Parallelisation and Performance Analysis of a TreeSPH Code for Galaxy Simulations

In cosmological simulations, the Lagrangian method Smoothed Particle Hydrodynamics is often applied to cover gas dynamics and combined with tree algorithms for long-range potentials like the Barnes-Hut method to include self-gravity and derive the nearest neighbour lists efficiently. In this thesis,...

Full description

Bibliographic Details
Main Author: Baloglu, Maximilian Volkan
Format: Others
Language:English
Published: KTH, Numerisk analys, NA 2014
Online Access:http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-147340
Description
Summary:In cosmological simulations, the Lagrangian method Smoothed Particle Hydrodynamics is often applied to cover gas dynamics and combined with tree algorithms for long-range potentials like the Barnes-Hut method to include self-gravity and derive the nearest neighbour lists efficiently. In this thesis, a so-called TreeSPH code is parallelized by using MPI and subsequently the performance is analysed. For the domain decomposition to the processes, the structure of an octree is examined and space filling curves are applied to achieve well-working dynamical load balancing. For an efficient parallel SPH calculation, a novel method with a localised boundary handling is proposed to reduce communication overhead === Inom kosmologiska simulationer är den Lagrangianska metoden Smoothed Particle Hydrodynamics en vanligt förekommande metod för att täcka gasdynamik och kombineras med trädalgortimer för långdistanspotentialer, exempelvis Barnes-Huts metod för att inkludera självgravitation och effektivt konstruera listor med de närmaste grannarna. I detta examensarbete parallelliseras en så kallad TreeSPH-kod med hjälp av MPI, därefter analyseras prestandan. Gällande domändekomposition av processerna så undersöks strukturen av en octree där rymdfyllande kurvor appliceras för att uppnå en väl fungerande dynamisk lastbalansering. För en effektiv parallell SPH beräkning föreslås en ny metod med lokal randbehandling för att reducera kommunikation.