DiFfRG::assembly_schedule_defaults Namespace Reference

Variables
constexpr double	chunk_work_ns = 8.0e3
constexpr double	worker_work_ns = 64.0e3
constexpr uint	chunk_cap = 256

Variable Documentation

chunk_cap

uint DiFfRG::assembly_schedule_defaults::chunk_cap = 256

inlineconstexpr

Ceiling on chunk_size. queue_length * chunk_size CopyData objects are allocated up front, so an unbounded chunk turns into allocation cost on a mesh that cannot use it.

chunk_work_ns

double DiFfRG::assembly_schedule_defaults::chunk_work_ns = 8.0e3

inlineconstexpr

Work a single chunk should carry for the pipeline hand-off around it to be negligible. ~ 8 us

worker_work_ns

double DiFfRG::assembly_schedule_defaults::worker_work_ns = 64.0e3

inlineconstexpr

Work that has to exist before another pipeline worker pays for itself.

Measured on some cheap loops, where waking workers on a small mesh is a net loss: 128 cells are assembled fastest by a single worker, 512 by two, and only past a few thousand does it pay to use the whole budget. ~ 64 us.