[
https://issues.apache.org/jira/browse/GIRAPH-275?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13425864#comment-13425864
]
Eli Reisman commented on GIRAPH-275:
------------------------------------
Since our IO formats/record readers fake the Hadoop side out when building
splits right now, Hadoop is not going to hand us the ready-made locations and
offsets we need. I will need to hand-check the block offsets etc. for our input
files to ensure the right blocks local to not just the file the split comes
from but the split itself are the ones we supply when we are populating the
locations info, but that code already exists in Hadoop and can be adapted
easily enough. i am also very interested in getting the data into the znode
path for now rather than having the master (or worse, every worker) read the
znode data to order its split list, but I will do what is needed to make this
work, and refine it once it does, if this turns out to be unworkable. I have a
couple of other ideas how to minimize the impact of these reads if this turns
out to be the case.
My main goal right now is to get this to work and run many tests see if it
moves the needle on speeding up the INPUT_SUPERSTEP and does in fact lower
network throughput at that stage for data load in (if its done right, I suspect
it will.) If this helps, I can do various things to make this a more palatable
solution at that point. Data locality seems like a real win for Giraph scale
out in general where it can be taken advantage of. Will post more soon...
> Restore data locality to workers reading InputSplits where possible without
> querying NameNode, ZooKeeper
> --------------------------------------------------------------------------------------------------------
>
> Key: GIRAPH-275
> URL: https://issues.apache.org/jira/browse/GIRAPH-275
> Project: Giraph
> Issue Type: Improvement
> Components: bsp, graph
> Affects Versions: 0.2.0
> Reporter: Eli Reisman
> Assignee: Eli Reisman
> Fix For: 0.2.0
>
> Attachments: GIRAPH-275-1.patch
>
>
> During INPUT_SUPERSTEP, workers wait on a barrier until the master has
> created a complete list of available input splits. Once the barrier is past,
> each worker iterates through this list of input splits, creating a znode to
> lay claim to the next unprocessed split the worker encounters.
> For a brief moment while the master is creating the input split znodes each
> worker iterates through, it has access to InputSplit objects that also
> contain a list of hostnames on which the blocks of the file are hosted. By
> including that list of locations in each znode pathname we can allow each
> worker reading the list of available splits to sort it so that splits the
> worker attempts to claim first are the ones that contain a block that is
> local to that worker's host.
> This allows the possibility for many workers to end up reading at least one
> split that is local to its own host. If the input split selected holds a
> local block, the RecordReader Hadoop supplies us with will automatically read
> from that block anyway. By supplying this locality data as part of the znode
> name rather than info inside the znode, we avoid reading the data from each
> znode while sorting, which is only currently done when a split is claimed and
> which is IO intensive. Sorting the string path data is cheap and faster, and
> making the final split znode's name longer doesn't seem to matter too much.
> By using the BspMaster's InputSplit data to include locality information in
> the znode path directly, we also avoid having to access the
> FileSystem/BlockLocations directly from either master or workers, which could
> also flood the name node with queries. This is the only place I've found
> where some locality information is already available to Giraph free of
> additional cost.
> Finally, by sorting each worker's split list this way, we get the
> contention-reduction of GIRAPH-250 for free, since only workers on the same
> host will be likely to contend for a split instead of the current situation
> in which all workers contend for the same input splits from the same list,
> iterating from the same index. GIRAPH-250 has already been logged as reducing
> pages of contention on the first pass (when using many 100's of workers) down
> to 0-3 contentions before claiming a split to read.
> This passes 'mvn verify' etc. I will post results of cluster testing ASAP. If
> anyone else could try this on an HDFS cluster where locality info is supplied
> to InputSplit objects, I would be really interested to see other folks'
> results.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
