giraph-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Claudio Martella (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (GIRAPH-528) Decouple vertex implementation from edge storage
Date Tue, 12 Mar 2013 11:53:14 GMT

    [ https://issues.apache.org/jira/browse/GIRAPH-528?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13599949#comment-13599949
] 

Claudio Martella commented on GIRAPH-528:
-----------------------------------------

That's exactly what I meant. In-memory, it would not change anything as it would keep the
current architecture. The big difference would be that we could stop assuming that the user
might have additional private data in his/her Vertex. What this buys us, is that we can write
back OOC only vertices that have changed state, and the same goes for Edges. We'd have to
hook the methods that change this state to get this done (setVertexValue, addEdge etc.).

Basically we would end up managing this data like the OOC messages, with multiple files (potentially
compacting them when they exceed a certain number?), as we would spill to disk sequentially
instead of overwriting old values with seeks. At loading time, you keep the values coming
from the most recent file.

Does this make sense? Any comments?


                
> Decouple vertex implementation from edge storage
> ------------------------------------------------
>
>                 Key: GIRAPH-528
>                 URL: https://issues.apache.org/jira/browse/GIRAPH-528
>             Project: Giraph
>          Issue Type: Improvement
>            Reporter: Alessandro Presta
>            Assignee: Alessandro Presta
>         Attachments: GIRAPH-528.patch, GIRAPH-528.patch, GIRAPH-528.patch, GIRAPH-528.patch,
GIRAPH-528.patch, GIRAPH-528.patch
>
>
> This is meant to address the following issues:
> 1) The Vertex hierarchy is too complex and sometimes hard to work with (Vertex, SimpleVertex,
MutableVertex, SimpleMutableVertex...).
> 2) Changing the underlying edge storage implementation for an existing algorithm requires
editing your vertex to extend a different one.
> 3) In the general case (e.g. when not using ByteArrayVertex with the current EdgeStore),
moving edges from the EdgeStore to the vertices is an additional step that can be avoided.
> My proposal is the following:
> - Make EdgeStore an interface. An implementation should deal with (concurrent) insertion
of edges during input superstep; iteration over a vertex's edges during computation; insertion/deletion
of edges during mutations (optional?); checkpointing.
> - The default EdgeStore will be the current byte-array implementation, which is generic
(works with any choice of <I, V, E, M>) and reasonably optimized.
> - Only one Vertex class, which the user extends for the sole purpose of defining compute().
I don't necessarily agree that it should be an interface, because we still want to provide
methods like getEdges(), sendMessage(), and those should delegate to the EdgeStore/MessageStore
of choice.
> - Switching edge storage implementation is done by passing the EdgeStore class as an
option. One can also define his own ad-hoc EdgeStore (e.g., backed by primitive arrays).
> I think we should also extend this idea to MessageStore, making it possible to override
that functionality too.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message