🚀 go-pugleaf

In article <8055@sun13.scri.fsu.edu> jac@ds8.scri.fsu.edu (Jim Carr) writes:
>
>I am curious what your Cray performance is.  What does hpm give you for 
>your Mflop rate and what is the vector/scalar breakdown?  

My performance is not spectacular, about 60 Mflops for a reasonably large
problem.  The code to generate the matrices is well vectorized, but it loses
it when it gets to solving the matrix (see below).

>I am not familiar with your matrix problem, but it might have some common 
>features with the one encountered by the Lattice group.  They use a 
>preconditioned conjugate gradient method to solve their system.  What 
>method do you use?  And whose Cray?

My matrix is fairly unstructured.  I have played a bit with CG methods, but
so far without any real success.  I haven't spent enormous time on it either
however.  The problem is that 1) the matrix is complex and 2) it is
devoid of the structure you usually look for when you apply iterative methods
(i.e., diagonal dominance, positive definite, even Hermitian).  Currently
I just solve it using a LU decomposition and Gaussian elimination.  I have
used both Linpack and Lapack on the Cray.  The Cray I have most recently
been using is the Y-MP at NCSA, though I have also had time at various
points on the Cray 2 at the Phillips Laboratory and at Los Alamos.  I also
have used the CM-2.  I haven`t gotten very impressive performance out of
the CM-2 however.  My biggest problem there is that I am using the CMSSL
software (QR solver with both the Paris and slicewise software models)
and it has been bombing me off the machine on matrices of order 1000 or so.
TMC is currently trying to figure out the problem.  Until they solve it,
I am not really in a position to run a large enough problem to really see
if I can get good performance off the CM-2.  I can do an order 1000 problem
on my Sparc 2 in minutes, so that isn't exactly a good test.  I am hoping
to run an order 10,000 this summer on the CM-2.  The largest I have done
on the Cray are order 6,000.

John
-- 
John K. Prentice                      | "We must make good the demands of
Department of Physics And Astronomy   | reason and create a life worthy of
University of New Mexico              | ourselves and of the goals we only
john@aquarius.unm.edu                 | dimly perceive."  - Andrei Sakharov