 |
NetCDF
4.9.2
|
|
NetCDF
4.9.2
|
To present the format more formally, we use a BNF grammar notation. In this notation:
Following the grammar, a few additional notes are included to specify format characteristics that are impractical to capture in a BNF grammar, and to note some special cases for implementers. Comments in the grammar point to the notes and special cases, and help to clarify the intent of elements of the format.
netcdf_file = header data
header = magic numrecs dim_list gatt_list var_list
magic = 'C' 'D' 'F' VERSION
VERSION = \x01 | // classic format
\x02 // 64-bit offset format
numrecs = NON_NEG | STREAMING // length of record dimension
dim_list = ABSENT | NC_DIMENSION nelems [dim ...]
gatt_list = att_list // global attributes
att_list = ABSENT | NC_ATTRIBUTE nelems [attr ...]
var_list = ABSENT | NC_VARIABLE nelems [var ...]
ABSENT = ZERO ZERO // Means list is not present
ZERO = \x00 \x00 \x00 \x00 // 32-bit zero
NC_DIMENSION = \x00 \x00 \x00 \x0A // tag for list of dimensions
NC_VARIABLE = \x00 \x00 \x00 \x0B // tag for list of variables
NC_ATTRIBUTE = \x00 \x00 \x00 \x0C // tag for list of attributes
nelems = NON_NEG // number of elements in following sequence
dim = name dim_length
name = nelems namestring
// Names a dimension, variable, or attribute.
// Names should match the regular expression
// ([a-zA-Z0-9_]|{MUTF8})([^\x00-\x1F/\x7F-\xFF]|{MUTF8})*
// For other constraints, see "Note on names", below.
namestring = ID1 [IDN ...] padding
ID1 = alphanumeric | '_'
IDN = alphanumeric | special1 | special2
alphanumeric = lowercase | uppercase | numeric | MUTF8
lowercase = 'a'|'b'|'c'|'d'|'e'|'f'|'g'|'h'|'i'|'j'|'k'|'l'|'m'|
'n'|'o'|'p'|'q'|'r'|'s'|'t'|'u'|'v'|'w'|'x'|'y'|'z'
uppercase = 'A'|'B'|'C'|'D'|'E'|'F'|'G'|'H'|'I'|'J'|'K'|'L'|'M'|
'N'|'O'|'P'|'Q'|'R'|'S'|'T'|'U'|'V'|'W'|'X'|'Y'|'Z'
numeric = '0'|'1'|'2'|'3'|'4'|'5'|'6'|'7'|'8'|'9'
// special1 chars have traditionally been
// permitted in netCDF names.
special1 = '_'|'.'|'@'|'+'|'-'
// special2 chars are recently permitted in
// names (and require escaping in CDL).
// Note: '/' is not permitted.
special2 = ' ' | '!' | '"' | '#' | '$' | '%' | '&' | '\'' |
'(' | ')' | '*' | ',' | ':' | ';' | '<' | '=' |
'>' | '?' | '[' | '\' | ']' | '^' | '‘’ | '{' |
'|' | '}' | '~'
MUTF8 = <multibyte UTF-8 encoded, NFC-normalized Unicode character>
dim_length = NON_NEG // If zero, this is the record dimension.
// There can be at most one record dimension.
attr = name nc_type nelems [values ...]
nc_type = NC_BYTE |
NC_CHAR |
NC_SHORT |
NC_INT |
NC_FLOAT |
NC_DOUBLE
var = name nelems [dimid ...] vatt_list nc_type vsize begin
// nelems is the dimensionality (rank) of the
// variable: 0 for scalar, 1 for vector, 2
// for matrix, ...
dimid = NON_NEG // Dimension ID (index into dim_list) for
// variable shape. We say this is a "record
// variable" if and only if the first
// dimension is the record dimension.
vatt_list = att_list // Variable-specific attributes
vsize = NON_NEG // Variable size. If not a record variable,
// the amount of space in bytes allocated to
// the variable's data. If a record variable,
// the amount of space per record. See "Note
// on vsize", below.
begin = OFFSET // Variable start location. The offset in
// bytes (seek index) in the file of the
// beginning of data for this variable.
data = non_recs recs
non_recs = [vardata ...] // The data for all non-record variables,
// stored contiguously for each variable, in
// the same order the variables occur in the
// header.
vardata = [values ...] // All data for a non-record variable, as a
// block of values of the same type as the
// variable, in row-major order (last
// dimension varying fastest).
recs = [record ...] // The data for all record variables are
// stored interleaved at the end of the
// file.
record = [varslab ...] // Each record consists of the n-th slab
// from each record variable, for example
// x[n,...], y[n,...], z[n,...] where the
// first index is the record number, which
// is the unlimited dimension index.
varslab = [values ...] // One record of data for a variable, a
// block of values all of the same type as
// the variable in row-major order (last
// index varying fastest).
values = bytes | chars | shorts | ints | floats | doubles
string = nelems [chars]
bytes = [BYTE ...] padding
chars = [CHAR ...] padding
shorts = [SHORT ...] padding
ints = [INT ...]
floats = [FLOAT ...]
doubles = [DOUBLE ...]
padding = <0, 1, 2, or 3 bytes to next 4-byte boundary>
// Header padding uses null (\x00) bytes. In
// data, padding uses variable's fill value.
// See "Note on padding", below, for a special
// case.
NON_NEG = <non-negative INT>
STREAMING = \xFF \xFF \xFF \xFF // Indicates indeterminate record
// count, allows streaming data
OFFSET = <non-negative INT> | // For classic format or
<non-negative INT64> // for 64-bit offset format
BYTE = <8-bit byte> // See "Note on byte data", below.
CHAR = <8-bit byte> // See "Note on char data", below.
SHORT = <16-bit signed integer, Bigendian, two's complement>
INT = <32-bit signed integer, Bigendian, two's complement>
INT64 = <64-bit signed integer, Bigendian, two's complement>
FLOAT = <32-bit IEEE single-precision float, Bigendian>
DOUBLE = <64-bit IEEE double-precision float, Bigendian>
// following type tags are 32-bit integers
NC_BYTE = \x00 \x00 \x00 \x01 // 8-bit signed integers
NC_CHAR = \x00 \x00 \x00 \x02 // text characters
NC_SHORT = \x00 \x00 \x00 \x03 // 16-bit signed integers
NC_INT = \x00 \x00 \x00 \x04 // 32-bit signed integers
NC_FLOAT = \x00 \x00 \x00 \x05 // IEEE single precision floats
NC_DOUBLE = \x00 \x00 \x00 \x06 // IEEE double precision floats
// Default fill values for each type, may be
// overridden by variable attribute named
// '_FillValue'. See "Note on fill values",
// below.
FILL_CHAR = \x00 // null byte
FILL_BYTE = \x81 // (signed char) -127
FILL_SHORT = \x80 \x01 // (short) -32767
FILL_INT = \x80 \x00 \x00 \x01 // (int) -2147483647
FILL_FLOAT = \x7C \xF0 \x00 \x00 // (float) 9.9692099683868690e+36
FILL_DOUBLE = \x47 \x9E \x00 \x00 \x00 \x00 \x00 \x00 //(double)9.9692099683868690e+36
Note on vsize: This number is the product of the dimension lengths (omitting the record dimension) and the number of bytes per value (determined from the type), increased to the next multiple of 4, for each variable. If a record variable, this is the amount of space per record (except that, for backward compatibility, it always includes padding to the next multiple of 4 bytes, even in the exceptional case noted below under “Note on padding”). The netCDF “record size” is calculated as the sum of the vsize's of all the record variables.
The vsize field is actually redundant, because its value may be computed from other information in the header. The 32-bit vsize field is not large enough to contain the size of variables that require more than 2^32 - 4 bytes, so 2^32 - 1 is used in the vsize field for such variables.
Note on names: Earlier versions of the netCDF C-library reference implementation enforced a more restricted set of characters in creating new names, but permitted reading names containing arbitrary bytes. This specification extends the permitted characters in names to include multi-byte UTF-8 encoded Unicode and additional printing characters from the US-ASCII alphabet. The first character of a name must be alphanumeric, a multi-byte UTF-8 character, or '_' (reserved for special names with meaning to implementations, such as the “_FillValue” attribute). Subsequent characters may also include printing special characters, except for '/' which is not allowed in names. Names that have trailing space characters are also not permitted.
Implementations of the netCDF classic and 64-bit offset format must ensure that names are normalized according to Unicode NFC normalization rules during encoding as UTF-8 for storing in the file header. This is necessary to ensure that gratuitous differences in the representation of Unicode names do not cause anomalies in comparing files and querying data objects by name.
Note on streaming data: The largest possible record count, 2^32 - 1, is reserved to indicate an indeterminate number of records. This means that the number of records in the file must be determined by other means, such as reading them or computing the current number of records from the file length and other information in the header. It also means that the numrecs field in the header will not be updated as records are added to the file. [This feature is not yet implemented].
Note on padding: In the special case when there is only one record variable and it is of type character, byte, or short, no padding is used between record slabs, so records after the first record do not necessarily start on four-byte boundaries. However, as noted above under “Note on vsize”, the vsize field is computed to include padding to the next multiple of 4 bytes. In this case, readers should ignore vsize and assume no padding. Writers should store vsize as if padding were included.
Note on byte data: It is possible to interpret byte data as either signed (-128 to 127) or unsigned (0 to 255). When reading byte data through an interface that converts it into another numeric type, the default interpretation is signed. There are various attribute conventions for specifying whether bytes represent signed or unsigned data, but no standard convention has been established. The variable attribute “_Unsigned” is reserved for this purpose in future implementations.
Note on char data: Although the characters used in netCDF names must be encoded as UTF-8, character data may use other encodings. The variable attribute “_Encoding” is reserved for this purpose in future implementations.
Note on fill values: Because data variables may be created before their values are written, and because values need not be written sequentially in a netCDF file, default “fill values” are defined for each type, for initializing data values before they are explicitly written. This makes it possible to detect reading values that were never written. The variable attribute “_FillValue”, if present, overrides the default fill value for a variable. If _FillValue is defined then it should be scalar and of the same type as the variable.
Fill values are not required, however, because netCDF libraries have traditionally supported a “no fill” mode when writing, omitting the initialization of variable values with fill values. This makes the creation of large files faster, but also eliminates the possibility of detecting the inadvertent reading of values that haven't been written.
The offset (position within the file) of a specified data value in a classic format or 64-bit offset data file is completely determined by the variable start location (the offset in the begin field), the external type of the variable (the nc_type field), and the dimension indices (one for each of the variable's dimensions) of the value desired.
The external size in bytes of one data value for each possible netCDF type, denoted extsize below, is:
The record size, denoted by recsize below, is the sum of the vsize fields of record variables (variables that use the unlimited dimension), using the actual value determined by dimension sizes and variable type in case the vsize field is too small for the variable size.
To compute the offset of a value relative to the beginning of a variable, it is helpful to precompute a “product vector” from the dimension lengths. Form the products of the dimension lengths for the variable from right to left, skipping the leftmost (record) dimension for record variables, and storing the results as the product vector for each variable.
For example:
At this point, the leftmost product, when rounded up to the next multiple of 4, is the variable size, vsize, in the grammar above. For example, in the non-record variable above, the value of the vsize field is 212 (210 rounded up to a multiple of 4). For the record variable, the value of vsize is just 72, since this is already a multiple of 4.
Let coord be the array of coordinates (dimension indices, zero-based) of the desired data value. Then the offset of the value from the beginning of the file is just the file offset of the first data value of the desired variable (its begin field) added to the inner product of the coord and product vectors times the size, in bytes, of each datum for the variable. Finally, if the variable is a record variable, the product of the record number, 'coord[0]', and the record size, recsize, is added to yield the final offset value.
A special case: Where there is exactly one record variable, we drop the requirement that each record be four-byte aligned, so in this case there is no record padding.
By using the grammar above, we can derive the smallest valid netCDF file, having no dimensions, no variables, no attributes, and hence, no data. A CDL representation of the empty netCDF file is
This empty netCDF file has 32 bytes. It begins with the four-byte “magic number” that identifies it as a netCDF version 1 file: ‘C’, ‘D’, ‘F’, ‘\x01’. Following are seven 32-bit integer zeros representing the number of records, an empty list of dimensions, an empty list of global attributes, and an empty list of variables.
Below is an (edited) dump of the file produced using the Unix command
Each 16-byte portion of the file is displayed with 4 lines. The first line displays the bytes in hexadecimal. The second line displays the bytes as characters. The third line displays each group of two bytes interpreted as a signed 16-bit integer. The fourth line (added by human) presents the interpretation of the bytes in terms of netCDF components and values.
As a less trivial example, consider the CDL
which corresponds to a 92-byte netCDF file. The following is an edited dump of this file:
The netCDF 64-bit offset format differs from the classic format only in the VERSION byte, ‘\x02’ instead of ‘\x01’, and the OFFSET entity, a 64-bit instead of a 32-bit offset from the beginning of the file. This small format change permits much larger files, but there are still some practical size restrictions. Each fixed-size variable and the data for one record's worth of each record variable are still limited in size to a little less that 4 GiB. The rationale for this limitation is to permit aggregate access to all the data in a netCDF variable (or a record's worth of data) on 32-bit platforms.
The netCDF-4 format implements and expands the netCDF-3 data model by using an enhanced version of HDF5 as the storage layer. Use is made of features that are only available in HDF5 version 1.8 and later.
Using HDF5 as the underlying storage layer, netCDF-4 files remove many of the restrictions for classic and 64-bit offset files. The richer enhanced model supports user-defined types and data structures, hierarchical scoping of names using groups, additional primitive types including strings, larger variable sizes, and multiple unlimited dimensions. The underlying HDF5 storage layer also supports per-variable compression, multidimensional tiling, and efficient dynamic schema changes, so that data need not be copied when adding new variables to the file schema.
Creating a netCDF-4/HDF5 file with netCDF-4 results in an HDF5 file. The features of netCDF-4 are a subset of the features of HDF5, so the resulting file can be used by any existing HDF5 application.
Although every file in netCDF-4 format is an HDF5 file, there are HDF5 files that are not netCDF-4 format files, because the netCDF-4 format intentionally uses a limited subset of the HDF5 data model and file format features. Some HDF5 features not supported in the netCDF enhanced model and netCDF-4 format include non-hierarchical group structures, HDF5 reference types, multiple links to a data object, user-defined atomic data types, stored property lists, more permissive rules for data object names, the HDF5 date/time type, and attributes associated with user-defined types.
A complete specification of HDF5 files is beyond the scope of this document. For more information about HDF5, see the HDF5 web site: http://hdf.ncsa.uiuc.edu/HDF5/.
The specification that follows is sufficient to allow HDF5 users to create files that will be accessible from netCDF-4.
The netCDF API maintains the creation order of objects that are created in the file. The same is not true in HDF5, which maintains the objects in alphabetical order. Starting in version 1.8 of HDF5, the ability to maintain creation order was added. This must be explicitly turned on in the HDF5 data file in several ways.
Each group must have link and attribute creation order set. The following code (from libsrc4/nc4hdf.c) shows how the netCDF-4 library sets these when creating a group.
Each dataset in the HDF5 file must be created with a property list for which the attribute creation order has been set to creation ordering. The H5Pset_attr_creation_order function is used to set the creation ordering of attributes of a variable.
The following example code (from libsrc4/nc4hdf.c) shows how the creation ordering is turned on by the netCDF library.
NetCDF-4 groups are the same as HDF5 groups, but groups in a netCDF-4 file must be strictly hierarchical. In general, HDF5 permits non-hierarchical structuring of groups (for example, a group that is its own grandparent). These non-hierarchical relationships are not allowed in netCDF-4 files.
In the netCDF API, the global attribute becomes a group-level attribute. That is, each group may have its own global attributes.
The root group of a file is named “/” in the netCDF API, where names of groups are used. It should be noted that the netCDF API (like the HDF5 API) makes little use of names, and refers to entities by number.
Until version 1.8, HDF5 did not have any capability to represent shared dimensions. With the 1.8 release, HDF5 introduced the dimension scale feature to allow shared dimensions in HDF5 files.
The dimension scale is unfortunately not exactly equivalent to the netCDF shared dimension, and this leads to a number of compromises in the design of netCDF-4.
A netCDF shared dimension consists solely of a length and a name. An HDF5 dimension scale also includes values for each point along the dimension, information that is (optionally) included in a netCDF coordinate variable.
To handle the case of a netCDF dimension without a coordinate variable, netCDF-4 creates dimension scales of type char, and leaves the contents of the dimension scale empty. Only the name and length of the scale are significant. To distinguish this case, netCDF-4 takes advantage of the NAME attribute of the dimension scale. (Not to be confused with the name of the scale itself.) In the case of dimensions without coordinate data, the HDF5 dimension scale NAME attribute is set to the string: "This is a netCDF dimension but not a netCDF variable."
In the case where a coordinate variable is defined for a dimension, the HDF5 dimscale matches the type of the netCDF coordinate variable, and contains the coordinate data.
A further difficulty arrises when an n-dimensional coordinate variable is defined, where n is greater than one. NetCDF allows such coordinate variables, but the HDF5 model does not allow dimension scales to be attached to other dimension scales, making it impossible to completely represent the multi-dimensional coordinate variables of the netCDF model.
To capture this information, multidimensional coordinate variables have an attribute named _Netcdf4Coordinates. The attribute is an array of H5T_NATIVE_INT, with the netCDF dimension IDs of each of its dimensions.
The _Netcdf4Coordinates attribute is otherwise hidden by the netCDF API. It does not appear as one of the attributes for the netCDF variable involved, except through the HDF5 API.
Starting with the netCDF-4.1 release, netCDF can read HDF5 files which do not use dimension scales. In this case the netCDF library assigns dimensions to the HDF5 dataset as needed, based on the length of the dimension.
When an HDF5 file is opened, each dataset is examined in turn. The lengths of all the dimensions involved in the shape of the dataset are determined. Each new (i.e. previously unencountered) length results in the creation of a phony dimension in the netCDF API.
This will not accurately detect a shared, unlimited dimension in the HDF5 file, if different datasets have different lengths along this dimension (possible in HDF5, but not in netCDF).
Note that this is a read-only capability for the netCDF library. When the netCDF library writes HDF5 files, they always use a dimension scale for every dimension.
Datasets must have either dimension scales for every dimension, or no dimension scales at all. Partial dimension scales are not, at this time, understood by the netCDF library.
In order to preserve creation order, the netCDF-4 library writes variables in their creation order. Since some variables are also dimension scales, their order reflects both the order of the dimensions and the order of the coordinate variables.
However, these may be different. Consider the following code:
In this case the order of the coordinate variables will be different from the order of the dimensions.
In practice, this should make little difference in user code, but if the user is writing code that depends on the ordering of dimensions, the netCDF library was updated in version 4.1 to detect this condition, and add the attribute _Netcdf4Dimid to the dimension scales in the HDF5 file. This attribute holds a scalar H5T_NATIVE_INT which is the (zero-based) dimension ID for this dimension.
If this attribute is present on any dimension scale, it must be present on all dimension scales in the file.
Variables in netCDF-4/HDF5 files exactly correspond to HDF5 datasets. The data types match naturally between netCDF and HDF5.
In netCDF classic format, the problem of endianness is solved by writing all data in big-endian order. The HDF5 library allows data to be written as either big or little endian, and automatically reorders the data when it is read, if necessary.
By default, netCDF uses the native types on the machine which writes the data. Users may change the endianness of a variable (before any data are written). In that case the specified endian type will be used in HDF5 (for example, a H5T_STD_I16LE will be used for NC_SHORT, if little-endian has been specified for that variable.)
The NC_CHAR type represents a single character, and the NC_STRING an array of characters. This can be confusing because a one-dimensional array of NC_CHAR is used to represent a string (i.e. a scalar NC_STRING).
An odd case may arise in which the user defines a variable with the same name as a dimension, but which is not intended to be the coordinate variable for that dimension. In this case the string "_nc4_non_coord_" is pre-pended to the name of the HDF5 dataset, and stripped from the name for the netCDF API.
Attributes in HDF5 and netCDF-4 correspond very closely. Each attribute in an HDF5 file is represented as an attribute in the netCDF-4 file, with the exception of the attributes below, which are hidden by the netCDF-4 API.
Each user-defined data type in an HDF5 file exactly corresponds to a user-defined data type in the netCDF-4 file. Only base data types which correspond to netCDF-4 data types may be used. (For example, no HDF5 reference data types may be used.)
The HDF5 library provides data compression using the zlib library and the szlib library. NetCDF-4 only allows users to create data with the zlib library (due to licensing restrictions on the szlib library). Since HDF5 supports the transparent reading of the data with either compression filter, the netCDF-4 library can read data compressed with szlib (if the underlying HDF5 library is built to support szlib), but has no way to write data with szlib compression.
With zlib compression (a.k.a. deflation) the user may set a deflation factor from 0 to 9. In our measurements the zero deflation level does not compress the data, but does incur the performance penalty of compressing the data. The netCDF API does not allow the user to write a variable with zlib deflation of 0 - when asked to do so, it turns off deflation for the variable instead. NetCDF can read an HDF5 file with deflation of zero, and correctly report that to the user.
Every classic and 64-bit offset file can be represented as a netCDF-4 file, with no loss of information. There are some significant benefits to using the simpler netCDF classic model with the netCDF-4 file format. For example, software that writes or reads classic model data can write or read netCDF-4 classic model format data by recompiling/relinking to a netCDF-4 API library, with no or only trivial changes needed to the program source code. The netCDF-4 classic model format supports this usage by enforcing rules on what functions may be called to store data in the file, to make sure its data can be read by older netCDF applications (when relinked to a netCDF-4 library).
Writing data in this format prevents use of enhanced model features such as groups, added primitive types not available in the classic model, and user-defined types. However performance features of the netCDF-4 formats that do not require additional features of the enhanced model, such as per-variable compression and chunking, efficient dynamic schema changes, and larger variable size limits, offer potentially significant performance improvements to readers of data stored in this format, without requiring program changes.
When a file is created via the netCDF API with a CLASSIC_MODEL mode flag, the library creates an attribute (_nc3_strict) in the root group. This attribute is hidden by the netCDF API, but is read when the file is later opened, and used to ensure that no enhanced model features are written to the file.
Starting with version 4.1, the netCDF libraries can read HDF4 SD (Scientific Dataset) files. Access is limited to those HDF4 files created with the Scientific Dataset API. Access is read-only.
Dataset types are translated between HDF4 and netCDF in a straightforward manner.
Generated on Tue Mar 14 2023 15:55:08 for NetCDF. NetCDF is
a Unidata library.