Friday, April 13, 2012

MadLINQ: New large scale matrix computation paper from MSR asia

I got this from Aapo Kyrola:

Please note that following paper:
"MadLINQ: Large-Scale Distributed Matrix Computation for the Cloud"http://dl.acm.org/citation.cfm?doid=2168836.2168857
It won the Eurosys best paper award, and really is an excellent paper from
MSR Asia. There are two reasons to read it:
1) It has similar application domain and motivation to GraphLab: they argue that many data mining
and ML algos can be represented as matrix algorithms. This work is focused
on dense/ish matrices and compares against ScalaPACK.  It is not direct competitor
to GraphLab other than maybe in the matrix factorization domain.
2) It is excellent Systems work and paper. It got very high praise from the
chairs of being a top-to-bottom systems work, from abstraction to verification
to detailed evaluation.

No comments:

Post a Comment