You know, the more that I work with map-reduce, the less I find that is
really hard to do with it.

Graph traversal is a great case.  Looks hard because you potentially have
lots of passes through MR to get your result.

But most problems of practical import scale the number of map-reduce passes
with the diameter of the graph.  For small world graphs, this diameter is
much less than log N (that is what 6 degrees of separation is all about!)
and in some cases looks more like log log N.

To put this in perspective, if the planet has 10^10 people (I call this 10
billion, but this is confusing) and each person has 100 computers, then log
N scaling for finding all shortest paths will require somewhere around 30
passes.   Log-log scaling will require 5.  This is not a big deal.


On Thu, Aug 28, 2008 at 7:31 AM, Robin Anil <[EMAIL PROTECTED]> wrote:

> I havent gone through the features of it. .But from the motive of it.
> Parallelizing it by threads will help is cases where the problem is not
> easy
> to reduce to Map-Reduce.(like Graph Traversal) There will be counter
> arguments also. Grid-Computing, p2p algorithms are tailor made for such a
> system
>
> Robin
>



-- 
ted
  • JPFF Grant Ingersoll
    • Re: JPFF Robin Anil
      • Re: JPFF Ted Dunning

Reply via email to