Strong-Stability preserving Runge-Kutta time-steppers ¶

The Gkyl DG solvers use SSP-RK time-steppers. Three steppers are implemented: SSP-RK2, SSP-RK3 and a four-stage SSP-RK3 that allows twice the CFL (for the cost of additional memory) as the other schemes. See [DurranBook] page 56. The schemes are described below. Here, the symbol \(\mathcal{F}\) is used to indicate a first-order Euler update:

\[\mathcal{F}[f,t] = f + \Delta t \mathcal{L}[f,t]\]

where \(\mathcal{L}[f]\) is the RHS operator from the spatial discretization of the DG scheme.

SSP-RK2 ¶

\[\begin{split}f^{(1)} &= \mathcal{F}[f^{n},t^n] \\ f^{n+1} &= \frac{1}{2} f^{n} + \frac{1}{2}\mathcal{F}[f^{(1)},t^n+\Delta t]\end{split}\]

with \(CFL \le 1\).

SSP-RK3 ¶

\[\begin{split}f^{(1)} &= \mathcal{F}[f^{n},t^n] \\ f^{(2)} &= \frac{3}{4} f^{n} + \frac{1}{4}\mathcal{F}[f^{(1)},t^n+\Delta t ] \\ f^{n+1} &= \frac{1}{3} f^{n} + \frac{2}{3}\mathcal{F}[f^{(2)},t^n+\Delta t/2]\end{split}\]

with \(CFL \le 1\). As this scheme has three stages instead of two, it will take about \(1.5X\) longer to run than the SSP-RK2 scheme.

Four stage SSP-RK3 ¶

\[\begin{split}f^{(1)} &= \frac{1}{2} f^{n} + \frac{1}{2} \mathcal{F}[f^{n},t^n] \\ f^{(2)} &= \frac{1}{2} f^{(1)} + \frac{1}{2} \mathcal{F}[f^{(1)},t^n+\Delta t/2] \\ f^{(3)} &= \frac{2}{3} f^{n} + \frac{1}{6} f^{(2)} + \frac{1}{6} \mathcal{F}[f^{(2)},t^n+\Delta t] \\ f^{n+1} &= \frac{1}{2} f^{(3)} + \frac{1}{2} \mathcal{F}[f^{(3)},t^n+\Delta t/2]\end{split}\]

with \(CFL\le 2\). Note that this scheme has four stages, but allows twice the time-step that SSP-RK2 and SSP-RK3, hence will result in a speed up of \(1.5X\) compared to the three-stage SSP-RK3 scheme.

Region of absolute stability ¶

For each of the above schemes, I have plotted below the region of absolute stability. Note that only the RK3 schemes are stable when there is no diffusion in the system, and hence should be prefered.