python api intro

Introduction {#section:python-api-examples:introduction}

• What is a action set
• How to set up a project
• ...

Parallelisation and GPU support

Follow the configuration recommendation of Section [section:particles:swift:hpc]{reference-type="ref" reference="section:particles:swift:hpc"} before you continue.

Domain decomposition

All domain decomposition is realised via your main routine, and you don't have to modify anything besides your main loop if you work shared memory only. Once you enable MPI, too, you will have to make some minor modifications.

 allows you to control the domain decomposition via explicit splits (and mergers) of subdomains. However, we rely on the load balancing toolbox here. The toolbox collects some global information and then triggers splits and merges accordingly. You have to perform the following steps:

• Include

  #include "toolbox/loadbalancing/RecursiveSubdivision.h"

• Create an instance of the load balancer on each MPI rank. It is convenient to insert

  toolbox::loadbalancing::RecursiveSubdivision loadBalancer;

  just before the initial tarch::mpi::Rank::getInstance().isGlobalMaster() check. The object accepts a lot of configuration parameters which you can look up in the source code documentation.

• You have to inform your load balancer when an individual step has terminated. This tells the balancer to collect some global statistics and (eventually) to trigger some rebalancing.

  loadBalancer.finishStep();

  I typically use the load balancer only within the grid construction. Totally dynamic load balancing is something this simple pre-manufactured balancer doesn't properly support anyway.

  int gridConstructionSteps = 0; while (peano4::parallel::SpacetreeSet::getInstance().getGridStatistics().getStationarySweeps()<5) examples::particles::observers::CreateGrid createObserver; peano4::parallel::SpacetreeSet::getInstance().traverse(createObserver); gridConstructionSteps++; loadBalancer.finishStep();