[
https://issues.apache.org/jira/browse/HDFS-395?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13058766#comment-13058766
]
Tomasz Nykiel commented on HDFS-395:
------------------------------------
To revive this discussion, I would have few comments about block reports.
Processing a single block report has mainly two parts. The first part is to
compute the difference between NN state and the report.
The second stage involves applying the diff to NN - which contains blocksToAdd,
blocksToRemove, blocksToInvalidat.
1. Now we have explicit Received requests sent form datanodes to the namenode,
hence the number of blocksToAdd should be very small.
2. blocksToInvalidate consist of blocks that do not belong to any file, which
is a abnormal situations.
Hence, the number thereof should also be very small.
3. We do not have explicit ACKs for blocks that are deleted from datanodes.
Hence, the diff will contain all such deleted blocks.
If we extend the block report interval to let's say 24 hours, we migh have a
huge number of blocks that have been deleted, and need to be processed.
I think it will be very beneficial to introduce explicit deletion acks.
Since we care more about the blockReceived in general, than blockDeleted, we
can send the deletion acks whenever we have any blockReceived acks to sent at
the datanode side.
Otherwise, we send the block deletion acks in some interval.
With this change, the block report interval can be extended to let's say two
times the basic interval.
-- Also a small change need to be made for sending blockDeleted ack.
When the block invalidate command comes to the datanode, it should
synchronously rename the block file (to some invalid block name).
Then it can immediately notify the namenode that the block has been deleted.
-- Another improvement can be made on NN side: We can process the entire list
of blocks within the FSNamesystem instead of processing them one-by-one.
This saves some repetitive computation.
I am attaching a diff that introduces these changes.
> DFS Scalability: Incremental block reports
> ------------------------------------------
>
> Key: HDFS-395
> URL: https://issues.apache.org/jira/browse/HDFS-395
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: dhruba borthakur
> Assignee: dhruba borthakur
> Attachments: blockReportPeriod.patch
>
>
> I have a cluster that has 1800 datanodes. Each datanode has around 50000
> blocks and sends a block report to the namenode once every hour. This means
> that the namenode processes a block report once every 2 seconds. Each block
> report contains all blocks that the datanode currently hosts. This makes the
> namenode compare a huge number of blocks that practically remains the same
> between two consecutive reports. This wastes CPU on the namenode.
> The problem becomes worse when the number of datanodes increases.
> One proposal is to make succeeding block reports (after a successful send of
> a full block report) be incremental. This will make the namenode process only
> those blocks that were added/deleted in the last period.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
