## IOHDF5

Date

Abstract

Thorn IOHDF5 provides an I/O method to output variables in HDF5 ﬁle format. It also implements checkpointing/recovery functionality using HDF5.

### 1 Purpose

Thorn IOHDF5 uses the standard I/O library HDF51 to output any type of CCTK grid variables (grid scalars, grid functions, and grid arrays of arbitrary dimension) in the HDF5 ﬁle format.

Output is done by invoking the IOHDF5 I/O method which thorn IOHDF5registers with the ﬂesh’s I/O interface at startup.

Data is written into ﬁles named "<varname>.h5". Such dataﬁles can be used for further postprocessing (eg. visualization) or fed back into Cactus via the ﬁlereader capabilities of thorn IOUtil.

### 2 IOHDF5 Parameters

Parameters to control the IOHDF5 I/O method are:

• IOHDF5::out_every (steerable)
How often to do periodic IOHDF5 output. If this parameter is set in the parameter ﬁle, it will override the setting of the shared IO::out_every parameter. The output frequency can also be set for individual variables using the out_every option in an option string appended to the IOHDF5::out_vars parameter.
• IOHDF5::out_vars (steerable)
The list of variables to output using the IOHDF5 I/O method. The variables must be given by their fully qualiﬁed variable or group name. The special keyword all requests IOHDF5 output for all variables. Multiple names must be separated by whitespaces.
An option string can be appended in curly braces to a group/variable name. Supported options are out_every (to set the output frequency for individual variables) and hyperslab options (see section 4 for details).
• IOHDF5::out_dir
The directory in which to place the IOHDF5 output ﬁles. If the directory doesn’t exist at startup it will be created.
If this parameter is set to an empty string IOHDF5 output will go to the standard output directory as speciﬁed in IO::out_dir.

### 3 Serial versus Parallel Output

According to the ouptput mode parameter settings (IO::out_mode, IO::out_unchunked,
IO::out_proc_every) of thorn IOUtil, thorn IOHDF5 will output distributed data either

• in serial into a single unchunked ﬁle
IO::out_mode      = "onefile"
IO::out_unchunked = "yes"

• in parallel, that is, into separate ﬁles containing chunks of the individual processors’ patches of the distributed array
IO::out_mode      = "proc|np"

The default is to output data in parallel, in order to get maximum I/O performance. If needed, you can recombine the resulting chunked dataﬁles into a single unchunked ﬁle using the recombiner utility program. See section 9 for information how to build the recombiner program.

### 4 Output of Hyperslab Data

By default, thorn IOHDF5 outputs multidimensional Cactus variables with their full contents resulting in maximum data output. This can be changed for individual variables by specifying a hyperslab as a subset of the data within the N-dimensional volume. Such a subset (called a hyperslab) is generally deﬁned as an orthogonal region into the multidimensional dataset, with an origin (lower left corner of the hyperslab), direction vectors (deﬁning the number of hyperslab dimensions and spanning the hyperslab within the N-dimensional grid), an extent (the length of the hyperslab in each of its dimensions), and an optional downsampling factor.

Hyperslab parameters can be set for individual variables using an option string appended to the variables’ full names in the IOHDF5::out_vars parameter.

Here is an example which outputs two 3D grid functions Grid::r and Wavetoy::phi. While the ﬁrst is output with their full contents at every 5th iteration (overriding the IOHDF5::out_every parameter for this variable), a two-dimensional hyperslab is deﬁned for the second grid function. This hyperslab deﬁnes a subvolume to output, starting with a 5 grid points oﬀset into the grid, spanning in the yz-plane, with an extent of 10 and 20 grid points in y- and z-direction respectively. For this hyperslab, only every other grid point will be output.

IOHDF5::out_every = 1
IOHDF5::out_vars  = "Grid::x{ out_every = 5 }
Wavetoy::phi{ origin     = {4 4 4}
direction  = {0 1 0
0 0 1}
extent     = {10 20}
downsample = {2 2}   }"

The hyperslab parameters which can be set in an option string are:

• origin[N]
This speciﬁes the origin of the hyperslab. It must be given as an array of integer values with $N$ elements. Each value speciﬁes the oﬀset in grid points in this dimension into the N-dimensional volume of the grid variable.
If the origin for a hyperslab is not given, if will default to 0.
• direction[N][M]
The direction vectors specify both the directions in which the hyperslab should be spanned (each vector deﬁnes one direction of the hyperslab) and its dimensionality ($=$ the total number of dimension vectors). The direction vectors must be given as a concatenated array of integer values. The direction vectors must not be a linear combination of each other or null vectors.
If the direction vectors for a hyperslab are not given, the hyperslab dimensions will default to $N$, and its directions are parallel to the underlying grid.
• extent[M]
This speciﬁes the extent of the hyperslab in each of its dimensions as a number of grid points. It must be given as an array of integer values with $M$ elements ($M$ being the number of hyperslab dimensions).
If the extent for a hyperslab is not given, it will default to the grid variable’s extent. Note that if the origin is set to a non-zero value, you should also set the hyperslab extent otherwise the default extent would possibly exceed the variable’s grid extent.
• downsample[M]
To select only every so many grid points from the hyperslab you can set the downsample option. It must be given as an array of integer values with $M$ elements ($M$ being the number of hyperslab dimensions).
If the downsample option is not given, it will default to the settings of the general downsampling parameters IO::out_downsample_[xyz] as deﬁned by thorn IOUtil.

### 5 IOHDF5 Output Restrictions

Due to the naming scheme used to build unique names for HDF5 datasets (see 7, the IOHDF5 I/O method currently has the restriction that it can output a given variable (with a speciﬁc timelevel) – or a hyperslab of it – only once per iteration.

As a workaround, you should request output in such a case by using the ﬂesh’s I/O API CCTK_OutputVarAsByMethod() routine with a diﬀerent alias name for each output. Note that this will create multiple output ﬁles for the same variable then.

### 6 Checkpointing & Recovery

Thorn IOHDF5 can also be used for creating HDF5 checkpoint ﬁles and recovering from such ﬁles later on.

Checkpoint routines are scheduled at several timebins so that you can save the current state of your simulation after the initial data phase, during evolution, or at termination. Checkpointing for thorn IOHDF5 is enabled by setting the parameter IOHDF5::checkpoint = "yes".

A recovery routine is registered with thorn IOUtil in order to restart a new simulation from a given HDF5 checkpoint. The very same recovery mechanism is used to implement a ﬁlereader functionality to feed back data into Cactus.

Checkpointing and recovery are controlled by corresponding checkpoint/recovery parameters of thorn IOUtil (for a description of these parameters please refer to this thorn’s documentation). The parameter IO::checkpoint_every_walltime_hours is not (yet) supported.

### 7 Importing External Data Into Cactus With IOHDF5

In order to import external data into Cactus (eg. to initialize some variable) you ﬁrst need to convert this data into an HDF5 dataﬁle which then can be processed by the registered recovery routine of thorn IOHDF5.

The following description explains the HDF5 ﬁle layout of an unchunked dataﬁle which thorn IOHDF5 expects in order to restore Cactus variables from it properly. There is also a well-documented example C program provided (IOHDF5/doc/CreateIOHDF5datafile.c) which illustrates how to create a dataﬁle with IOHDF5 ﬁle layout. This working example can be used as a template for building your own data converter program.

1. Actual data is stored as multidimensional datasets in an HDF5 ﬁle. There is no nested grouping structure, every dataset is located in the root group.
A dataset’s name must match the following naming pattern which guarantees to generate unique names:
"<full variable name> timelevel <timelevel> at iteration <iteration>"

IOHDF5’s recovery routine parses a dataset’s name according to this pattern to determine the Cactus variable to restore, along with its timelevel. The iteration number is just informative and not needed here.

2. The type of your data as well as its dimensions are already inherited by a dataset itself as metainformation. But this is not enough for IOHDF5 to safely match it against a speciﬁc Cactus variable. For that reason, the variable’s groupname, its grouptype, and the total number of timelevels must be attached to every dataset as attribute information.
3. Finally, the recovery routine needs to know how the dataﬁle to recover from was created:
• Does the ﬁle contain chunked or unchunked data ?
• How many processors were used to produce the data ?
• How many I/O processors were used to write the data ?
• What Cactus version is this dataﬁle compatible with ?

Such information is put into as attributes into a group named "Global Attributes". Since we assume unchunked data here the processor information isn’t relevant — unchunked data can be fed back into a Cactus simulation running on an arbitrary number of processors.
The Cactus version ID must be present to indicate that grid variables with multiple timelevels should be recovered following the new timelevel scheme (as introduced in Cactus beta 10).

The example C program goes through all of these steps and creates a dataﬁle x.h5 in IOHDF5 ﬁle layout which contains a single dataset named "grid::x timelevel 0 at iteration 0", with groupname "grid::coordinates", grouptype CCTK_GF (thus identifying the variable as a grid function), and the total number of timelevels set to 1.

The global attributes are set to "unchunked" $=$ "yes", nprocs $=$ 1, and ioproc_every $=$ 1.

Once you’ve built and ran the program you can easily verify if it worked properly with

h5dump x.h5

which lists all objects in the dataﬁle along with their values. It will also dump the contents of the 3D dataset. Since it only contains zeros it would probably not make much sense to feed this dataﬁle into Cactus for initializing your x coordinate grid function :-)

### 8 Building A Cactus Conﬁguration with IOHDF5

The Cactus distribution does not contain the HDF5 header ﬁles and library which is used by thorn IOHDF5. You have to include the thorn HDF5 (located in the Cactus ExternalLibraries arrangement). This thorn will either build its own HDF5 library, or use an installed version in some cases.

Thorn IOHDF5 inherits from IOUtil and IOHDF5Util so you need to include these thorns in your thorn list to build a conﬁguration with IOHDF5.

### 9 Utility Programs provided by IOHDF5

Thorn IOHDF5 provides the following utility programs:

• hdf5_recombiner
Recombines chunked HDF5 dataﬁle(s) into a single unchunked HDF5 dataﬁle. By applying the -single_precision command line option, double precision ﬂoating-point datasets can be converted into single precision during the recombination.
• hdf5_convert_from_ieeeio
Converts a dataﬁle created by thorn IOFlexIO into an HDF5 dataﬁle. Your thornlist must include this thorn in its thornlist in order to build the FlexIO-to-HDF5 utility program.
• hdf5_convert_from_sdf
Converts a dataﬁle created by thorn CactusIO/IOSDF or other Cactus-external programs into an HDF5 dataﬁle. Your thornlist must include this thorn in its thornlist in order to build the SDF-to-HDF5 utility program.

All utility programs are located in the src/util/ subdirectory of thorn IOHDF5. To build the utilities just do a