Performance and Scaling

CosmicGrain cost is governed by both the initial zoom geometry and the dust population spawned during evolution. Nominal resolution alone is therefore not a sufficient resource estimate.

Initial-condition size

The accepted suite ranges from 2.12 million particles for halo 3879 at \(512^3\) to 517.54 million for halo 1534 at \(4096^3\). Measure each file directly before choosing MPI tasks, memory limits, or storage allocations.

Dust time bins

At high resolution, newly spawned dust can become a substantial fraction of the gravity tree. spawn_dust_particle() assigns a synchronized birth timebin. The ongoing gravity timestep criterion then allows dust to migrate toward appropriately long synchronized bins after its first force evaluation. This prevents a newly created particle from entering an unsynchronized bin and causing a collective gravity hang.

Dust-physics cadence

The expensive dust update is cadence controlled rather than called on every gravity step. Processes using the scaled interval employ exact exponential updates where applicable, such as \(1-\exp(-\Delta t/\tau)\), rather than unstable linear approximations.

FOF and SUBFIND

FOF/SUBFIND can dominate runtime when executed frequently. Snapshot cadence and group-finding cadence are separate scientific choices: retain enough halo catalogs to follow the target, but avoid invoking SUBFIND at every fine output unless the analysis requires it.

Before a production launch

1. run through the first domain decomposition and force calculation;
2. record peak resident memory per rank;
3. check particle imbalance and top-level tree balance;
4. estimate output size using the actual HDF5 IC;
5. confirm restart writing; and
6. repeat the test when changing resolution, halo, MPI layout, or major particle-creation parameters.