[SPARK-4480] Avoid many small spills in external data structures #3353

andrewor14 · 2014-11-19T02:08:02Z

Summary. Currently, we may spill many small files in ExternalAppendOnlyMap and ExternalSorter. The underlying root cause of this is summarized in SPARK-4452. This PR does not address this root cause, but simply provides the guarantee that we never spill the in-memory data structure if its size is less than a configurable threshold of 5MB. This config is not documented because we don't want users to set it themselves, and it is not hard-coded because we need to change it in tests.

Symptom. Each spill is orders of magnitude smaller than 1MB, and there are many spills. In environments where the ulimit is set, this frequently causes "too many open file" exceptions observed in SPARK-3633.

14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4792 B to disk (292769 spills so far)
14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4760 B to disk (292770 spills so far)
14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4520 B to disk (292771 spills so far)
14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4560 B to disk (292772 spills so far)
14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4792 B to disk (292773 spills so far)
14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4784 B to disk (292774 spills so far)

Reproduction. I ran the following on a small 4-node cluster with 512MB executors. Note that the back-to-back shuffle here is necessary for reasons described in SPARK-4522. The second shuffle is a reduceByKey because it performs a map-side combine.

sc.parallelize(1 to 100000000, 100)
  .map { i => (i, i) }
  .groupByKey()
  .reduceByKey(_ ++ _)
  .count()

Before the change, I notice that each thread may spill up to 1000 times, and the size of each spill is on the order of 10KB. After the change, each thread spills only up to 20 times in the worst case, and the size of each spill is on the order of 1MB.

andrewor14 · 2014-11-19T02:08:44Z

@mateiz

SparkQA · 2014-11-19T02:14:55Z

Test build #23577 has started for PR 3353 at commit a919776.

This patch merges cleanly.

SparkQA · 2014-11-19T03:20:04Z

Test build #23577 has finished for PR 3353 at commit a919776.

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

AmplabJenkins · 2014-11-19T03:20:08Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23577/
Test FAILed.

SparkQA · 2014-11-19T04:15:16Z

Test build #23591 has started for PR 3353 at commit 23f2a2e.

This patch merges cleanly.

Unfortunately we have to expose the config here to keep the test duration reasonably low.

andrewor14 · 2014-11-19T04:38:04Z

retest this please

SparkQA · 2014-11-19T04:40:16Z

Test build #23594 has started for PR 3353 at commit f4736e3.

This patch merges cleanly.

SparkQA · 2014-11-19T04:42:30Z

Test build #23595 has started for PR 3353 at commit f4736e3.

This patch merges cleanly.

SparkQA · 2014-11-19T05:16:23Z

Test build #23591 has finished for PR 3353 at commit 23f2a2e.

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

AmplabJenkins · 2014-11-19T05:16:27Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23591/
Test FAILed.

SparkQA · 2014-11-19T06:03:07Z

Test build #23594 has finished for PR 3353 at commit f4736e3.

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

AmplabJenkins · 2014-11-19T06:03:12Z

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23594/
Test PASSed.

SparkQA · 2014-11-19T06:05:28Z

Test build #23595 has finished for PR 3353 at commit f4736e3.

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

AmplabJenkins · 2014-11-19T06:05:31Z

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23595/
Test PASSed.

mateiz · 2014-11-19T07:25:59Z

LGTM. Feel free to merge it.

Conflicts: core/src/main/scala/org/apache/spark/util/collection/Spillable.scala

SparkQA · 2014-11-19T18:27:57Z

Test build #23614 has started for PR 3353 at commit 27d6966.

This patch merges cleanly.

This is the branch-1.1 version of #3353. This requires a separate PR because the code in master has been refactored a little to eliminate duplicate code. I have tested this on a standalone cluster. The goal is to merge this into 1.1.1. Author: Andrew Or <andrew@databricks.com> Closes #3354 from andrewor14/avoid-small-spills-1.1 and squashes the following commits: f2e552c [Andrew Or] Fix tests 7012595 [Andrew Or] Avoid many small spills

arahuja · 2014-11-19T19:44:44Z

Was this not going into master?

andrewor14 · 2014-11-19T19:45:03Z

Whoops I accidentally closed this without merging into master. I'll re-open it.

SparkQA · 2014-11-19T19:51:37Z

Test build #23614 has finished for PR 3353 at commit 27d6966.

This patch fails MiMa tests.
This patch merges cleanly.
This patch adds no public classes.

AmplabJenkins · 2014-11-19T19:51:41Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23614/
Test FAILed.

SparkQA · 2014-11-19T19:52:43Z

Test build #23621 has started for PR 3353 at commit 27d6966.

This patch merges cleanly.

andrewor14 · 2014-11-19T19:53:25Z

Argh, tests won't pass because MIMA checks are broken in master. I'll send a hot fix.

SparkQA · 2014-11-19T21:18:43Z

Test build #23621 has finished for PR 3353 at commit 27d6966.

This patch fails MiMa tests.
This patch merges cleanly.
This patch adds no public classes.

AmplabJenkins · 2014-11-19T21:18:46Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23621/
Test FAILed.

…into avoid-small-spills

SparkQA · 2014-11-19T22:20:12Z

Test build #23633 has started for PR 3353 at commit 49f380f.

This patch merges cleanly.

SparkQA · 2014-11-20T00:20:13Z

Test build #23633 timed out for PR 3353 at commit 49f380f after a configured wait of 120m.

AmplabJenkins · 2014-11-20T00:20:16Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23633/
Test FAILed.

andrewor14 · 2014-11-20T00:37:14Z

retest this please

SparkQA · 2014-11-20T00:42:49Z

Test build #23645 has started for PR 3353 at commit 49f380f.

This patch merges cleanly.

SparkQA · 2014-11-20T02:05:49Z

Test build #23645 has finished for PR 3353 at commit 49f380f.

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

AmplabJenkins · 2014-11-20T02:05:53Z

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/23645/
Test PASSed.

andrewor14 · 2014-11-20T02:07:06Z

Finally. I'm merging this into master and 1.2.

**Summary.** Currently, we may spill many small files in `ExternalAppendOnlyMap` and `ExternalSorter`. The underlying root cause of this is summarized in [SPARK-4452](https://issues.apache.org/jira/browse/SPARK-4452). This PR does not address this root cause, but simply provides the guarantee that we never spill the in-memory data structure if its size is less than a configurable threshold of 5MB. This config is not documented because we don't want users to set it themselves, and it is not hard-coded because we need to change it in tests. **Symptom.** Each spill is orders of magnitude smaller than 1MB, and there are many spills. In environments where the ulimit is set, this frequently causes "too many open file" exceptions observed in [SPARK-3633](https://issues.apache.org/jira/browse/SPARK-3633). ``` 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4792 B to disk (292769 spills so far) 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4760 B to disk (292770 spills so far) 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4520 B to disk (292771 spills so far) 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4560 B to disk (292772 spills so far) 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4792 B to disk (292773 spills so far) 14/11/13 19:20:43 INFO collection.ExternalSorter: Thread 60 spilling in-memory batch of 4784 B to disk (292774 spills so far) ``` **Reproduction.** I ran the following on a small 4-node cluster with 512MB executors. Note that the back-to-back shuffle here is necessary for reasons described in [SPARK-4522](https://issues.apache.org/jira/browse/SPARK-4452). The second shuffle is a `reduceByKey` because it performs a map-side combine. ``` sc.parallelize(1 to 100000000, 100) .map { i => (i, i) } .groupByKey() .reduceByKey(_ ++ _) .count() ``` Before the change, I notice that each thread may spill up to 1000 times, and the size of each spill is on the order of 10KB. After the change, each thread spills only up to 20 times in the worst case, and the size of each spill is on the order of 1MB. Author: Andrew Or <andrew@databricks.com> Closes #3353 from andrewor14/avoid-small-spills and squashes the following commits: 49f380f [Andrew Or] Merge branch 'master' of https://git-wip-us.apache.org/repos/asf/spark into avoid-small-spills 27d6966 [Andrew Or] Merge branch 'master' of github.com:apache/spark into avoid-small-spills f4736e3 [Andrew Or] Fix tests a919776 [Andrew Or] Avoid many small spills (cherry picked from commit 0eb4a7f) Signed-off-by: Andrew Or <andrew@databricks.com>

This is blocking apache#3353 and other patches. Author: Andrew Or <andrew@databricks.com> Closes apache#3371 from andrewor14/mima-hot-fix and squashes the following commits: 842d059 [Andrew Or] Move excludes to the right section c4d4f4e [Andrew Or] MIMA hot fix

Avoid many small spills

a919776

andrewor14 mentioned this pull request Nov 19, 2014

[SPARK-4480] Avoid many small spills in external data structures (1.1) #3354

Closed

Fix tests

f4736e3

Unfortunately we have to expose the config here to keep the test duration reasonably low.

andrewor14 force-pushed the avoid-small-spills branch from 23f2a2e to f4736e3 Compare November 19, 2014 04:34

Merge branch 'master' of github.com:apache/spark into avoid-small-spills

27d6966

Conflicts: core/src/main/scala/org/apache/spark/util/collection/Spillable.scala

andrewor14 closed this Nov 19, 2014

andrewor14 deleted the avoid-small-spills branch November 19, 2014 19:27

andrewor14 restored the avoid-small-spills branch November 19, 2014 19:44

andrewor14 mentioned this pull request Nov 19, 2014

Please disregard this PR in favor of #3353 #3370

Closed

andrewor14 reopened this Nov 19, 2014

andrewor14 mentioned this pull request Nov 19, 2014

[HOT FIX] MiMa tests are broken #3371

Closed

Merge branch 'master' of https://git-wip-us.apache.org/repos/asf/spark …

49f380f

…into avoid-small-spills

asfgit closed this in 0eb4a7f Nov 20, 2014

andrewor14 deleted the avoid-small-spills branch November 20, 2014 02:32

peter-toth mentioned this pull request Jun 21, 2020

[SPARK-29375][SPARK-28940][SPARK-32041][SQL] Whole plan exchange and subquery reuse #28885

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SPARK-4480] Avoid many small spills in external data structures #3353

[SPARK-4480] Avoid many small spills in external data structures #3353

andrewor14 commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

mateiz commented Nov 19, 2014

SparkQA commented Nov 19, 2014

arahuja commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 20, 2014

AmplabJenkins commented Nov 20, 2014

andrewor14 commented Nov 20, 2014

SparkQA commented Nov 20, 2014

SparkQA commented Nov 20, 2014

AmplabJenkins commented Nov 20, 2014

andrewor14 commented Nov 20, 2014

[SPARK-4480] Avoid many small spills in external data structures #3353

[SPARK-4480] Avoid many small spills in external data structures #3353

Conversation

andrewor14 commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

mateiz commented Nov 19, 2014

SparkQA commented Nov 19, 2014

arahuja commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

andrewor14 commented Nov 19, 2014

SparkQA commented Nov 19, 2014

AmplabJenkins commented Nov 19, 2014

SparkQA commented Nov 19, 2014

SparkQA commented Nov 20, 2014

AmplabJenkins commented Nov 20, 2014

andrewor14 commented Nov 20, 2014

SparkQA commented Nov 20, 2014

SparkQA commented Nov 20, 2014

AmplabJenkins commented Nov 20, 2014

andrewor14 commented Nov 20, 2014