5.1.9 iparm(_JT_ resid_) - determines how often to compute and/or use true residual (b - Ax) and how often to compute error estimate based on true residual (errt)

• Current index: 9
• Intent: In
• Default: 0
• Values:

  Variable	Meaning	Value
  none	compute and use only as needed by the method	0
  none	compute b - Ax and errt each time a new iterate is obtained and use as appropriate for the method	1
  none	replace r with b - Ax every MAX(10,$\sqrt{n}$) iterations	2
  none	compute b - Ax every iteration, but only replace r with b - Ax every MAX(10,$\sqrt{n}$) iterations	3

• NOTES:
  • iparm(_JT_ resid_) = 1 and iparm(_JT_ resid_) = 3 are mainly for investigating the behavior of solvers, since time is wasted calculating but not using b - Ax.
  • Use of iparm(_JT_ resid_) > 1 can be very beneficial to CG, and is recommended. For production use and when CPU time is a consideration, use iparm(_JT_ resid_) = 2.
  • Values of iparm(_JT_ resid_) > 1 are ignored by GMRES, since b - Ax is recalculated automatically each iparm(_JT_ nold_) iterations.
  • Values of iparm(_JT_ resid_) > 1 are meaningless for stationary methods (Jacobi, SOR, SSOR), since they do not generate a residual during the course of the algorithm.
  • Values of iparm(_JT_ resid_) > 1 are ignored by BCGS and TFQMR. It doesn't seem to benefit these methods, and seems to harm both in some situations.