Example UDF (CUDA) - Sum of Squares¶

The following implements the sum-of-squares algorithm as a distributed UDF using the CUDA-based UDF C++ API. The initialization script will be run separately from the UDF itself, and the execution script will need to target the UDF C++ binary.

This example will take a list of input tables and corresponding output tables (must be the same number) and, for each record of each input table, sums the squares of all input table columns and saves the result to the first column of the corresponding output table record; i.e.:

in.a² + in.b² + ... + in.n² -> out.a

This setup assumes the UDF is being developed on the Kinetica host (or head node host, if a multi-node Kinetica cluster); and that the Python database API is available at /opt/gpudb/api/python and the C++ UDF API is available at /opt/gpudb/udf/api/cpp.

This example will contain the following (click to download):

udf_sos_cpp_init.py: creates the input & output tables and loads test data
udf_sos_cpp_proc.cu: implements the UDF itself
makefile: compiles the UDF
udf_sos_cpp_exec.py: creates & executes the UDF

Note

All commands should be run as the gpudb user.

After copying these four files to a gpudb-accessible directory on the Kinetica head node, the example can be run as follows:

$ /opt/gpudb/bin/gpudb_python udf_sos_cpp_init.py
$ make
$ /opt/gpudb/bin/gpudb_python udf_sos_cpp_exec.py

The results of the run can be checked via Kinetica Administration Application (GAdmin). There should exist two tables, udf_sos_in_table & udf_sos_out_table, each holding 10,000 records; the former containing pairs of numbers and the latter containing the sums of squares of those numbers. Each table will carry an id, which can be used to associate input values to output sums.

To verify the existence of the tables, in GAdmin, click Data > Tables. Both tables should appear in the listing, each with 10,000 records.

To verify the calculations, click Query > KiSQL. Enter the following query into the SQL Statement box:

SELECT
   sos_in.id,
   STRING(x1) || '^2 + ' || STRING(x2) || '^2 = ' || STRING(y) AS "Equation"
FROM
   udf_sos_in_table sos_in,
   udf_sos_out_table sos_out
WHERE
   sos_in.id = sos_out.id
ORDER BY
   id;

The Query Result box should show each of the 10,000 calculations made.

Execution Detail¶

While the example UDF itself can run against multiple tables, the example run will use a single table, udf_sos_in_table, as input and a matching table, udf_sos_out_table, for output.

The input table will contain two float columns and be populated with 10,000 pairs of randomly-generated numbers. The output table will contain one float column that will hold the sums calculated by the UDF. Both tables will also contain an int column that is the calculation identifier, allowing the input data to be matched up with the output data after the UDF has run.

Note

The UDF will assume the first column of the input table, as defined in the original table creation process, is the identifier field. All of the remaining columns after the first will be used in the sum-of-squares calculation.

The UDF will calculate the sum of the squares of each of the 10,000 pairs of numbers and insert into the output table the corresponding 10,000 sums.

udf_sos_cpp_init.py¶

This initialization script creates the input & output tables and populates the input data using the standard Kinetica Python API, all outside of the UDF execution framework.

Several aspects of the initialization process are noteworthy:

The external database connection, indicative of the use of the standard Kinetica Python API--the UDF will not have this, as it runs within the database:

h_db = GPUdb(encoding = 'BINARY', host = KINETICA_HOST, port = KINETICA_PORT)

Input & output table creation:

columns = []
columns.append(GPUdbRecordColumn("id", GPUdbRecordColumn._ColumnType.INT, [GPUdbColumnProperty.PRIMARY_KEY, GPUdbColumnProperty.INT16]))
columns.append(GPUdbRecordColumn("x1", GPUdbRecordColumn._ColumnType.FLOAT))
columns.append(GPUdbRecordColumn("x2", GPUdbRecordColumn._ColumnType.FLOAT))

input_table = GPUdbTable(columns, INPUT_TABLE, db = h_db)

columns = []
columns.append(GPUdbRecordColumn("id", GPUdbRecordColumn._ColumnType.INT, [GPUdbColumnProperty.PRIMARY_KEY, GPUdbColumnProperty.INT16]))
columns.append(GPUdbRecordColumn("y", GPUdbRecordColumn._ColumnType.FLOAT))

GPUdbTable(columns, OUTPUT_TABLE, db = h_db)

udf_sos_cpp_proc.cu¶

This is the UDF itself. It uses the Kinetica C++ UDF API to compute the sums of squares of input table columns and output those sums to an output table. It runs within the UDF execution framework, and as such, is not called directly--instead, it is registered and launched by udf_sos_cpp_exec.py.

Noteworthy in the UDF are the following:

The initial call to ProcData() to access the database:

        kinetica::ProcData* procData = kinetica::ProcData::get();

The size of the output table must be specified before writing to it:

            size_t recordCount = inputTable.getSize();

            outputTable.setSize(recordCount);

The final call to complete() to mark the process as finished and ready for clean-up:

        procData->complete();

makefile¶

This standard makefile is used to compile the C++ UDF source before registering the UDF with udf_sos_cpp_exec.py.

Noteworthy in the makefile are the following:

The assumption of the C++ UDF library installed at the default location of a typical Kinetica deployment, and the location of a local CUDA install:

UDF_LIB=/opt/gpudb/udf/api/cpp/kinetica
CUDADIR := /usr/local/cuda

The inclusion of Proc.cpp & Proc.hpp as the only UDF-centric compilation dependence:

${TARGET}: makefile ${TARGET}.cu ${UDF_LIB}/Proc.cpp ${UDF_LIB}/Proc.hpp
	nvcc -o ${TARGET} ${TARGET}.cu ${UDF_LIB}/Proc.cpp -I${UDF_LIB} -m64

udf_sos_cpp_exec.py¶

The execution script uses the standard Kinetica Python API to register the UDF in the database and then execute it.

The registration step associates a name with the UDF execution code compiled from udf_sos_cpp_proc.cu, the command (the name of the compiled executable, referenced locally) to use to run it, and that it will run in distributed mode.

response = h_db.create_proc(proc_name, 'distributed', files, './' + file_name, [], {})

The execution step invokes the UDF by name, passing in the input & output table names against which the UDF will execute.

response = h_db.execute_proc(proc_name, {}, {}, [INPUT_TABLE], {}, [OUTPUT_TABLE], {})

Table Of Contents