Register to reply

Matrix Condition Number

by nurfherder
Tags: convergence, iterative solver, matrix
Share this thread:
nurfherder
#1
Jan8-13, 02:39 PM
P: 4
Hello all,

I am new to this forum but am glad I found it, I have a quick question about condition numbers and order of operations.

Given a symmetric positive-definitive matrix with a initial condition number α, is it possible to improve that condition number with a different access pattern? For example, if I access the matrix within the context of an iterative solver (e.g., Jacobin) in column-major order would it improve the condition number over access done in row-major order?

I am doing some personal research into iterative solvers and convergence rates and I would like to know if the condition number can be improved, thus lower the total number of iterations to converge, significantly with something so small.

Thank you.
Phys.Org News Partner Science news on Phys.org
Climate change increases risk of crop slowdown in next 20 years
Researcher part of team studying ways to better predict intensity of hurricanes
New molecule puts scientists a step closer to understanding hydrogen storage
nurfherder
#2
Jan9-13, 10:13 AM
P: 4
Never mind about my initial post, the order of operations as applied to column major versus row major will have no effect on the condition number of the matrix. The access order should not effect the Eigenvalue Spectral Radius.

Does anyone have any clue as to why a iterative algorithm, such as Jacobin, would have less iterations to converge when executed on the GPU versus the CPU? The model and tolerance is exactly the same in both cases, so I cannot understand how the GPU has less iterations using a Krylov search space. I have executed the SAME code for CPU and GPU (except of course the CPU has NO calls to the GPU) on two different sets of CPUs and GPUs (one double precision - Tesla, and one not - Quadro) and get exactly the same result.

Any ideas would be great, I think I might have broke one of my neurons on this one.

Thanks.
AlephZero
#3
Jan9-13, 11:19 AM
Engineering
Sci Advisor
HW Helper
Thanks
P: 6,932
Quote Quote by nurfherder View Post
Does anyone have any clue as to why a iterative algorithm, such as Jacobin, would have less iterations to converge when executed on the GPU versus the CPU? The model and tolerance is exactly the same in both cases, so I cannot understand how the GPU has less iterations using a Krylov search space.
That doesn't make any sense to me. If you do the EXACT same operations, you should get exactly the same results.

The explanation may be something to do with compiler optimisation, compiler bugs, library routines, different implementations of floating point arithmetic, etc. The only way to nail that is compare your two calculations step by step. If there are differences, they will probably show up (if only in the last decimal place) on small matrices as well as on big ones.

nurfherder
#4
Jan9-13, 11:45 AM
P: 4
Matrix Condition Number

You are right - it doesn't make sense to me either. I was just wondering if there was an obvious and therefore easy reason.

Thank you for your help and time.
nurfherder
#5
Jan10-13, 04:41 PM
P: 4
I found the problem.

Turns out that the GPU typically has some round-off error that will benefit the GPU for iterative solvers such that the higher precision of the CPU will take more iterations to converge. The small inaccuracies of the GPU become magnified when doing large sets of summations - such as those found in the Dot-product of the iterative solver I am using.

It is sneaky and is typically solved by using double precision (CUDA arch. 1.3 or greater) or algorithmically with the Kahan approximation.
AlephZero
#6
Jan10-13, 06:08 PM
Engineering
Sci Advisor
HW Helper
Thanks
P: 6,932
Quote Quote by nurfherder View Post
Turns out that the GPU typically has some round-off error that will benefit the GPU for iterative solvers such that the higher precision of the CPU will take more iterations to converge.
Hm.... long before the days of GPUs, I remember a CFD software guru trying to convince me that his code worked better in 32 bit precision arithmetic than in 64 (In fact it didn't work at all in 64, on most problems).

Maybe he gave up trying to sell his CFD software and went into GPU design ...


Register to reply

Related Discussions
Condition number of a matrix is 1 but determinant is almost 0 General Math 2
Matrix condition number question Calculus 1
Condition number of a matrix Calculus & Beyond Homework 0
Condition number Linear & Abstract Algebra 0
Condition number (or estimate) of a very large matrix Linear & Abstract Algebra 0