Draft: Performance mega boost - queue per app #1990
Draft
+471
−108
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
* Add MemoryLimit as option that will override spark pods limits (by using webhook) * queue per spark app - improved performance by far * added logs * prevent concurrent access to the appsQueues map * use RWmutex when accessing the appQueues --------- Co-authored-by: Netanel Levine <netanel.levine@mobileye.com> Co-authored-by: Eran Ben Ami <eranba@mobileye.com>
# Conflicts: # pkg/controller/sparkapplication/controller.go
Merge kubeflow
Updated
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: The full list of commands accepted by this bot can be found here.
Needs approval from an approver in each of these files:
Approvers can indicate their approval by writing |
# Conflicts: # Dockerfile
Fix tests
|
/assign @ChenYi015 Please review the changes when you get a chance |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Issues related
#1138
#1574
#783
#1489
Purpose of this PR
First of all - this PR is mainly a Draft which we think should be discussed, and that's why we are submitting it even though we haven't added documentation not unit tests.
The original design of spark-operator posses one queue which is used for all the applications.
This design is causing a huge latency issue when trying to deal with hundreds / thousands of applications concurrently.
In benchmarks me and my fellows from Mobileye preformed, we showed clearly a linear latency increase depend on the amount of apps handled.
When getting to more than 500 applications, the avg time from creating a spark application object until pod creation is ~130 seconds for app to start. When going up to 1200 apps concurrently it can go Up to 20 minutes on average for each spark application to be created.
Scale up vertically would not be helpful as the CPU is doing nothing, most of the time is spent on the queue mutex.
The change we are presenting here is to create a queue for each app.
It required a big change all around the code, but it is not changing the main flow in any way.
Our benchmarks showed that even with 1000 apps concurrently, Avg time for application creation is ~7 seconds
We also added a nice feature of using memoryLimit for driver / executor which is larger than the request by using the admission webhook.
Proposed changes:
Change Category
What are we still missing:
I would point out that this code currently runs on our production environment with massive scale without any issues.