At 10:51 PM 3/28/00 -0500, you wrote:
>
> Hello!
>
>> > >It seems to be a fact that numerically intensive code spends a lot of time
>> > >performing matrix maths. If this means `array performance' is the
>limiting
>> > >factor in performance why it it `tunnel vision' to concentrate on it?
>
>> >because matrices in real codes are sparse and rarely implemented
>> >with "arrays".
>
>> go get 'em andy! that's giving them a taste of the real (aka 'sparse')
>> world.:)
>
> I agree that matrix arithmetics is not the issue (but still arrays, or
>what do you use?)
A clarification: sparse LA does not use an array to represent an entire matrix. Indeed,
there may be 1D arrays used to store, say, a row of a matrix, or a portion
of a row, as a component of a larger, much more complicated data structure (and
yes, sometimes little 2D arrays to store blocks in a block format; but they are little).
This is *particularly* so in massively parallel, distributed memory computing,
which is my field and by which I am biased. The point was really to say that
even were you to provide an efficient array class, even one that spanned across
distributed memory, it wouldn't be anywhere near sufficient to do sparse LA.
Matrices, in the LA sense, and arrays in the data structure sense, are vastly
different beasts, basically.
Having said all that: you all may be dismissing too quickly the important of
sparse LA in scientific and engineering codes. Yes, there are other kernels, but
this remains probably the single largest consumer of flops on such machines.
Cheers,
Andy
Dr. Andrew J. Cleary
Computational Scientist
Center for Applied Scientific Computing
Lawrence Livermore National Lab
L-561, PO Box 808, Livermore, CA 94551
925-422-1939
--------------------- Object Oriented Numerics List --------------------------
* To subscribe/unsubscribe: use the handy web form at
http://oonumerics.org/oon/
* If this doesn't work, please send a note to owner-oon-list@oonumerics.org
This archive was generated by hypermail 2b29 : Wed Feb 20 2002 - 03:20:11 EST