Jed, Can you simply your test into a small petsc example that repeats this behavior? I want test it myself to make sure the differences are not result of bug in the new implementation.
Hong On Feb 24, 2010, at 2:55 AM, Jed Brown wrote: > I notice low-bit differences between the new ILU and versions from > last > month. From one of my tests: > > -Algebraic residual |x|_1 9.35e-16 |x|_2 2.59e-16 |x|_inf > 1.67e-16 > +Algebraic residual |x|_1 1.09e-15 |x|_2 3.00e-16 |x|_inf > 1.67e-16 > -Algebraic residual |x|_1 6.18e-12 |x|_2 2.59e-13 |x|_inf > 4.08e-14 > +Algebraic residual |x|_1 6.19e-12 |x|_2 2.59e-13 |x|_inf > 4.08e-14 > -Algebraic residual |x|_1 4.85e-12 |x|_2 1.95e-13 |x|_inf > 4.73e-14 > +Algebraic residual |x|_1 4.84e-12 |x|_2 1.95e-13 |x|_inf > 4.73e-14 > -Algebraic residual |x|_1 3.36e-13 |x|_2 1.56e-14 |x|_inf > 3.06e-15 > +Algebraic residual |x|_1 3.32e-13 |x|_2 1.54e-14 |x|_inf > 3.05e-15 > > I'm guessing this comes from traversing rows in the other order during > backsolve. This is no problem at all since both are equally valid, I > just wanted to note that some rounding effects are visible. > > Jed
