[
https://issues.apache.org/jira/browse/HDDS-17862?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Andika updated HDDS-17862:
-------------------------------
Description:
This is just an idea for discussion.
Ozone currently makes the client to be in charge of writing the transmit the EC
stripes to all the datanodes similar to HDFS. However, other systems like Minio
/ RustFS and SeaweedFS implements the EC stripes writing to one of the datanode.
I can see multiple benefits
* Assume that a client machine is further to the datanodes than the datanodes
themselves. Making the striping done in the datanode makes the latency to be
improved.
* Assume that the datanode machine is a lot more powerful than client
(reference architecture does not prescribe client machine minimum requirement),
having datanode handles EC related logic would improve performance
**
[https://ozone.apache.org/docs/next/administrator-guide/installation/reference-architecture/]
(Ozone reference architecture)
* Clients do not need to open many socket to different datanodes which can
save resources.
We can consider the approach where the datanode is in charge of spreading the
stripes to other datanodes. Client can simply uploads the block to one datanode
(and with other peers locations) and that datanode will be in charge of
calculating and then spreading the blocks to multiple EC shards.
was:
This is just an idea for discussion.
Ozone currently makes the client to be in charge of writing the transmit the EC
stripes to all the datanodes similar to HDFS. However, other systems like Minio
/ RustFS and SeaweedFS implements the EC stripes writing to one of the datanode.
I can see multiple benefits
* Assume that a client machine is further to the datanodes than the datanodes
themselves. Making the striping done in the datanode makes the latency to be
improved.
* Assume that the datanode machine is a lot more powerful than client (from
the reference architecture it seems it does not prescribe client machine
minimum requirement), having datanode handles EC related logic would improve
performance
**
[https://ozone.apache.org/docs/next/administrator-guide/installation/reference-architecture/]
(Ozone reference architecture)
* Clients do not need to open many socket to different datanodes which can
save resources.
We can consider the approach where the datanode is in charge of spreading the
stripes to other datanodes. Client can simply uploads the block to one datanode
(and with other peers locations) and that datanode will be in charge of
calculating and then spreading the blocks to multiple EC shards.
> Support server-side EC fan-out
> ------------------------------
>
> Key: HDDS-17862
> URL: https://issues.apache.org/jira/browse/HDDS-17862
> Project: Apache Ozone
> Issue Type: Improvement
> Reporter: Ivan Andika
> Assignee: Ivan Andika
> Priority: Major
>
> This is just an idea for discussion.
> Ozone currently makes the client to be in charge of writing the transmit the
> EC stripes to all the datanodes similar to HDFS. However, other systems like
> Minio / RustFS and SeaweedFS implements the EC stripes writing to one of the
> datanode.
> I can see multiple benefits
> * Assume that a client machine is further to the datanodes than the
> datanodes themselves. Making the striping done in the datanode makes the
> latency to be improved.
> * Assume that the datanode machine is a lot more powerful than client
> (reference architecture does not prescribe client machine minimum
> requirement), having datanode handles EC related logic would improve
> performance
> **
> [https://ozone.apache.org/docs/next/administrator-guide/installation/reference-architecture/]
> (Ozone reference architecture)
> * Clients do not need to open many socket to different datanodes which can
> save resources.
> We can consider the approach where the datanode is in charge of spreading the
> stripes to other datanodes. Client can simply uploads the block to one
> datanode (and with other peers locations) and that datanode will be in charge
> of calculating and then spreading the blocks to multiple EC shards.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]