giraph-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nick West <>
Subject Re: Termination Conditions
Date Tue, 14 Aug 2012 21:08:30 GMT
Thank you all for the replies.

I think Kaushik has the easiest solution for me.  I'll give that a try.

Avery: I want to the vertices to re-activate when they receive a new message, unless all vertices
have voted to halt that iteration.  So not reactivating after voting to halt would not solve
this problem.


On Aug 4, 2012, at 1:49 PM, Avery Ching wrote:

I think it would be pretty straightforward to add an option to not re-activate a halted vertex
on new messages.  Would that satisfy your issue Nick?


On 8/4/12 2:42 AM, Sebastian Schelter wrote:
Hi Nick,

I think you are running into the issue that vertex
activation/scheduling and data transfer are not separated in the
Pregel/Giraph model. I ran into the same problem when trying to
implement Adaptive PageRank on Giraph.

I think the problem comes from the vast generality of Pregel's BSP
model: vertices can send any kind of message to any other vertex.
However, in a lot of algorithms, vertices only send their state to
their neighbors. Other platforms such as GraphLab separate
activation/scheduling from messaging, in this system a vertex can only
schedule its neighbors and once one of the neighbors is invoked, it
can read the state of its direct neighbors.

I think that the Pregel model is simply not suited for such algorithms
(unfortunately), let me know if you find a clever workaround.


2012/8/3 Nick West <<>>
Thanks for the reply.

Is there an easy modification that I can make to remove condition 2?  Can you point me to
the code that addresses this?

The problem I am facing is the following:  At every iteration a non-halted vertex needs messages
from all of its neighbors.  When deciding to send messages, a given vertex doesn't know if
its neighbors will vote-to-halt in the current superstep, thus it must send a message to each
of its neighbors.  In the case that all vertices have voted to stop, the sending of a messages
by any vertex will cause the algorithm to continue, yet in this situation it is desirable
to terminate.

I have worked out a few solutions that involve either increasing the amount of data a vertex
saves each iteration or augmenting the messages sent with additional information, but I think
it would be beneficial, and more general, to allow this type of termination instead.

Do you have any thoughts on this?


On Aug 3, 2012, at 10:14 AM, Alessandro Presta wrote:

Hi Nick,

You are pretty much correct, except that not all vertices need to vote to halt at the same
time: some vertices might have voted to halt at a previous superstep and never received any
messages after then, in which case they are never reactivated.

In other words, I think you can rephrase that as:

All vertices are halted after a given superstep
No messages were sent in that superstep

Hope it helps.


From: Nick West <<>>
Reply-To: "<>" <<>>
Date: Friday, August 3, 2012 2:48 PM
To: "<>" <<>>
Subject: Termination Conditions

Excuse me if this is stated somewhere obvious, but I haven't been unable to find it.  What
are the exact termination criteria for the global algorithm?

Reading the documentation on voteToHalt, looking at the Shortest Path Example code, and looking
at the results of my own application, these two conditions must both hold for the global BSP
algorithm to terminate:

1) All vertices vote to halt in a given superstep
2) No messages are sent in that supersetp

Is that correct?

Nick West

Benchmark Solutions
101 Park Avenue - 7th Floor
New York, NY 10178
Tel + | Mobile +1.646.267.4324<>


Nick West

Benchmark Solutions
101 Park Avenue - 7th Floor
New York, NY 10178
Tel + | Mobile +1.646.267.4324<>

Nick West

Benchmark Solutions
101 Park Avenue - 7th Floor
New York, NY 10178
Tel + | Mobile +1.646.267.4324 <>

View raw message