[netcdf-java] performance issue

• To: "netcdf-java@xxxxxxxxxxxxxxxx" <netcdf-java@xxxxxxxxxxxxxxxx>
• Subject: [netcdf-java] performance issue
• From: Mark Balonkin <mbalonkin@xxxxxxx>
• Date: Thu, 11 Dec 2008 09:18:54 +0100

Hi,
I'm using java NetCDF 2.2.20. One of the issues which I have related to very 
low performance comparing to .Net wrapper on "C" library 3.6.1
For instance I made simple straight forward test case:

-          Create a file

-          1 dimension

-          1 variable type Float

-          Dump values as arrays of 5 elements


Java code:
             try {
             NetcdfFileWriteable ncFile = 
NetcdfFileWriteable.createNew("d:\\duck2.nc", false);
            Dimension timeDim = ncFile.addDimension(timeSeriesDimensionName, 
-1, true, true, false);
            Dimension[] dim1 = new Dimension[1];
            dim1[0] = timeDim;
          ncFile.addVariable("avar", DataType.FLOAT, dim1);
          ncFile.create();
          int c=0;
            float[] arr;
            arr=new float[5];
            arr[0]=50;
            arr[1]=40;
            arr[2]=30;
            arr[3]=20;
            arr[4]=10;
            int i, j;
            for(j=0;j<2000;j++)
            for(i=0;i<50000;i++){
                  int[] timeOrigin = new int[] {c};
                  ncFile.write("avar", timeOrigin, Array.factory(arr));
                  c+=5;

            }
            ncFile.flush();
            ncFile.close();
            }
             catch(Exception ex){

             }


And similar VB.NET code

        Dim ncid As Integer
        Dim dimid As Integer
        Dim varid As Integer

        Dim v(4) As Single
        v(1) = 10
        v(2) = 20
        v(3) = 30
        v(4) = 40
        v(0) = 50
        Dim vSize(0) As Integer
        vSize(0) = 5

        Dim res As Integer
        Dim dims(0) As Integer

        res = NetCDF.nc_create("d:\duck.nc", NetCDF.cmode.NC_WRITE, ncid)
        res = NetCDF.nc_def_dim(ncid, "big_dim", 0, dimid)
        dims(0) = dimid
        res = NetCDF.nc_def_var(ncid, "big_var", NetCDF.nc_type.NC_FLOAT, 1, 
dims, varid)
        NetCDF.nc_enddef(ncid)

        Dim idx(0) As Integer
        Dim i As Integer
        Dim j As Integer

        idx(0) = 0
        For j = 0 To 2000
            For i = 0 To 50000
                res = NetCDF.nc_put_vara_float(ncid, varid, idx, vSize, v)
                idx(0) = idx(0) + vSize(0)

            Next
        Next

***************************


Java code is approximately 40 times slower.
Apparently, making a bigger buffer (for instance arrays of 10000 instead of 5 
elements) in Java case will reduce the difference, but anyway it very 
significant.
Any suggestion how to improve it?

Thanks,
Mark Balonkin

• Follow-Ups:
  • Re: [netcdf-java] performance issue
    • From: John Caron