Dark Matter Haloes

Source

Halo data is read from the same merger tree files SAGE26 consumes at runtime, in either of two formats — ViSAGE detects which from the file itself, so no extra configuration is needed:

• lhalo_binary (TreeName.N, e.g. trees_063.0) — the packed C struct described below.
• lhalo_hdf5 (TreeName.N.hdf5) — one HDF5 group per tree (Tree0, Tree1, …), each holding the halo properties as separate datasets.

Fields loaded by ViSAGE

Field	Type	Units (raw)	Units (viewer)
Pos[3]	float32	Mpc/h	Mpc/h (unchanged)
Mvir	float32	10¹⁰ Msun/h	Msun (×1e10/h)
SnapNum	int32	—	—

Only Pos and Mvir are used for rendering. All other fields in the struct (Vel, Spin, VelDisp, Vmax, etc.) are read but discarded after the per-snapshot filter.

Halo mass is not always called Mvir

Tree files disagree on which column carries the halo mass, so ViSAGE tries several and uses the first one that is actually populated (present-but-all-zero columns are skipped):

Format	Candidates, in order
lhalo_binary	Mvir, M_TopHat, M_Mean200
lhalo_hdf5	Mvir, Group_M_Crit200, M_Crit200, Mass_200crit, M200c, Group_M_TopHat200, M_TopHat200, M_TopHat, Group_M_Mean200, M_Mean200, Mass_200mean

The load line names the column whenever it isn't Mvir:

Haloes: 79,890 loaded (halo mass from Group_M_Crit200)

Rvir, Vvir, the mass floor, colouring and filtering all follow the column that was chosen. Positions and masses in HDF5 trees are likewise resolved by alias (SubhaloPos, SubhaloVMax, …), and positions stored in kpc/h are converted to Mpc/h automatically — detected by comparing against BoxSize.

Host haloes only

ViSAGE renders FOF centrals. In lhalo_binary trees satellites carry Mvir = 0, so the mass floor excludes them on its own; in HDF5 trees the mass columns are populated for satellites too, so centrals are selected explicitly via FirstHaloInFOFGroup. For The300 at z=0 this gives 79,890 host haloes — exactly the number of Type == 0 galaxies SAGE writes for that snapshot.

Mass floor

The default mass floor is 1e10 Msun. This removes very low-mass substructure that would otherwise dominate point count without contributing visible structure. Adjust with --min-halo-mass.

lhalo_binary file format

Each tree file has the following binary layout:

int32   nforests
int32   nhalos_total
int32[nforests]  nhalos_per_forest
HaloStruct[nhalos_total]

HaloStruct is 19 fields; the full dtype is defined in visage/io/halo_reader.py:HALO_DTYPE.

Parallel loading

ViSAGE reads the N lhalo_binary tree files in parallel (joblib). For miniMillennium with 8 files and 8 CPUs, all files load simultaneously.

HDF5 trees are handled differently: h5py serialises the per-tree reads, so parallelising them gains nothing (measured: threads no faster, processes ~4× slower). Instead each file is scanned once into an in-memory index of every snapshot, and each snapshot is then served from it. For The300 — 152k trees, 5.5M host haloes — that scan takes ~40 s on first load, after which every snapshot resolves in milliseconds, and preloading all 129 snapshots costs no further I/O.