qfoxzjd opened a new pull request #32900:
URL: https://github.com/apache/spark/pull/32900
<!--
Thanks for sending a pull request! Here are some tips for you:
1. If this is your first time, please read our contributor guidelines:
https://spark.apache.org/contributing.html
2. Ensure you have added or run the appropriate tests for your PR:
https://spark.apache.org/developer-tools.html
3. If the PR is unfinished, add '[WIP]' in your PR title, e.g.,
'[WIP][SPARK-XXXX] Your PR title ...'.
4. Be sure to keep the PR description updated to reflect all changes.
5. Please write your PR title to summarize what this PR proposes.
6. If possible, provide a concise example to reproduce the issue for a
faster review.
7. If you want to add a new configuration, please read the guideline first
for naming configurations in
'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'.
8. If you want to add or modify an error message, please read the
guideline first:
https://spark.apache.org/error-message-guidelines.html
-->
### What changes were proposed in this pull request?
<!--
Please clarify what changes you are proposing. The purpose of this section
is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR. See the examples below.
1. If you refactor some codes with changing classes, showing the class
hierarchy will help reviewers.
2. If you fix some SQL features, you can provide some references of other
DBMSes.
3. If there is design documentation, please add the link.
4. If there is a discussion in the mailing list, please add the link.
-->
### Why are the changes needed?
<!--
Please clarify why the changes are needed. For instance,
1. If you propose a new API, clarify the use case for a new API.
2. If you fix a bug, you can clarify why it is a bug.
-->
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such as
the documentation fix.
If yes, please clarify the previous behavior and the change this PR proposes
- provide the console output, description and/or an example to show the
behavior difference if possible.
If possible, please also clarify if this is a user-facing change compared to
the released Spark versions or within the unreleased branches such as master.
If no, write 'No'.
-->
No
### How was this patch tested?
<!--
If tests were added, say they were added here. Please make sure to add some
test cases that check the changes thoroughly including negative and positive
cases if possible.
If it was tested in a way different from regular unit tests, please clarify
how you tested step by step, ideally copy and paste-able, so that other
reviewers can test and check, and descendants can verify in the future.
If tests were not added, please describe why they were not added and/or why
it was difficult to add.
-->
Currently, there is only eviction strategy for cached RDD partition in Spark.
The default RDD eviction strategy is LRU (with an additional rule that do
not replacing another block that belongs to the same RDD like current creating
partition).
When memory space not sufficient for RDD caching, several partitions will be
evicted, if these partitions are used again latterly, they will be reproduced
by the Lineage information and cached in memory again. The reproduce phase will
bring in additional cost. However, LRU has no guarantee for the lowest
reproduce cost.
The first RDD that needed to be cached is usually generated by reading from
HDFS and doing several transformations. The reading operation usually cost
longer time than other Spark transformations.
For example, in one stage we having the following DAG structure: hdfs -> [A]
-> B -> [C] -> D - > [E] -> [F], RDD A, C, E, F needed to be cached in memory,
F is creating during this stage while A, B and E had already been created in
previous. When using the LRU eviction strategy, partition of A will be evicted
first. However, the time cost in\ [A] -> B -> [C] may be much less than hdfs
->\ [A], so evict [C] may be better than evict [A].
A eviction strategy based on the creation cost may be better than LRU, by
statisticing each transformation's time during the creation of cached RDD
partition (e.g. [E] only need to statistic time cost in [C] -> D and D -> [E])
and time cost in needed shuffle reading. When memory for RDD storage not
sufficient, partition with the least creation cost may be evicted first. So
this strategy for be called as LCS. My current demo show better performance
gain than default LRU.
This strategy needs to consider the following situation:
1. Unified Memory Management is provided after Spark 1.6, memory for
execution during recomputing a partition may be pretty different than the first
time the partition created. So before better thought, LCS may not be allowed in
UMM mode. (Though my demo also show improvement in LCS than LRU in UMM mode).
2. MEMORY_AND_DISK_SER or other similar storage level may serialize RDD
partition. By estimating ser/deserialize cost and compare to creation cost, if
the ser/deserialize cost even larger than recreation, not serialize but
directly removed from memory. As existing storage level only allowed for the
whole RDD, so a new storage level may be needed for RDD partition to directly
determine whether to serialize or just remove from memory.
Besides LCS, FIFO or LFU is easy to be implemented.
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
