Schmitt, Andreas (2018)
Numerical Investigation of Parallel-in-Time Methods for Dominantly Hyperbolic Equations.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
Simulations aid in many scientific and industrial applications. A general ambition for these simulations is to keep the time-to-solution as small as possible while maintaining a desired accuracy. Besides with high computational power, this can be achieved by employing multiple processing units with parallelization. Today’s state of the art is the spatial parallelization which provides a very good parallel efficiency.
However, such a parallelization introduces communication and synchronization overheads leading to a maximal number of processing units which can be used efficiently. Applying a parallelization in time on top of the parallelization in space makes using more processing units possible. An issue of the parallel-in-time methods is their problem dependent efficiency. It tends to be generally bad for dominantly hyperbolic problems. The viscous Burgers equation, which for small viscosities falls into that category, is used to investigate two methods of parallelization in time.
First, a look is taken at the Adomian decomposition method (ADM) and possibilities of exploiting additional degrees of parallelism within this method. Its viability is questioned by comparing its discrete version (DADM) to the explicit Runge-Kutta method (ERK). The comparison shows similar restric- tions regarding their maximal time step size for both methods. Furthermore, the DADM leads to larger errors with increasing order of accuracy compared to the ERK. However, discussing the parallelization within the DADM shows a reduction of the runtime complexity from quadratic to linear is possible. With this reduction in the runtime DADM seems to be a viable competitor to the ERK. This is especially true for high-order schemes, as fewer function evaluations have to be run serial. Increasing the order of accuracy is also embarrassingly easy with the DADM compared to the ERK.
The second method investigated in this thesis is the Parareal algorithm. Here, the focus lies on the potential of the implicit Runge-Kutta method with semi-Lagrangian advection (SLIRK) as the coarse solver for the Parareal algorithm. Its potential compared to using the explicit Runge-Kutta method (ERK) and the implicit-explicit Runge-Kutta method (IMEX) is tested with three different benchmarks. The comparison shows the ERK is in contrast to the other two methods not able to provide speedup potential with the chosen benchmarks. For advection dominated problems SLIRK performs better than IMEX due to its stability. The stability of SLIRK leads to speedup potential for a larger range of viscosities with the Parareal algorithm. Still, the instability of Parareal itself causes a decreasing potential with a decreasing viscosity. With an inviscid case the number of iterations to convergence for Parareal is too large to yield a reasonable speedup. An additional result worth mentioning is it was possible to show the importance of predicting the phase of the solution correctly for the convergence of Parareal.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2018 | ||||
Autor(en): | Schmitt, Andreas | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Numerical Investigation of Parallel-in-Time Methods for Dominantly Hyperbolic Equations | ||||
Sprache: | Englisch | ||||
Referenten: | Schäfer, Prof. Dr. Michael ; Schöps, Prof. Dr. Sebastian | ||||
Publikationsjahr: | 23 Juli 2018 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 10 Oktober 2018 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/8328 | ||||
Kurzbeschreibung (Abstract): | Simulations aid in many scientific and industrial applications. A general ambition for these simulations is to keep the time-to-solution as small as possible while maintaining a desired accuracy. Besides with high computational power, this can be achieved by employing multiple processing units with parallelization. Today’s state of the art is the spatial parallelization which provides a very good parallel efficiency. However, such a parallelization introduces communication and synchronization overheads leading to a maximal number of processing units which can be used efficiently. Applying a parallelization in time on top of the parallelization in space makes using more processing units possible. An issue of the parallel-in-time methods is their problem dependent efficiency. It tends to be generally bad for dominantly hyperbolic problems. The viscous Burgers equation, which for small viscosities falls into that category, is used to investigate two methods of parallelization in time. First, a look is taken at the Adomian decomposition method (ADM) and possibilities of exploiting additional degrees of parallelism within this method. Its viability is questioned by comparing its discrete version (DADM) to the explicit Runge-Kutta method (ERK). The comparison shows similar restric- tions regarding their maximal time step size for both methods. Furthermore, the DADM leads to larger errors with increasing order of accuracy compared to the ERK. However, discussing the parallelization within the DADM shows a reduction of the runtime complexity from quadratic to linear is possible. With this reduction in the runtime DADM seems to be a viable competitor to the ERK. This is especially true for high-order schemes, as fewer function evaluations have to be run serial. Increasing the order of accuracy is also embarrassingly easy with the DADM compared to the ERK. The second method investigated in this thesis is the Parareal algorithm. Here, the focus lies on the potential of the implicit Runge-Kutta method with semi-Lagrangian advection (SLIRK) as the coarse solver for the Parareal algorithm. Its potential compared to using the explicit Runge-Kutta method (ERK) and the implicit-explicit Runge-Kutta method (IMEX) is tested with three different benchmarks. The comparison shows the ERK is in contrast to the other two methods not able to provide speedup potential with the chosen benchmarks. For advection dominated problems SLIRK performs better than IMEX due to its stability. The stability of SLIRK leads to speedup potential for a larger range of viscosities with the Parareal algorithm. Still, the instability of Parareal itself causes a decreasing potential with a decreasing viscosity. With an inviscid case the number of iterations to convergence for Parareal is too large to yield a reasonable speedup. An additional result worth mentioning is it was possible to show the importance of predicting the phase of the solution correctly for the convergence of Parareal. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-83286 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften und Maschinenbau | ||||
Fachbereich(e)/-gebiet(e): | 16 Fachbereich Maschinenbau 16 Fachbereich Maschinenbau > Fachgebiet für Numerische Berechnungsverfahren im Maschinenbau (FNB) Exzellenzinitiative Exzellenzinitiative > Graduiertenschulen Exzellenzinitiative > Graduiertenschulen > Graduate School of Computational Engineering (CE) |
||||
Hinterlegungsdatum: | 10 Feb 2019 20:55 | ||||
Letzte Änderung: | 10 Feb 2019 20:55 | ||||
PPN: | |||||
Referenten: | Schäfer, Prof. Dr. Michael ; Schöps, Prof. Dr. Sebastian | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 10 Oktober 2018 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |