Snapshots

The snapshots are stored using the HDF5 format and are almost compatible with Gadget-2 (fully compatible outside of cosmological runs). They do, however, contain a large set of extensions including units, meta-data about the code and runs as well as facilities to quickly access the particles in a specific region of the simulation volume.

Header

The header group (/Header) contains basic information about the simulation and the number of particles. For compatibility reasons, the SWIFT snapshot’s header format is identical to the Gadget-2 one with small additions.

In addition to the standard quantities, the header contains a field Code that is always set to the string SWIFT, which can be used to identify SWIFT-generated snapshots and hence make use of all the extensions to the file format described below.

The most important quantity of the header is the array NumPart_Total which contains the number of particles of each type in this snapshot. This is an array of 6 numbers; one for each of the supported types. Following the Gadget-2 convention, if that number is larger than 2^31, SWIFT will use the NumPart_HighWord field to store the high-word bits of the total number of particles. The field NumPart_ThisFile contains the number of particles in this sub-snapshot file when the user asked for distributed snapshots (see Snapshots); otherwise it contains the same information as NumPart_Total. Note however, that there is no high-word for this field. We store it as a 64-bits integer [1]. The field NumFilesPerSnapshot specifies the number of sub-snapshot files (always 1 unless a distributed snapshot was asked for) and ThisFile the id of that specific file (always 0 unless a distributed snapshot was asked for).

The field TotalNumberOfParticles gives the total number of particles of each type as a 64 bit integer. This allows the total number of particles to be read directly with no calculation required even if there are 2^31 or more particles. This field is equal to NumPart_ThisFile if the snapshot is not distributed over multiple files.

The field InitialMassTable contains the mean initial mass of each of the particle types present in the initial conditions. This can be used as estimator of the mass resolution of the run. The masses are expressed in internal units.

The field OutputType contains information about the kind of output this snapshot is. The possible values are:

OutputType	Definition
FullVolume	Regular vanilla snapshot
SubSampled	Snapshot where some particle types were sub-sampled
LineOfSight	Line-of-sight snapshot
FOF	Friends-Of-Friends Halo Catalogue

The RunName field contains the name of the simulation that was specified as the run_name in the Meta Data section of the YAML parameter file.

The System field contains the name of the machine where the MPI rank 0 was placed. This name is whatever UNIX’s gethostname() function returns on that system. Similarly, the SnapshotDate field contains the date and time when the file was written.

The TimeBase_dloga field contains the change in logarithm of the scale-factor corresponding to a time-step of length 1 on the integer time-line. This is the smallest time-step size that the code can use. This field is zero in non-cosmological runs. Similarly, the field TimeBase_dt contains the smallest time-step size (in internal units) that the code can take. This would be the increase in time a particle in the time-bin one would have. Note that in cosmological runs this quantity evolves with redshift as the (logarithm of the) scale-factor is used on the integer time-line.

The field SelectOutput will contain the name of the Output Selection used for this specific output and will take the value Default if no such selection (or the default one) was used.

If a sub-sampling of the particle fields was used, then the header additionally contains a field describing the fraction of the particles of each type that were written to the snapshot. Note, however, that when sub-sampling the fields NumPart_Total, NumPart_HighWord, and NumPart_ThisFile contain the number of particles actually written (i.e. after sub-sampling), not the total number of particles in the run.

The field CanHaveTypes contains information about whether a given particle type is to be expected in snapshots of the run. For instance, a simulation with star formation switched on, the code may not have formed a star yet but might in future snapshots. This allows reading tools to distinguish fields they will never expect to find in a given simulation from fields that may be present in other outputs.

Finally, the shift that may have been applied to all the particles upon reading the ICs (See Initial Conditions) is added to the header in the field Shift. This is expressed in internal units.

Meta-data about the code and run

Several groups at the root of the files only contain attributes and are used to store some meta-data about the simulation and the code itself.

Code

The group /Code contains basic information about the version of the code that was used to run the simulation that dumped this snapshot. Versions of the libraries used to compile the code as well as information about the compiler and the flags used are stored. The most important element here are the git SHA and configuration parameters of the code. Alongside the compiler flags, policies and used parameters, these allow to reproduce exactly an older run.

Cosmology

The group /Cosmology contains information about the cosmological model used for this simulation. The first important field is the attribute Cosmological run which is set to 1 for cosmological runs and to 0 otherwise. This allows users to quickly distinguish between these two main modes. Most values in this section only make sense for cosmological runs.

All quantities are expressed in the internal system of units (note that this may differ from the units used in the particle arrays). Values like the look-back time are given for the redshift (or scale-factor) of this snapshot.

Policy

The group /Policy list the engine policies (defined in src/engine.h) that were activated in the run that dumped this snapshot. The policies roughly translate to the main run-time parameters of SWIFT.

GravityScheme

HydroScheme

StarsScheme

SubgridScheme

This section of the meta-data mainly contains basic information about the flavour of sub-grid schemes used in the simulation. This is typically a list of attributes describing the parameters of the model. Users willing to add information can edit the functions chemistry_write_flavour(), cooling_write_flavour(), etc. located in the i/o header of each scheme.

The other important output stored in that group is the NamedColumns sub-group. In it, we store the names of the columns of larger particle arrays that are stored as large n-dimensional arrays. For instance, in the EAGLE model, the individual chemical element fractions of each particles are stored as a Nx9 array, where N is the number of particles (See Chemical tracers). This array is labeled ElementMassFractions and is used instead of 9 individual 1-d arrays. In the NamedColumns sub-group we store as an array of strings the name of each of the 9 individual columns. In this case, the name of the 9 elements traced by the model. This array has the same name as the particle array it corresponds to; here ElementMassFractions. The same mechanism is used for other quantities stored in a similar fashion. This allows external tools reading SWIFT snapshots to give meaningful names to more complex entries of the particle arrays.

Unit systems

The snapshots contain two groups at the root containing information about the unit systems used in the snapshots.

The main one Units contains the units used in the snapshot. In a similar fashion to what is done for the parameter files (see Internal Unit System), SWIFT specifies only the basic units. These are the units of mass (U_M), length (U_L), time (U_t), electric current (U_I) and temperature (U_T). These are specified in units of their CGS equivalents (gram, centimeter, second, Ampere, Kelvin). All the quantities present in the particle arrays are expressed in this system of units. For each quantity, SWIFT gives the conversion factor in terms of these units. For instance, the internal energy per unit mass would be expressed as U_L^2 U_t^-2, which in the CGS unit system translates to \(cm/s^2 = erg/g\).

The second group InternalCodeUnits contains the unit system that was used internally by the code when running the simulation. This is in most cases the same system as given in Units but since users can specify a different system for the snapshots, there might be cases where they differ. As this system only relates to what was used inside the code and not in the snapshots themselves, this group is mostly here to report on the code’s run-time behaviour and is used to express all the quantities in the meta-data (e.g. in the cosmology group or the softening lengths in the gravity group).

Used and unused run-time parameters

The groups /Parameters and UnusedParameters located at the root of the file contain the list of all the run-time parameters used by the run with their values and the list of parameters that were in the YAML but were not read. The content of these two groups is identical to the used_parameters.yml and unused_parameters.yml files produced by SWIFT when starting a run (See the File format and basic information section of the documentation).

Structure of the particle arrays

There are several groups that contain ‘auxiliary’ information, such as Header. Particle data is placed in separate groups depending of the type of the particles. There are currently 6 particle types available. The type use the naming convention of Gadget-2 (with the OWLS and EAGLE extensions). A more intuitive naming convention is given in the form of aliases within the file. The aliases are shown in the third column of the table.

HDF5 Group Name	Physical Particle Type	HDF5 alias	In code enum part_type
/PartType0/	Gas	/GasParticles/	swift_type_gas
/PartType1/	Dark Matter	/DMParticles/	swift_type_dark_matter
/PartType2/	Background Dark Matter	/DMBackgroundParticles/	swift_type_dark_matter_background
/PartType3/	Sinks	/SinkParticles/	swift_type_sink
/PartType4/	Stars	/StarsParticles/	swift_type_star
/PartType5/	Black Holes	/BHParticles/	swift_type_black_hole
/PartType6/	Neutrino Dark Matter	/NeutrinoParticles/	swift_type_neutrino

The last column in the table gives the enum value from part_type.h corresponding to a given entry in the files.

For completeness, the list of particle type names is stored in the snapshot header in the array /Header/PartTypeNames. The number of types (aka. the length of this array) is stored as the attribute /Header/NumPartTypes.

Each group contains a series of arrays corresponding to each field of the particles stored in the snapshots. The exact list of fields depends on what compile time options were used and what module was activated. A full list can be obtained by running SWIFT with the -o runtime option (See Output Selection for details). Each field contains a short description attribute giving a brief summary of what the quantity represents.

Note that the HDF5 names of some fields differ from the GADGET-2 format for initial condition files (see Initial Conditions) that mixes singular and plural names, which in snapshot files are all plural by default (e.g. InternalEnergies in snapshots versus InternalEnergy in initial conditions).

All the individual arrays created by SWIFT have had the Fletcher 32 check-sum filter applied by the HDF5 library when writing them. This means that any eventual data corruption on the disks will be detected and reported by the library when attempting to read the data.

Additionally, some compression filter may have been applied to the fields. See the Snapshots section of the parameter file description for more details.

Unit information for individual fields

Each particle field contains meta-data about the units and how to convert it to CGS in physical or co-moving frames. The meta-data is in part designed for users to directly read and in part for machine reading of the information. Each field contains the exponent of the scale-factor, reduced Hubble constant [2] and each of the 5 base units that is required to convert the field values to physical CGS units. The base assumption is that all fields are written in the co-moving frame (see below for exceptions). These fields are:

Meta-data field name	Description
U_L exponent	Power of the length unit
U_M exponent	Power of the mass unit
U_t exponent	Power of the time unit
U_I exponent	Power of the current unit
U_T exponent	Power of the temperature unit
a-scale exponent	Power of the scale-factor
h-scale exponent	Power of the reduced Hubble constant

These are used by the swiftsimio python library to read units and we encourage users to use this meta-data directly in their automated tools.

As an example, the fluid densities (which are written in the co-moving frame) have the following (non-zero) conversion factors:

• U_L exponent: -3

• U_M exponent: 1

• a-scale exponent: -3

This condensed information is stored in the string Expression for physical CGS units, which in the case of the densities would read a^-3 U_M U_L^-3 [ g cm^-3 ]. The values of the U_x can be found in the Units System group at the root of the snapshot (see above). Note that only unit factors with non-zero exponents are printed to this string.

Additionally, the meta-data contains the numerical conversion factor from the field to co-moving CGS and physical CGS assuming the units in the Unit System group. These are:

• Conversion factor to CGS (not including cosmological corrections

• Conversion factor to phyical CGS (including cosmological corrections)

These are designed for the users to directly use if they don’t want to compute the individual exponents themselves. As an example, in the case of the densities and assuming the usual system of units (\(10^{10} \rm{M}_\odot\), \(100 \rm{km/s}\), \(\rm{Mpc}\)) at redshift 0.1, the conversion factors are:

• Conversion to CGS: \(6.76814403 \times 10^{-31}\)

• Conversion to physical CGS: \(9.00808555 \times 10^{-31}\)

In the case of a non-cosmological simulation, these two expressions are identical since \(a=1\).

In some special cases, the fields cannot be meaningfully expressed as co-moving quantities. In these exceptional circumstances, we set the value of the attribute Value stored as physical to 1. And we additionally set the attribute Property can be converted to comoving to 0.

Particle splitting metadata

When particle splitting is turned on (see File format and basic information; by using particle_splitting=1 in the parameter file) some particles in the output may have been created from the ‘splitting’ of a single, over-massive, particle.

There are three fields, associated with all gas, star, and black hole particles, that can be used to understand if, and how, these particles were split.

These three fields are:

• ProgenitorIDs, the IDs of the gas particles in the initial conditions that is the direct progenitor of this particle.

• SplitCounts, the number of times this gas particle has been split; or, if a star or black hole, how many times the gas particle that became this star (or black hole seed) was split before becoming so.

• SplitTrees, a binary tree (encoded as a 64 bit integer) showing how this particle was split. Each item in the tree shows whether this particle retained its original ID (encoded as 0) or was given a new ID (encoded as 1) in the splitting event. This data is enough to completely reconstruct the splitting history of the particles.

For example, if a particle has been split 5 times (SplitCounts=5 for this particle), and has a binary tree of “10010”, it retained its original ID in the first event, was given a new one in the second event, for the next two events it retained its new ID (obtained in the second event), and finally was given a new ID in the final event. Throughout this process, the value of ProgenitorIDs remained the same. Through this system, we can ensure that the combination of ProgenitorID and this binary tree corresponds to a fully traceable, unique, identifier for every particle in the simulation volume.

Note that we can only track 64 splitting events for a given particle, and after this the binary tree is meaningless. In practice, however, such a high number of splitting events is extremely unlikely to occur. The logging of extra splits can optionally be activated. When particles reach 64 splits, their tree information is reset but the status prior to the reset is stored in a log file allowing for the reconstruction of the full history even in the cases where the maximum is reached.

An example is provided in examples/SubgridTests/ParticleSplitting, with a figure showing how one particle is split (eventually) into 16 descendants that makes use of this metadata.

Quick access to particles via hash-tables

The particles are not sorted in a specific order when they are written to the snapshots. However, the particles are sorted into the top-level cell structure used internally by the code every time a tree rebuild is triggered. The top-level cells are a coarse-grained mesh but knowing which particle belongs to which cell can nevertheless be useful to rapidly access particles in a given region only.

One important caveat is that particles are free to drift out of their cells between rebuilds of the tree (but not by more than one cell-length). If one wants to have all the particles in a given cell, one has to read all the neighbouring cells as well. We note that for image making purposes, for instance to generate a slice, this is typically not necessary and reading just the cells of interest is sufficient.

At the root of the HDF5 file, the Cells group contains all the relevant information. The dimension of the top-level grid (a triplet of integers) is given by the attribute Cells/Meta-data/dimension and the size of each cell (a triplet of floating-point numbers) is given by the attribute Cells/Meta-data/size. All the cells have the same size but for non-cubic simulation volumes the cells themselves can have different sizes along each axis.

The /Cells/Centres array gives the centre of each of the top-level cells in the simulation volume. Both the cell sizes and positions of the centres are expressed in the unit system used for the snapshots (see above) and are hence consistent with the particle positions themselves.

Once the cell(s) containing the region of interest has been located, users can use the /Cells/Files/PartTypeN/, /Cells/Counts/PartTypeN/ and /Cells/OffsetsInFile/PartTypeN/ to retrieve the location of the particles of type N in the /PartTypeN arrays. These contain information about which file contains the particles of a given cell. It also gives the offset from the start of the /PartTypeN array in that file at which the particles of that cell are located and how many particles are in the cell. This allows to read a single contiguous section of the whole array by directly reading the slab starting at the offset and with the given length.

The cells, files, offsets in file and counts arrays are sorted spatially using C-style ordering. That means the inner-most loop runs over the z axis, then y axis and x is the slowest varying dimension.

In the case of a single-file snapshot, the Files array is just an array of zeroes since all the particles will be in the 0-th file. Note also that in the case of a multi-files snapshot, a cell is always contained in a single file.

As noted above, particles can (slightly) drift out of their cells. This can be problematic in cases where one wants to find precisely all the particles in a given region. To help with this, the meta-data also contains a “cell bounding box”. The arrays /Cells/MinPositions/PartTypeN and /Cells/MaxPositions/PartTypeN contain the minimal (maximal) x,y,z coordinates of all the particles of this type in the cells. Note that these coordinates can be outside of the cell itself. When using periodic boundary conditions, no box-wrapping is applied.

If a snapshot used a sub-sampled output, then the counts and offsets are adjusted accordingly and correspond to the actual content of the file (i.e. after the sub-sampling was applied).

As an example, if one is interested in retriving all the densities of the gas particles in the cell around the position [1, 1, 1] in a single-file snapstshot one could use a piece of code similar to:

import numpy as np
import h5py

snapshot_file = h5py.File("snapshot.hdf5", "r")

my_pos = [1, 1, 1]

# Read in the cell centres and size
nr_cells = f["/Cells/Meta-data"].attrs["nr_cells"]
centres = f["/Cells/Centres"][:,:]
size = f["/Cells/Meta-data"].attrs["size"]
half_size = size / 2.

# Look for the cell containing the position of interest.
#
# Note that since the cells are sorted spatially, we would formally
# not need to do this search and could jump directly to the correct 'i'.
my_cell = -1
for i in range(nr_cells):
   if my_pos[0] > centres[i, 0] - half_size[0] and my_pos[0] < centres[i, 0] + half_size[0] and
      my_pos[1] > centres[i, 1] - half_size[1] and my_pos[1] < centres[i, 1] + half_size[1] and
      my_pos[2] > centres[i, 2] - half_size[2] and my_pos[2] < centres[i, 2] + half_size[2]:
      my_cell = i
      break

# Print the position of the centre of the cell of interest
centre = snapshot_file["/Cells/Centres"][my_cell, :]
print("Centre of the cell:", centre)

# Retrieve the offset and counts
my_offset = snapshot_file["/Cells/OffsetsInFile/PartType0"][my_cell]
my_count = snapshot_file["/Cells/Counts/PartType0"][my_cell]

# Get the densities of the particles in this cell
rho = snapshot_file["/PartType0/Density"][my_offset:my_offset + my_count]

For large simulations, this vastly reduces the amount of data that needs to be read from the disk.

Note that this is all automated in the swiftsimio python library and we highly encourage its use.

Meta-file for distributed snapshots

If distributed snapshots are chosen for an MPI parallel run (see Snapshots), N snapshot files are produced, where N is the number of MPI ranks. When HDF5 1.10.0 or higher is available, an additional meta-snapshot is produced that uses HDF5’s virtual dataset feature to present these N files as if they were a single, regular snapshot file.

The meta-snapshot contains all the meta-data (including the top level cell hash-tables) contained in a regular snapshot, but does not store any actual particle data. Instead, the particle datasets contain virtual links to the corresponding particle data in the distributed snapshot files. Since this is a feature of the HDF5 library itself, this is entirely transparent to modules like h5py that try to read the data. A user only needs to access the meta-snapshot, and the HDF5 library takes care of the rest.

The virtual links in the meta-snapshot only work if the HDF5 library knows the location of the distributed snapshots. These are stored within the meta-snapshot as relative paths. When SWIFT produces a distributed snapshot, all files are placed within the same directory. This means that the meta-snapshot can only be safely read if the other N files are also present in the same directory.

The header of a meta-snapshot looks exactly like the header of a normal, non-distributed snapshot (i.e. NumFilesPerSnapshot is 1). However, the attribute Virtual is set to 1 to distinguish it from a normal snapshot file.

[1]

In the rare case where an output selection (see Output Selection) disabling a given particle type in its entirety was used, the corresponding entry in NumPart_ThisFile will be 0 whilst the NumPart_Total field will still contain the number of particles present in the run.

[2]

Note that all quantities in SWIFT are always “h-free” in the sense that they are expressed in units withouy any h terms. This implies that the h-scale exponent field value is always 0. SWIFT nevertheless includes this field to be comprehensive and to prevent confusion with other software packages that express their quantities with h-full units.