diff options
Diffstat (limited to 'Docs/source/running_cpp/parameters.rst')
| -rw-r--r-- | Docs/source/running_cpp/parameters.rst | 28 |
1 files changed, 18 insertions, 10 deletions
diff --git a/Docs/source/running_cpp/parameters.rst b/Docs/source/running_cpp/parameters.rst index 9e38ab133..f50d59d41 100644 --- a/Docs/source/running_cpp/parameters.rst +++ b/Docs/source/running_cpp/parameters.rst @@ -207,7 +207,7 @@ Distribution across MPI ranks and parallelization When using mesh refinement, this number applies to the subdomains of the coarsest level, but also to any of the finer level. -* ``warpx.load_balance_intervals`` (`string`) optional (default `0`) +* ``algo.load_balance_intervals`` (`string`) optional (default `0`) Using the `Intervals parser`_ syntax, this string defines the timesteps at which WarpX should try to redistribute the work across MPI ranks, in order to have better load balancing. @@ -220,27 +220,35 @@ Distribution across MPI ranks and parallelization This relies on each MPI rank handling several (in fact many) subdomains (see ``max_grid_size``). -* ``warpx.load_balance_with_sfc`` (`0` or `1`) optional (default `0`) - If this is `1`: use a Space-Filling Curve (SFC) algorithm in order to - perform load-balancing of the simulation. - If this is `0`: the Knapsack algorithm is used instead. - -* ``warpx.load_balance_efficiency_ratio_threshold`` (`float`) optional (default `1.1`) +* ``algo.load_balance_efficiency_ratio_threshold`` (`float`) optional (default `1.1`) Controls whether to adopt a proposed distribution mapping computed during a load balance. If the the ratio of the proposed to current distribution mapping *efficiency* (i.e., average cost per MPI process; efficiency is a number in the range [0, 1]) is greater than the threshold value, the proposed distribution mapping is adopted. The suggested - range of values is ``warpx.load_balance_efficiency_ratio_threshold >= 1``, which ensures + range of values is ``algo.load_balance_efficiency_ratio_threshold >= 1``, which ensures that the new distribution mapping is adopted only if doing so would improve the load balance efficiency. The higher the threshold value, the more conservative is the criterion for adoption of a proposed distribution; for example, with - ``warpx.load_balance_efficiency_ratio_threshold = 1``, the proposed distribution is + ``algo.load_balance_efficiency_ratio_threshold = 1``, the proposed distribution is adopted *any* time the proposed distribution improves load balancing; if instead - ``warpx.load_balance_efficiency_ratio_threshold = 2``, the proposed distribution is + ``algo.load_balance_efficiency_ratio_threshold = 2``, the proposed distribution is adopted only if doing so would yield a 100% to the load balance efficiency (with this threshold value, if the current efficiency is ``0.45``, the new distribution would only be adopted if the proposed efficiency were greater than ``0.9``). +* ``algo.load_balance_with_sfc`` (`0` or `1`) optional (default `0`) + If this is `1`: use a Space-Filling Curve (SFC) algorithm in order to + perform load-balancing of the simulation. + If this is `0`: the Knapsack algorithm is used instead. + +* ``algo.load_balance_knapsack_factor`` (`float`) optional (default `1.24`) + Controls the maximum number of boxes that can be assigned to a rank during + load balance when using the 'knapsack' policy for update of the distribution + mapping; the maximum is + `load_balance_knapsack_factor*(average number of boxes per rank)`. + For example, if there are 4 boxes per rank and `load_balance_knapsack_factor=2`, + no more than 8 boxes can be assigned to any rank. + * ``algo.load_balance_costs_update`` (`Heuristic` or `Timers`) optional (default `Timers`) If this is `Heuristic`: load balance costs are updated according to a measure of particles and cells assigned to each box of the domain. The cost :math:`c` is |