[
https://issues.apache.org/jira/browse/HADOOP-1134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12511270
]
Raghu Angadi commented on HADOOP-1134:
--------------------------------------
> I don't follow. If block sizes are always a multiple of bytesPerChecksum, as
> we agreed, then a chunk should never cross a block boundary, right?
DFSOutputStream needs to have its own write() in any case, since it needs to
decide how much can be written to current block and write the rest of the data
to next block. This is the case even if block size a multiple of
bytesPerChecksum. So DFSOutputStream needs to have a 'BlockWriter' class that
subclasses OutputSummer (similar to BlockReader in DFSInputStream).
Regd, enforcing block size to be multiple of bytesperchecksum, I did not know
if there was a consensus reached . I am more than happy to enforce it. I think
Owen thought it is not needed yet.
> Also, using a one-byte-buffer may still impact performance, since it must
> call multiple methods per call to write(int).
> It can make a significant performance difference if write(int) does much more
> than check a buffer boundary and
> buffer[count++] = byte.
It depends on what we are aiming to optimize. Yes, this write has couple of
extra calls (mainly checksum.update()).
If you think we have such a large percent of single byte writes, then this
might just be part of a bigger performance issue.
> Block level CRCs in HDFS
> ------------------------
>
> Key: HADOOP-1134
> URL: https://issues.apache.org/jira/browse/HADOOP-1134
> Project: Hadoop
> Issue Type: New Feature
> Components: dfs
> Reporter: Raghu Angadi
> Assignee: Raghu Angadi
> Attachments: BlockLevelCrc-07032007.patch,
> BlockLevelCrc-07052007.patch, BlockLevelCrc-07062007.patch,
> BlockLevelCrc-07062007.patch, BlockLevelCrc-07062007.patch,
> DfsBlockCrcDesign-05305007.htm, readBuffer.java
>
>
> Currently CRCs are handled at FileSystem level and are transparent to core
> HDFS. See recent improvement HADOOP-928 ( that can add checksums to a given
> filesystem ) regd more about it. Though this served us well there a few
> disadvantages :
> 1) This doubles namespace in HDFS ( or other filesystem implementations ). In
> many cases, it nearly doubles the number of blocks. Taking namenode out of
> CRCs would nearly double namespace performance both in terms of CPU and
> memory.
> 2) Since CRCs are transparent to HDFS, it can not actively detect corrupted
> blocks. With block level CRCs, Datanode can periodically verify the checksums
> and report corruptions to namnode such that name replicas can be created.
> We propose to have CRCs maintained for all HDFS data in much the same way as
> in GFS. I will update the jira with detailed requirements and design. This
> will include same guarantees provided by current implementation and will
> include a upgrade of current data.
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
