Skip to content

Latest commit

 

History

History
106 lines (85 loc) · 5.59 KB

PLACEMENT.md

File metadata and controls

106 lines (85 loc) · 5.59 KB

Where the partitioner lives

The cell graph, the partition, and the geometric halo belong in linkcell. The rank share, the ring search, and ice clustering stay in seams-core. The placement keeps the linkcell graph here and leaves seams::domain in seams-core.

What each repository owns today

linkcell owns the periodic cell, the fractional bin mesh, the linked-cell walk, lattice shifts, pairs_within, and the gpulite workspace. Bin counts use the perpendicular widths, so a triclinic cell is binned in fractional space. The C application binary interface (lc_*) is the Library of Data Exchange (LODE) hourglass. C++ is a Resource Acquisition Is Initialization (RAII) header over that interface. Python crosses it with DLPack.

seams-core calls the linkcell mesh and the linked-cell walk from the analysis engine (cloned from https://github.com/d-SEAMS/seams-core, default branch). src/neighbours.cpp calls linkcell::pairs_within and linkcell::knearest when SEAMS_HAS_LINKCELL is set. Meson asks for linkcell >= 0.4.0. src/ring.cpp and src/include/internal/ring.hpp identify primitive rings (Franzblau) and tag them with the bulk and confined ice criteria. src/cluster.cpp and clump::Domain are Stoddard's connected components of an ice mask. That Domain names an ice-mask cluster.

The linkcell mesh in src/domain.cpp already decomposes one analysis frame across ranks. The Hilbert curve orders the cells. The cut makes contiguous runs of near equal atom count (Springel, Monthly Notices of the Royal Astronomical Society (MNRAS) 364, 1105, 2005). A rank's halo collects every cell within a caller distance of a cell it owns. reduceRings folds the owned rings with the Message Passing Interface (MPI). The hop length follows the cutoff. maxLvl equals maxDepth / 2. hops equals maxLvl + (maxLvl - 1) / 2. The domain grid is a second fractional binning. kCellsPerRank, kMaxBits, kHaloCells, and kMaxHaloCells cap that grid.

Other consumers

eOn calls linkcell.pairs_within for a fully periodic cutoff list. Vesin covers an open direction. neighbor_list_linkcell is that pair list. eOn stays on the pair list and leaves seams-core unloaded.

rgpot (OmniPotentRPC/rgpot) builds pairs through vesin. Its docs and sources link to the linkcell documentation from the site nav. Pair potentials use rgpot::nlist::PairListCache over vesin, with an internal linked-cell search once a box contains enough cells. A partitioner inside seams-core would sit behind a dependency on the analysis engine, so rgpot would still miss it.

readcon-core (lode-org/readcon-core) reads and writes configuration (CON) frames. It shares the hourglass application binary interface shape. readcon-core stops at configuration frames. The only mentions of linkcell are documentation cross-links.

The dependency already runs from seams-core to linkcell. Reversing it would put Eigen, Franzblau, and PointCloud under every neighbour caller.

Geometry versus analysis

A part is a set of linkcell bins. A cell weight counts the points in that bin. An edge weight counts the pairs_within rows whose endpoints fall in those two bins, periodic images included. Those three objects are properties of the mesh and the cutoff list.

The halo of a part is the set of bins that can hold a point within a caller length of an owned point under the minimum image. The length is cutoff * depth. The bin radius along direction i is the smallest integer r with r * bin_width_i >= length. The halo statement covers bins, the cutoff, and the minimum image.

What is analysis, and stays in seams-core:

  • Share (local, owned, owner) on a PointCloud
  • the ring hop count that chooses depth
  • rings, gatherRings, reduceRings, RingSummary
  • Message Passing Interface collectives
  • subCloud
  • clump::Domain and the ice-mask component

Interface

linkcell, in Rust, C (lc_*), C++ Resource Acquisition Is Initialization, and Python (DLPack):

  • cell_grid returns nx, ny, nz, the bin widths, and a point count per cell. The edge is the one pairs_within uses for the same cutoff and hint.
  • cell_graph adds the periodic adjacency. Each undirected edge's weight is the pair count from pairs_within (half set).
  • partition and lc_partition return one domain id per cell. Methods are a Morton curve, a Hilbert curve, weighted recursive coordinate bisection, and a Fiduccia-Mattheyses pass under a balance bound. With balance <= 0, both use 1 + parts * w_max / total. w_max is the heaviest cell weight. total is the sum of the cell weights. A weighted median assigns that indivisible cell to one side, so the bound matches w_max. The d-SEAMS domain driver derives this ratio for each frame. A caller may pass a looser bound. Every system uses the ratio by default.
  • halo_width(cutoff, depth, bin_width) returns the three bin radii.
  • halo returns the ghost cell ids of one domain.
  • Serial Scotch and parallel Scotch (PT-Scotch) stay optional. The default build leaves libscotch and libptscotch out.
  • The device path computes cell weights, curve keys, and the prefix split through gpulite when a Compute Unified Device Architecture (CUDA) driver is present. Copies both ways sit inside that time.

seams-core will call this partition interface later:

  • build xyz from the cloud and call partition
  • set depth from the ring hop count, then call halo_width and halo
  • map cell ids to atom indices and fill Share
  • leave rings and the rank reduction as they are

eOn passes cell weights through this partition and halo, and through the ring depth, on a packed frame. Callers hand over that packed frame.