Update configuration-acceleration.md

Add recommendation to use restart mode in `IQN-IMVJ` method for large-scale simulations.

pages/docs/configuration/configuration-acceleration.md

@@ -87,5 +87,6 @@ For quasi-Newton, the configuration is more involved and requires some attention
 * In the first iteration, quasi-Newton methods don’t provide an estimate for the Jacobian yet. Thus, the first iteration is an under-relaxed fixed-point iteration, for which we have to define the parameter <tt>initial-relaxation</tt>. 0.1 is a robust choice. Too small values can render the information from the first iteration too coarse for the calculation of a good Jacobian estimate. Too large values might lead to stability problems.
 * The parameter <tt>max-used-iterations</tt> specifies the maximum number of previous iterations used to generate the data basis for Jacobian estimation. In particular for small simulations with only few degrees for freedom, this is an important parameter. It should be chosen to be smaller than half of the total number of degrees of freedom at the interface. For large-scale simulations 100 is a robust choice.
 * The parameter `time-windows-reused` also limits the number of previous iterations, but in a per-time-window fashion. Iterations from older time windows than `time-windows-reused` are dropped. Note that, as we don’t know the number of iterations per time window a priori, this is not equivalent to setting max-used-iterations. For `IQN-IMVJ`, this parameter can be set to 0 as information from past time windows is implicitly used in the modified Jacobian norm minimization. For `IQN-ILS`, this parameter is an important tuning parameter, in particular if no filtering or filtering with a very low threshold is used. The optimum highly depends on the application, the used solvers and also the grid resolution. We recommend to choose a rather large value (10-30) and combine it with effective filtering (e.g., `QR2` with limit 1e-2) as a starting point for further optimization. With increasing degree of non-linearity of the considered application, the optimal value for `time-windows-reused` is expected to decrease.
+* While the `IQN-IMVJ` method exhibits excellent convergence properties as a result of implicit reuse of prior information, it also causes quadratic storage and runtime complexity because of the explicit computation of Jacobian matrix. To make this method practical for large-scale simulations, we introduced a set of `imvj-restart-mode` variants, which reduce the method's complexity to a linear one. In particular, the `RS-SVD` mode gives good performance comparing to other options while retaining the original MV convergence properties, although its performance slightly depends on the choose of the truncation parameter.
 
 Quasi-Newton acceleration is a good choice for strong interactions. Please note that a necessary prerequisite for convergence of the implicit coupling loop is the proper convergence of each participant internally. Inner convergence measure (e.g. of the fluid solver) should be two orders of magnitude stricter than the coupling convergence-measure to achieve good performance with quasi-Newton.