incubator-hama-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Edward J. Yoon" <edwardy...@apache.org>
Subject Fwd: Hama
Date Sun, 26 Apr 2009 23:29:40 GMT
Just FYI,

I thought that the simple marking is enough for A(T), the transpose of
matrix A. But, Map/Reduce will only read sequentially. So I guess
communication cost will be increased between nodes for next steps. For
example, A(T)A. The mat-mat mult designed for the sequential process.

Any advices are welcome.

---------- Forwarded message ----------
From: Maksim Tsvetovat <maksim@slonimgroup.com>
Date: Wed, Apr 22, 2009 at 1:18 AM
Subject: Re: Hama
To: "Edward J. Yoon" <edwardyoon@apache.org>

Ed -
we're starting to enter this phase of development, so I'll be working
with Hama soon.

I was actually going to respond to your blog post about matrix
transposition... I think there's something wrong with the idea that
matrix transpose should take 10 minutes on a cluster. After all, this
may be the simplest matrix op in existence. If a dense matrix has been
represented as an array, all I'd have to do is swap the row and column
iterators - which is of course a constant time operation. Do you think
there may be a to do a similar thing in the HBase representation? (of
course, if we have to copy matrices into a new storage that may be
more expensive)

On a sparse matrix you simply reverse the edge direction; if the
representation is a doubly-linked list or something similar then
transposition is also constant-time (+ copying if needed).

I feel strongly about transposition, in my application I'll be using
it a whole lot, and on much larger (mostly sparse) matrices.  I have
always regarded it as an instant op (the way it is in, say, Matlab)
rather then a linear-time op.

Max

ps. just thinking here... maybe I am wrong and there's no better
way... but gotta try!

-- 
Best Regards, Edward J. Yoon @ NHN, corp.
edwardyoon@apache.org
http://blog.udanax.org

Mime
View raw message