Comments (18)
I think it's necessary feature, what do u think? @jerqi
Besides, I have implement one policy of read directly
and want to contribute.
from incubator-uniffle.
I think it's the most important thing that we don't official deployment plan now.
from incubator-uniffle.
Yes. we need a detailed deployment instruction for users in doc. But I dont think this feature depends on the former.
from incubator-uniffle.
Yes. we need a detailed deployment instruction for users in doc. But I dont think this feature depends on the former.
Sorry, i means that we don't official yarn deployment plan now. I think this feature depends yarn.
from incubator-uniffle.
Oh, this feature is not bound to Yarn.
We will deploy shuffle-servers on Yarn nodemanager machine using ansible, instead of using YARN Applications to deploy.
from incubator-uniffle.
Motivation
When uniffle shuffle servers are co-located with Yarn nodemanagers, we could use the short-circuit read to improve performance and reduce the overhead.
How to do
There are two options to solve this
- Directly read shuffle-data files by client side.
- Use the domain socket.
Option 1 - Directly read
This is the fastest way, but the local-files' read permission should be open for client side. This maybe have security problem.
Option2 - Domain socket
This way could be avoid security problem, but will be slower than above. And its implementation will be complex.
Conclusion
In my opinion, above two ways could be all supported in uniffle, which could be as different policies for users to choose.
Which option do you prefer?
from incubator-uniffle.
Which option do you prefer?
I prefer directly read
in the initial version. What do u think? Maybe we could reserve the pluggable policy interface to implement for domain socket.
from incubator-uniffle.
Which option do you prefer?
I prefer
directly read
in the initial version.
How to know the local directory for Spark client?
from incubator-uniffle.
How to know the local directory for Spark client?
The stored local path should be retrieved from shuffle server by client side. So we need to add a new grpc api.
from incubator-uniffle.
When the taskScheduler allocate task, we don't consider reduce partition distribution, fewer task can use the feature.
from incubator-uniffle.
When the taskScheduler allocate task, we don't consider reduce partition distribution, fewer task can use the feature.
If having this feature, maybe we could update the Spark data-locality policy of preferLocation
for uniffle. Anyway, this is a improvement. Do u think so?
from incubator-uniffle.
When the taskScheduler allocate task, we don't consider reduce partition distribution, fewer task can use the feature.
If having this feature, maybe we could update the Spark data-locality policy of
preferLocation
for uniffle. Anyway, this is a improvement. Do u think so?
Although it's an improvement, if it may bring more complexity and less effectiveness, I prefer not merging it.
from incubator-uniffle.
Got your thought.
complexity
This feature only introduces the short-circuit read handler. In my POC, it wont break down the current implementation. Besides, we could introduce the config to control whether to enable this.
effectiveness
The effectiveness depends on the Spark knowledge about the shuffle-data-locality. I think this can be improved by the later patches for Spark, like Alluxio does.
from incubator-uniffle.
Do u have some ideas? @jerqi
from incubator-uniffle.
Do u have some ideas? @jerqi
This feature need to increase a new rpc interface. The interfaces are very important. I need to guarantee the compatibility of interfaces. I think this is the biggest complexity which this feature brings. For effectiveness, we don't have data locality, we can't get too much performance improvement. And we have no plans about data locality, if we want to have this feature, we should achieve the feature of data locality.
from incubator-uniffle.
For your thought, do I need to submit a data-locality shuffle patch about Spark firstly? As I know, we need to change the Spark codebase about MapOutputTracker.getPreferredLocationsForShuffle
and ShuffleRowRDD.getPreferredLocations
. When enable RSS, the perferred locations could be gotten from the dep.shuffleHandle(RssShuffleHandle of Uniffle partitionToServers vars). If I'm wrong, plz point out it. Thanks
I still think this is a necessary feature for co-location deployment. Besides this has been implemented by other RSS project.
from incubator-uniffle.
For your thought, do I need to submit a data-locality shuffle patch about Spark firstly? As I know, we need to change the Spark codebase about
MapOutputTracker.getPreferredLocationsForShuffle
andShuffleRowRDD.getPreferredLocations
. When enable RSS, the perferred locations could be gotten from the dep.shuffleHandle(RssShuffleHandle of Uniffle partitionToServers vars). If I'm wrong, plz point out it. ThanksI still think this is a necessary feature for co-location deployment. Besides this has been implemented by other RSS project.
Not only the spark patch, but also we need to modify the strategy of coordinator. I know the Bytedance RSS implement this feature, if the cluster is big enough, the feature will be useless.
from incubator-uniffle.
but also we need to modify the strategy of coordinator
Yes, we’d better to assign shuffle servers with data locality. This is a long way to get the best optimization.
from incubator-uniffle.
Related Issues (20)
- [FEATURE] Support rpc audit log for coordinator
- [Bug] java.lang.NoSuchMethodError: java.nio.ByteBuffer.position(I)Ljava/nio/ByteBuffer
- [FEATURE] Client send sparkConf through registerApplicationInfo rpc to Coordinator HOT 2
- [Bug] The server page of dashboard cannot display server info when refresh the page HOT 2
- [Bug] [TEZ] Application may get stuck when exception occurs.
- [FEATURE] Add mock data to the Dashboard front end.
- [Improvement] introduce Lombok
- [FEATURE] Introduce swagger to support restapi HOT 2
- [Improvement] Add Netty support for TEZ tasks in Uniffle
- [Subtask] Report HTTP port of Shuffle Server to the coordinator HOT 1
- [Bug] [dashboard] Add exclude node not work HOT 1
- Unify `ShuffleManageClient` config `rss.client.rpc.timeout.ms` and `rss.rpc.client.type.grpc.timeout`
- [Flaky Test] org.apache.uniffle.test.GetReaderTest
- [Improvement] Support Netty for MR integration test
- [Improvement] Support config to exclude verbose operation from rpc audit log HOT 1
- [DOCS] Add Troubleshooting log documents
- [Bug] org.apache.uniffle.common.exception.RssException: There isn't enough shuffle servers
- [Subtask] move the rest-client from cli module to common module
- [Improvement] Fix typo issue as a batch
- [Bug] spark on yarn throws NoClassDefFoundError HOT 4
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from incubator-uniffle.