check in SparkConf
44e489e

jerryshao mentioned this pull request Feb 2, 2018
[LIVY-435] Allow floating point numbers to be used for the driver and executor core counts apache/livy#74
Closed

Also allow spark.task.cpus to have fractional values
b21e864

liyinan926 closed this Feb 3, 2018

foxish mentioned this pull request Feb 27, 2018
[SPARK-23285][K8S] Add a config property for specifying physical executor cores #20553
Closed

liyinan926 mentioned this pull request Apr 20, 2018
Sync with apache/spark/master kubeflow/spark-operator#129
Closed
8 tasks

Uh oh!

Conversation

liyinan926 commented Jan 31, 2018• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

foxishJan 31, 2018

Choose a reason for hiding this comment

Uh oh!

srowenJan 31, 2018

Choose a reason for hiding this comment

Uh oh!

liyinan926Jan 31, 2018

Choose a reason for hiding this comment

Uh oh!

foxish left a comment

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jan 31, 2018

Uh oh!

SparkQA commented Jan 31, 2018

Uh oh!

srowenJan 31, 2018

Choose a reason for hiding this comment

Uh oh!

liyinan926Jan 31, 2018

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jan 31, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

felixcheung commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

srowenFeb 1, 2018

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

jiangxb1987 commented Feb 1, 2018

Uh oh!

jiangxb1987 commented Feb 1, 2018• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liyinan926 commented Feb 1, 2018

Uh oh!

jiangxb1987 commented Feb 1, 2018

Uh oh!

liyinan926 commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

SparkQA commented Feb 1, 2018

Uh oh!

jerryshao commented Feb 2, 2018

Uh oh!

SparkQA commented Feb 2, 2018

Uh oh!

jerryshao commented Feb 2, 2018

Uh oh!

jiangxb1987 commented Feb 2, 2018

Uh oh!

liyinan926 commented Feb 2, 2018

Uh oh!

felixcheung commented Feb 2, 2018

Uh oh!

liyinan926 commented Feb 2, 2018• edited Loading Uh oh! There was an error while loading. Please reload this page.

liyinan926 commented Jan 31, 2018•
edited
Loading

jiangxb1987 commented Feb 1, 2018•
edited
Loading

liyinan926 commented Feb 2, 2018•
edited
Loading

zenglian commented Sep 7, 2018•
edited
Loading