Scalable tuning engine

• US 9,785,951 B1
• Filed: 02/28/2006
• Issued: 10/10/2017
• Est. Priority Date: 02/28/2006
• Status: Active Grant

• Current Assignee: Acoustic, L.P.
• Original Assignee: International Business Machines Corporation
• Inventors: Coleman, Alan, Kneubuhl, Joshua
• Primary Examiner(s): Li, Sun
• Application Number:
  US11/364,966
• Time in Patent Office: 4,242 Days
• Field of Search: 705 1- 30, 705 144, 705 1471, 705400, 705 729, 705 711, 705 731, 709 1- 20, 707 1- 20, 395800, 712225-273
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising
• CPC Class Codes: G06Q 30/0206 Price or cost determination based on market factors

Abstract

A computer implemented method for processing data is provided. At least one dataflow comprising transformational and numerical steps is defined. The flow is decomposed into distinct executable segments along process domains. The flow is decomposed into distinct executable segments along data domains. Parallel execution paths are identified across the executable segments. The executable segments are executed across a plurality of execution units.

Claims

1. A computer implemented method for processing sales data on a network of a plurality of computers, comprising:

receiving data including point of sale data, cost data and product data;

partitioning the received data into data portions each residing in a different section of a data source;

defining a dataflow comprising transformational and numerical steps on a first computer of the plurality of computers;

decomposing, using the first computer, the dataflow along process domains on the first computer by decomposing the dataflow into one or more from a group of distinct executable segments for parallel execution accepting inputs that are the same and distinct executable segments for parallel execution accepting inputs lacking dependencies between each other, wherein the distinct executable segments along process domains include more than one econometric operation and further wherein one process domain includes a modeling segment for generation of a demand model;

decomposing, using the first computer, the dataflow along data domains on the first computer by decomposing the dataflow into distinct executable segments for parallel execution based on dependencies between records within the data indicated by identifying operators and dividing the distinct executable segments along the data domains by demand groups, and wherein demand groups are groupings of highly substitutable products;

executing the distinct executable segments in parallel on a second computer of the plurality of computers and a third computer of the plurality of computers, wherein executing the distinct executable segments comprises;

interpreting a script with programming language statements and generating from an interpretation of the script an executable graph of the dataflow comprising the transformational and numerical steps to process the received data;

executing the executable graph via a graph execution engine that distributes the distinct executable segments among the second computer and the third computer;

reading, in a non-sequential manner during a single reading operation, the data portions from the different sections of the data source to corresponding ones of a plurality of data buffers in parallel to increase a speed of data access, wherein each data buffer corresponds to a distinct executable segment;

monitoring an amount of data in each of the data buffers to determine when a data buffer becomes filled for processing by the corresponding distinct executable segment, wherein at least two of the data buffers become filled and are accessed by the corresponding distinct executable segments at different times;

executing the distinct executable segments on the second computer and the third computer in parallel, wherein each distinct executable segment is responsive to the monitoring and retrieves and processes data from a corresponding data buffer independent of other distinct executable segments when that corresponding data buffer is filled; and

receiving the processed data from each of the distinct executable segments in parallel.

2. A computer implemented method for processing data, useful in association with a profit optimization system, the data processing method comprising:

receiving data including point of sale data, cost data and product data;

partitioning the received data into data portions each residing in a different section of a data source;

defining a dataflow comprising transformational and numerical steps;

decomposing, using a computer, the dataflow along process domains by decomposing the dataflow into one or more from a group of distinct executable segments for parallel execution accepting inputs that are the same and distinct executable segments for parallel execution accepting inputs lacking dependencies between each other, wherein the distinct executable segments along process domains include more than one econometric operation, and further wherein one process domain includes a modeling segment for generation of a demand model;

decomposing, using the computer, the dataflow along data domains by decomposing the dataflow into distinct executable segments for parallel execution based on dependencies between records within the data indicated by identifying operators and dividing the distinct executable segments along the data domains by demand groups, and wherein demand groups are groupings of highly substitutable products;

executing the distinct executable segments across a plurality of execution units in parallel, wherein executing the distinct executable segments comprises;

interpreting a script with programming language statements and generating from an interpretation of the script an executable graph of the dataflow comprising the transformational and numerical steps to process the received data;

executing the executable graph via a graph execution engine that distributes the distinct executable segments among the plurality of execution units;

reading, in a non-sequential manner during a single reading operation, the data portions from the different sections of the data source to corresponding ones of a plurality of data buffers in parallel to increase a speed of data access, wherein each data buffer corresponds to a distinct executable segment;

monitoring an amount of data in each of the data buffers to determine when a data buffer becomes filled for processing by the corresponding distinct executable segment, wherein at least two of the data buffers become filled and are accessed by the corresponding distinct executable segments at different times;

executing the distinct executable segments on the plurality of execution units in parallel, wherein each distinct executable segment is responsive to the monitoring and retrieves and processes data from a corresponding data buffer independent of other distinct executable segments when that corresponding data buffer is filled; and

receiving the processed data from each of the distinct executable segments in parallel.

3. The computer implemented method, as recited in claim 2, wherein the execution units are threads in the computer.

4. The computer implemented method, as recited in claim 2, wherein the execution units are computers in a network.

5. The computer implemented method, as recited in claim 4, further comprising, for each executable segment:

defining a dataflow comprising transformational and numerical steps for the executable segment;

decomposing the dataflow of the executable segment into distinct executable sub-segments along process domains, wherein the distinct executable sub-segments along process domains include more than one econometric operation, and further wherein one process domain includes a modeling sub-segment for generation of a demand model;

decomposing the dataflow of the executable segment into distinct executable sub-segments along data domains, wherein the distinct executable sub-segments along data domains are divided by demand groups, and wherein demand groups are groupings of highly substitutable products;

identifying parallel execution paths across the executable sub-segments;

executing the executable sub-segments across a plurality of threads.

6. The computer implemented method, as recited in claim 5, wherein a distinct executable segment of the distinct executable segments comprises a matrix operation.

7. The computer implemented method, as recited in claim 6, wherein the matrix operation operates on a matrix in a scratch storage.

8. The computer implemented method, as recited in claim 5, wherein on a first thread of the plurality of threads an executable sub-segment for a transformation operation is being performed on a matrix in scratch data in parallel with an executable sub-segment on a second thread of the plurality of threads executing an executable sub-segment of a mathematical operation on scratch data.

9. The computer implemented method, as recited in claim 5, further comprising reading data from an initial input location with a parallel reader that reads from the initial input location in a nonsequential manner and providing the data to a plurality of executable sub-segments along data domains through scratch storage.

10. The computer implemented method, as recited in claim 9, wherein the sub-segments along the process domains are state machines associated with scratch storage, and which automatically execute when data is placed in associated scratch storage.

11. The computer implemented method, as recited in claim 10, further comprising writing data from scratch storage to a final output location using a parallel writer, and wherein the final output location is a shared network storage location.

12. The computer implemented method, as recited in claim 2, further comprising reading data from an initial input location in a nonsequential manner to provide a plurality of distinct executable segments along data domains.

13. The computer implemented method, as recited in claim 12, wherein the segments along the process domains are state machines associated with scratch storage, and which automatically execute when data is placed in associated scratch storage.

14. The computer implemented method, as recited in claim 13, further comprising writing data from scratch storage to a final output location using a parallel writer.

15. The computer implemented method, as recited in claim 2, wherein the segments along the process domains are state machines associated with scratch storage, and which automatically execute when data is placed in associated scratch storage.

16. The computer implemented method, as recited in claim 15, further comprising writing data from scratch storage to a final output location using a parallel writer.

17. An apparatus for processing data, useful in association with a profit optimization system, the data processing apparatus comprising:

a first computer of a plurality of computers, the first computer configured to;

receive data including point of sale data, cost data and product data;

partition the received data into data portions each residing in a different section of a data source;

define a dataflow comprising transformational and numerical steps;

decompose the dataflow along process domains by decomposing the dataflow into one or more from a group of distinct executable segments for parallel execution accepting inputs that are the same and distinct executable segments for parallel execution accepting inputs lacking dependencies between each other, wherein the distinct executable segments along process domains include more than one econometric operation, and further wherein one process domain includes a modeling segment for generation of a demand model;

decompose the dataflow along data domains by decomposing the dataflow into distinct executable segments for parallel execution based on dependencies between records within the data indicated by identifying operators and dividing the distinct executable segments along the data domains by demand groups, and wherein demand groups are groupings of highly substitutable products;

a second computer of the plurality of computers and a third computer of the plurality of computers, wherein the second computer and third computer are configured to execute the distinct executable segments in parallel, wherein the plurality of computers are configured to;

interpret a script with programming language statements and generate from an interpretation of the script an executable graph of the dataflow comprising the transformational and numerical steps to process the received data;

execute the executable graph via a graph execution engine that distributes the distinct executable segments among the second computer and the third computer;

read, in a non-sequential manner during a single reading operation, the data portions from the different sections of the data source to corresponding ones of a plurality of data buffers in parallel to increase a speed of data access, wherein each data buffer corresponds to a distinct executable segment;

monitor an amount of data in each of the data buffers to determine when a data buffer becomes filled for processing by the corresponding distinct executable segment, wherein at least two of the data buffers become filled and are accessed by the corresponding distinct executable segments at different times;

execute the distinct executable segments on the second computer and the third computer in parallel, wherein each distinct executable segment is responsive to the monitoring and retrieves and processes data from a corresponding data buffer independent of other distinct executable segments when that corresponding data buffer is filled; and

receive the processed data from each of the distinct executable segments in parallel.

18. The apparatus, as recited in claim 17, wherein threads execute the distinct executable segments in parallel.

19. The apparatus, as recited in claim 18, wherein a distinct executable segment of the distinct executable segments comprises a matrix operation, and wherein the matrix operation operates on a matrix in a scratch storage.

20. The apparatus, as recited in claim 19, wherein a first thread of a plurality of threads is an executable sub-segment for a transformation operation which is being performed on a matrix in scratch data in parallel with an executable sub-segment on a second thread of the plurality of threads executing an executable sub-segment of a mathematical operation on scratch data.

21. The apparatus, as recited in claim 20, further comprising a parallel reader configured to read data from an initial input location in a nonsequential manner and provide the data to a plurality of executable sub-segments along data domains through scratch storage.

22. The apparatus, as recited in claim 17, wherein the segments along the process domains are state machines associated with scratch storage, and which automatically execute when data is placed in associated scratch storage.

23. The apparatus, as recited in claim 22, further comprising a parallel writer configured to write data from scratch storage to a final output location, and wherein the final output location is a shared network storage location.