[
https://issues.apache.org/jira/browse/HDFS-1070?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13014348#comment-13014348
]
Matt Foley commented on HDFS-1070:
----------------------------------
I was talking with Sanjay and he would like us to consider making the
serialized tree format more robust. With the old "full names" format, an
arbitrary chunk of the file could be lost or corrupted, and the remainder would
still be recoverable. With this tree format, if a chunk is lost or corrupted,
everything thereafter is difficult or impossible to reconstruct because the
tree information is all implicit in the ordering.
Things that would help are
* making more of the tree-structure information explicit
* providing re-synchronization points or labels in the format
The simplest idea that occurred to me (and I haven't had time to discuss it
with Sanjay yet) is to go back to a breadth-first ordering so that the members
of each directory are grouped together, and output the full-path name of the
directory at the beginning of each such group. The full-path string provides
both the structure info and the re-synch capability, just as the full-names do
now. Maybe we can avoid actually parsing that string in most cases. This
would be going back to something like the old saveImage(), but only output the
local name with each inode, and output the full-path name once when beginning
each new breadth-first directory.
This will decrease the benefit of the change, but if we believe that number of
files is typically much larger than number of directories, there could still be
a big benefit. What do you think? Would it be hard to re-run your experiment
with this mod to see the cost of this change?
> Speedup NameNode image loading and saving by storing local file names
> ---------------------------------------------------------------------
>
> Key: HDFS-1070
> URL: https://issues.apache.org/jira/browse/HDFS-1070
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: name-node
> Reporter: Hairong Kuang
> Assignee: Hairong Kuang
> Attachments: trunkLocalNameImage.patch, trunkLocalNameImage1.patch,
> trunkLocalNameImage3.patch, trunkLocalNameImage4.patch,
> trunkLocalNameImage5.patch
>
>
> Currently each inode stores its full path in the fsimage. I'd propose to
> store the local name instead. In order for each inode to identify its parent,
> all inodes in a directory tree are stored in the image in in-order. This
> proposal also requires each directory stores the number of its children in
> image.
> This proposal would bring a few benefits as pointed below and therefore
> speedup the image loading and saving.
> # Remove the overhead of converting java-UTF8 encoded local name to
> string-represented full path then to UTF8 encoded full path when saving to an
> image and vice versa when loading the image.
> # Remove the overhead of traversing the full path when inserting the inode to
> its parent inode.
> # Reduce the number of temporary java objects during the process of image
> loading or saving and therefore reduce the GC overhead.
> # Reduce the size of an image.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
