[
https://issues.apache.org/jira/browse/HDFS-17061?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Shilun Fan updated HDFS-17061:
------------------------------
Target Version/s: 3.4.1 (was: 3.4.0)
> EC: Let data blocks and parity blocks on DNs more balanced
> ----------------------------------------------------------
>
> Key: HDFS-17061
> URL: https://issues.apache.org/jira/browse/HDFS-17061
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: balancer & mover, erasure-coding, hdfs
> Reporter: WangYuanben
> Priority: Minor
> Attachments: figure1, unbalanced traffic load on DNs.png, figure2,
> balanced traffic load on DNs.png
>
>
> When choosing DN for placing data block or parity block, the existing number
> of data block and parity block on datanode is not taken into consideration.
> This may lead to *uneven traffic load*.
> As shown in the figure 1, when reading block group A, B, C, D and E from five
> different EC files without any missing block, datanodes like DN1 and DN2 will
> have high traffic load. However, datanodes like DN3, DN4 and DN5 may have low
> or even no traffic load.
> !figure1, unbalanced traffic load on DNs.png|width=600,height=333!
> +If we can let data blocks and parity blocks on DNs more balanced, the
> traffic load in cluster will be more balanced and the peak traffic load on DN
> will be reduced+. Here "balance" refers to the matching of the number of data
> blocks and parity blocks on DN with its EC policy. In the ideal state, each
> DN has a balanced traffic load just like what figure 2 shows.
> !figure2, balanced traffic load on DNs.png|width=600,height=333!
> Then how to reduce this imbalance? I think it's related to EC policy and the
> ratio of data blocks to parity blocks on datanode. For RS-3-2-1024k, it's
> appropriate to let the ratio close to 3:2.
> There are two solutions:
> 1.Improve the block placement policy.
> 2.Improve the Balancer.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
