[
https://issues.apache.org/jira/browse/HDFS-17868?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
YUBI LEE updated HDFS-17868:
----------------------------
Description:
I was inspired by the ideas presented in https://dan.naver.com/25/sessions/692
and https://www.youtube.com/watch?v=1h4k_Dbt0t8 and implemented a new block
placement policy called BlockPlacementPolicyCrossDC.
Thanks to [~acedia28] for the original ideas. It would be great if [~acedia28]
could share an improved or more mature version of this block placement policy.
This implementation introduces the following configuration options
(default values are shown in parentheses):
{code}
dfs.block.replicator.cross.dc.async.enabled (false)
dfs.block.replicator.cross.dc.preferred.datacenter
dfs.block.replicator.cross.dc.bandwidth.limit.mb (5120)
dfs.block.replicator.cross.dc.bandwidth.refill.period.sec (1)
dfs.block.replicator.cross.dc.sync.paths
dfs.block.replicator.cross.dc.limited.sync.paths
{code}
Based on the ideas from the session mentioned above, this policy supports three
different HDFS block write modes:
1. Synchronous write
The standard HDFS behavior, where block replicas are synchronously written to
all target DataNodes.
2. Limited synchronous write
Uses bucket4j to control cross-datacenter traffic.
Writes below the configured threshold are performed synchronously, while writes
exceeding the threshold fall back to asynchronous replication.
3. Asynchronous write
The client is initially given only DataNode candidates located in the same
datacenter.
With asynchronous write, the hdfs client app will not suffer from write latency
caused by cross-datacenter communication.
Any under-replicas are created later through background HDFS replication to
other datacenters.
was:
I was inspired by the ideas presented in https://dan.naver.com/25/sessions/692
and https://www.youtube.com/watch?v=1h4k_Dbt0t8 and implemented a new block
placement policy called BlockPlacementPolicyCrossDC.
Thanks to [~acedia28] for the original ideas. It would be great if [~acedia28]
could share an improved or more mature version of this block placement policy.
This implementation introduces the following configuration options
(default values are shown in parentheses):
{code}
dfs.block.replicator.cross.dc.async.enabled (false)
dfs.block.replicator.cross.dc.preferred.datacenter
dfs.block.replicator.cross.dc.bandwidth.limit.mb (5120)
dfs.block.replicator.cross.dc.bandwidth.refill.period.sec (1)
dfs.block.replicator.cross.dc.sync.paths
dfs.block.replicator.cross.dc.limited.sync.paths
{code}
Based on the ideas from the session mentioned above, this policy supports three
different HDFS block write modes:
1. Synchronous write
The standard HDFS behavior, where block replicas are synchronously written to
all target DataNodes.
2. Limited synchronous write
Uses bucket4j to control cross-datacenter traffic.
Writes below the configured threshold are performed synchronously, while writes
exceeding the threshold fall back to asynchronous replication.
3. Asynchronous write
The client is initially given only DataNode candidates located in the same
datacenter.
Any under-replicas are created later through background HDFS replication to
other datacenters.
> introduce BlockPlacementPolicyCrossDC for multi datacenter stretched hdfs
> cluster
> ---------------------------------------------------------------------------------
>
> Key: HDFS-17868
> URL: https://issues.apache.org/jira/browse/HDFS-17868
> Project: Hadoop HDFS
> Issue Type: New Feature
> Components: block placement
> Reporter: YUBI LEE
> Priority: Major
> Labels: pull-request-available
>
> I was inspired by the ideas presented in
> https://dan.naver.com/25/sessions/692 and
> https://www.youtube.com/watch?v=1h4k_Dbt0t8 and implemented a new block
> placement policy called BlockPlacementPolicyCrossDC.
> Thanks to [~acedia28] for the original ideas. It would be great if
> [~acedia28] could share an improved or more mature version of this block
> placement policy.
> This implementation introduces the following configuration options
> (default values are shown in parentheses):
> {code}
> dfs.block.replicator.cross.dc.async.enabled (false)
> dfs.block.replicator.cross.dc.preferred.datacenter
> dfs.block.replicator.cross.dc.bandwidth.limit.mb (5120)
> dfs.block.replicator.cross.dc.bandwidth.refill.period.sec (1)
> dfs.block.replicator.cross.dc.sync.paths
> dfs.block.replicator.cross.dc.limited.sync.paths
> {code}
> Based on the ideas from the session mentioned above, this policy supports
> three different HDFS block write modes:
> 1. Synchronous write
> The standard HDFS behavior, where block replicas are synchronously written to
> all target DataNodes.
> 2. Limited synchronous write
> Uses bucket4j to control cross-datacenter traffic.
> Writes below the configured threshold are performed synchronously, while
> writes exceeding the threshold fall back to asynchronous replication.
> 3. Asynchronous write
> The client is initially given only DataNode candidates located in the same
> datacenter.
> With asynchronous write, the hdfs client app will not suffer from write
> latency caused by cross-datacenter communication.
> Any under-replicas are created later through background HDFS replication to
> other datacenters.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
