giraph-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dionysios Logothetis <dlogothe...@gmail.com>
Subject Re: Is it normal that a same algorithm for 4 profiles calculus implemented both in giraph and graphlab gives different output with same input?
Date Wed, 08 Aug 2018 23:02:57 GMT
Hi Francesco,

Here are few thoughts on reasons this may happen:

- I haven't written any graphlab application myself, but the synchronous
vs. asynchronous model can generally be a reason for differences. It is
possible that such differences do not manifest on smaller graphs. What's
the size of the smaller input graphs where the result is equal?
- More generally, and taking into consideration the differences in the
model, how certain are you that the implementations are equivalent?
- For the bigger graphs, do you have a reference result that you can use,
to see what the output should be?

Also, are you planning to contribute your Giraph implementation to Apache
Giraph? That'd be great.


Dionysios



On Mon, Aug 6, 2018 at 5:44 AM, Francesco Sclano <sclano.francesco@gmail.com
> wrote:

> Hi,
> I implemented in giraph the calculus of 4 profiles (
> https://people.csail.mit.edu/jshun/papers/ESBD16.pdf) already implemented
> here https://github.com/eelenberg/4-profiles with graphlab.
> I'm able to run both graphlab e giraph programs. I know well differences
> between graphlab and giraph but I have a doubt: when I use small input
> graphs both programs give exactly the same output. When I use bigger input
> graphs (like http://snap.stanford.edu/data/p2p-Gnutella08.html) they
> produce a slightly different output. It is normal? Maybe it depends by the
> way to partitionate the input graph? Or simply there is a bug in my giraph
> implementation?
>
> --
> Francesco Sclano
>

Mime
View raw message