[
https://issues.apache.org/jira/browse/HDFS-14617?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16883532#comment-16883532
]
Daryn Sharp commented on HDFS-14617:
------------------------------------
I'm definitely not opposed to startup gains because 1-2h startup is rough.
I've been vocal about PB performance issues but I'm skeptical of the following:
{quote}The time taken to read and parse the protobuf messages seems to dominate
the runtime
{quote}
Have you profiled to determine if PB parsing is the dominator? Perhaps the cpu
cycle dominator but I'd expect I/O latency to be the wall clock dominator.
But, performance problems are rarely intuitive.
{quote}but we will need to somehow read and decode the protobuf in parallel to
get significant speedup
{quote}
The biggest problem is usually the GC overhead from all the objects vomited by
PB marshaling. Parallelism is likely to exasperate the GC overhead but you
appear to be seeing healthy performance gains. It's a question/experiment of
can we do better?
Perhaps the performance gain is simply due to the pipelining that decouples the
I/O latency from the computation to update the fsdir related structures. I'd
be very interested in the relative performance to a single thread allowed to
read as fast as the os/disk/page cache allows (while vomiting voluminous
PB-related objects) while another updates the fsdir unimpeded by
synchronization. It might be similar or higher.
> Improve fsimage load time by writing sub-sections to the fsimage index
> ----------------------------------------------------------------------
>
> Key: HDFS-14617
> URL: https://issues.apache.org/jira/browse/HDFS-14617
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: namenode
> Reporter: Stephen O'Donnell
> Assignee: Stephen O'Donnell
> Priority: Major
> Attachments: HDFS-14617.001.patch
>
>
> Loading an fsimage is basically a single threaded process. The current
> fsimage is written out in sections, eg iNode, iNode_Directory, Snapshots,
> Snapshot_Diff etc. Then at the end of the file, an index is written that
> contains the offset and length of each section. The image loader code uses
> this index to initialize an input stream to read and process each section. It
> is important that one section is fully loaded before another is started, as
> the next section depends on the results of the previous one.
> What I would like to propose is the following:
> 1. When writing the image, we can optionally output sub_sections to the
> index. That way, a given section would effectively be split into several
> sections, eg:
> {code:java}
> inode_section offset 10 length 1000
> inode_sub_section offset 10 length 500
> inode_sub_section offset 510 length 500
>
> inode_dir_section offset 1010 length 1000
> inode_dir_sub_section offset 1010 length 500
> inode_dir_sub_section offset 1010 length 500
> {code}
> Here you can see we still have the original section index, but then we also
> have sub-section entries that cover the entire section. Then a processor can
> either read the full section in serial, or read each sub-section in parallel.
> 2. In the Image Writer code, we should set a target number of sub-sections,
> and then based on the total inodes in memory, it will create that many
> sub-sections per major image section. I think the only sections worth doing
> this for are inode, inode_reference, inode_dir and snapshot_diff. All others
> tend to be fairly small in practice.
> 3. If there are under some threshold of inodes (eg 10M) then don't bother
> with the sub-sections as a serial load only takes a few seconds at that scale.
> 4. The image loading code can then have a switch to enable 'parallel loading'
> and a 'number of threads' where it uses the sub-sections, or if not enabled
> falls back to the existing logic to read the entire section in serial.
> Working with a large image of 316M inodes and 35GB on disk, I have a proof of
> concept of this change working, allowing just inode and inode_dir to be
> loaded in parallel, but I believe inode_reference and snapshot_diff can be
> make parallel with the same technique.
> Some benchmarks I have are as follows:
> {code:java}
> Threads 1 2 3 4
> --------------------------------
> inodes 448 290 226 189
> inode_dir 326 211 170 161
> Total 927 651 535 488 (MD5 calculation about 100 seconds)
> {code}
> The above table shows the time in seconds to load the inode section and the
> inode_directory section, and then the total load time of the image.
> With 4 threads using the above technique, we are able to better than half the
> load time of the two sections. With the patch in HDFS-13694 it would take a
> further 100 seconds off the run time, going from 927 seconds to 388, which is
> a significant improvement. Adding more threads beyond 4 has diminishing
> returns as there are some synchronized points in the loading code to protect
> the in memory structures.
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
