Mapping between the CDM and NetCDF-4 Data Models

last modified: June 2014

The CDM data model is close to, but not identical to the NetCDF-4 data model. When reading netCDF-4 files, one is interested in the mapping from netCDF-4 to CDM. This mapping is relatively stable. As of version 4.3, the CDM can write to netCDF-4 files, and one is interested in the mapping from CDM to netCDF-4. This mapping is still being developed, eg to give users some control where needed.

NetCDF-4 intentionally supports a simpler data model than HDF5, which means there are HDF5 files that cannot be converted to netCDF-4. See: http://www.unidata.ucar.edu/software/netcdf/docs/faq.html#fv15


Data Model Differences

DataTypes

From netCDF-4 to CDM

  • A netCDF-4 Compound is a CDM Structure. Both can be arbitrarily nested. The components of the Compound type are called fields in NetCDF-4 and member variables in the CDM.
  • A netCDF-4 Enum is a CDM enum1, enum2, or enum4 and references a EnumTypedef which holds the (enum, String) map.
  • A netCDF-4 Vlen is mapped to a CDM variable length Dimension.
  • A netCDF-4 Opaque type is a CDM opaque type, but the length of the data cannot be found until you read the data.
  • NetCDF-4 signed and unsigned byte, short, int, long are mapped to CDM byte, short, int, long. If unsigned, the attribute _Unsigned = "true" is added to the CDM Variable, and Variable.isUnsigned() is true, as well as Array.isUnsigned(), from any data that is read from the variable.

From CDM to netCDF-4

  • A CDM array of Opaque may have a different length for each Opaque object. May have to read to find maximum length.
  • A CDM Structure may have member Variables that use shared dimensions. NetCDF4 / HDF5 does not support shared dimensions in Compound fields. If writing CDM to netCDF4, currently we just anonymize the shared dimensions.
  • A CDM Structure member Variables may have attributes. NetCDF4 / HDF5 does not support attributes on fields in Compound fields. The CDM uses these conventions for specifying attributes on compound fields.

Type Definitions

From netCDF-4 to CDM

  • A netCDF-4 Enumeration Type becomes a CDM EnumTypedef.
  • All other netCDF-4 type definitions are repeated for each CDM variable that uses them. The attribute _Typedef = "typename" is added to the CDM Variable, where typename is the name of the netCDF-4 type.

From CDM to netCDF-4

  • A CDM EnumTypedef becomes a netCDF-4 Enumeration Type.
  • If a CDM Variable has an attribute _Typedef = "typename", then the Variables' definition is made into a netCDF-4 type.

Attributes

In CDM, an attribute type may only be a scalar or 1D array of signed byte, short, int, long, float, double, or String. A char type is mapped to a String.

From netCDF-4 to CDM

  • An attribute of compound type in netCDF-4 is flattened, by making each field a separate attribute, with name attName.fieldName in the CDM.
  • If the compound attribute is for a compound variable, and the field name of the attribute matches a field name of the variable, the attribute is added to that field instead of being flattened.
  • An attribute of enum type in netCDF-4 becomes a String type in the CDM. ???
  • An attribute of opaque type in netCDF-4 becomes a byte type in the CDM.
  • An attribute of vlen of type in netCDF-4 becomes an array of type in the CDM.
  • An attribute of an unsigned byte, short, int in netCDF-4 is promoted to a signed short, int, or long in the CDM.

From CDM to netCDF-4

  • Attributes on member variables of Structures are made into a compound attribute on the parent Structure.

Differences between netCDF-4 C and Java libraries for netCDF4 files

Unsigned types

Enum Typedefs

Attributes

Creation Order

Compound field Types


Differences between netCDF-4 C and Java libraries for HDF5 files

Fixed length Strings with anonymous dimension

Anonymous dimensions

Time datatype (HDF type 2)

 


Internal Notes

1) char arrays are interpreted as UTF-8 bytes array (Strings) when they are attributes . but data arrays are not, they are run through unsignedToShort() and cast to char. this seems like trouble.

2) nc4 allows arbitrary composition of vlen. cdm tries to map these to a variable length dimension, to get a ragged array, not part of the data type. But Arrays are rectangular, so its a difficult fit.

could define ArrayRagged which maps to C multidim arrays.

its natural to map

 int data(x,y,*) -> int(*) data(x,y)

but it doesn't generalize well to nested vlens. nc4 solution is to declare each type separately and chain them:

 int(*) type1;
 type1(*) type2;
 type2 data(x,y);

Array.isVariableLength(). IOSP might return ArrayInteger from int data(*). Needs to return ArrayObject for int data(3,*), with Array.isVariableLength() true.

int(*)     returns ArrayInt
int(3,*)   returns ArrayObject(3) with ArrayInt(*) inside
int(*,3)  returns Array(n,3), whatever n happens to be.
int(3,*,*) returns ArrayObject(3) with ArrayObject(*) inside with ArrayInt(*) inside.
int(*,3,*) returns ArrayObject(n) with ArrayObject(3) inside with ArrayInt(*) inside.
int(*,*,3) returns ArrayObject(n) with ArrayInt(*,3) inside. OR  ArrayObject(n) with ArrayObject(*) with ArrayInt(3) inside.

struct {
  int i1;
  float vf(*);
} s(3);

is like float(3,*) -> ArrayObject(3) with ArrayFloat(*), inside the ArrayStructure.
this is getting out of control

3) attributes : n4 can be user defined types, cdm: 1 dim array of primitive or String.

netcdf tst_enums {
  types:
    ubyte enum Bradys {Mike = 8, Carol = 7, Greg = 6, Marsha = 5, Peter = 4, Jan = 3, Bobby = 2, Whats-her-face = 1, Alice = 0} ;

// global attributes:
  Bradys :brady_attribute = Alice, Peter, Mike ;
}

netcdf R:/testdata/netcdf4/nc4/tst_enums.nc {
 types:
  enum Bradys { 'Alice' = 0, 'Whats-her-face' = 1, 'Bobby' = 2, 'Jan' = 3, 'Peter' = 4, 'Marsha' = 5, 'Greg' = 6, 'Carol' = 7, 'Mike' = 8};

 :brady_attribute = "Alice", "Peter", "Mike";
}