[
https://issues.apache.org/jira/browse/HDFS-15610?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Karthik Palanisamy updated HDFS-15610:
--------------------------------------
Description:
There is a kernel overhead on datanode upgrade. If datanode with millions of
blocks and 10+ disks then block-layout migration will be super expensive during
its hardlink operation. Slowness is observed when running with large hardlink
threads(dfs.datanode.block.id.layout.upgrade.threads, default is 12 thread for
each disk) and its runs for 2+ hours.
I.e 10*12=120 threads (for 10 disks)
Small test:
RHEL7, 32 cores, 20 GB RAM, 8 GB DN heap
||dfs.datanode.block.id.layout.upgrade.threads||Blocks||Disks||Time taken||
|12|3.3 Million|1|2 minutes and 59 seconds|
|6|3.3 Million|1|2 minutes and 35 seconds|
|3|3.3 Million|1|2 minutes and 51 seconds|
Tried same test twice and 95% is accurate (only a few sec difference on each
iteration). Using 6 thread is faster than 12 thread because of its overhead.
was:
There is a kernel overhead on datanode upgrade. If datanode with millions of
blocks and 10+ disks then block-layout migration will be super expensive during
its hardlink operation. Slowness is observed when running with large hardlink
threads(dfs.datanode.block.id.layout.upgrade.threads, default is 12 thread for
each disk) and its runs for 2+ hours.
I.e 10*12=120 threads (for 10 disks)
Small test.
||dfs.datanode.block.id.layout.upgrade.threads||Blocks||Disks||Time taken||
|12|3.3 Million|1|2 minutes and 59 seconds|
|6|3.3 Million|1|2 minutes and 35 seconds|
|3|3.3 Million|1|2 minutes and 51 seconds|
Tried same test twice and 95% is accurate (only a few sec difference on each
iteration). Using 6 thread is faster than 12 thread because of its overhead.
> Reduce datanode upgrade/hardlink thread
> ---------------------------------------
>
> Key: HDFS-15610
> URL: https://issues.apache.org/jira/browse/HDFS-15610
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode
> Affects Versions: 3.0.0, 3.1.4
> Reporter: Karthik Palanisamy
> Assignee: Karthik Palanisamy
> Priority: Major
>
> There is a kernel overhead on datanode upgrade. If datanode with millions of
> blocks and 10+ disks then block-layout migration will be super expensive
> during its hardlink operation. Slowness is observed when running with large
> hardlink threads(dfs.datanode.block.id.layout.upgrade.threads, default is 12
> thread for each disk) and its runs for 2+ hours.
> I.e 10*12=120 threads (for 10 disks)
> Small test:
> RHEL7, 32 cores, 20 GB RAM, 8 GB DN heap
> ||dfs.datanode.block.id.layout.upgrade.threads||Blocks||Disks||Time taken||
> |12|3.3 Million|1|2 minutes and 59 seconds|
> |6|3.3 Million|1|2 minutes and 35 seconds|
> |3|3.3 Million|1|2 minutes and 51 seconds|
> Tried same test twice and 95% is accurate (only a few sec difference on each
> iteration). Using 6 thread is faster than 12 thread because of its overhead.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
