Systems and methods to adaptively select execution modes

Patent metadata

Field	Value
Publication number	US-9939792-B2
Application number	US-201414585738-A
Country	US
Kind code	B2
Filing date	Dec 30, 2014
Priority date	Dec 30, 2014
Publication date	Apr 10, 2018
Grant date	Apr 10, 2018

Abstract

Methods and systems that facilitate efficient and effective adaptive execution mode selection are described. The adaptive execution mode selection is performed in part on-the-fly and changes to an execution mode (e.g., sequential, parallel, etc.) for a program task can be made. An intelligent adaptive selection can be made between a variety execution modes. The adaptive execution mode selection can also include selecting parameters associated with the execution modes. A controller receives historical information associated with execution mode selection, engages in training regarding execution mode selection, and adaptively selects an execution mode on-the-fly. The training can use an approach similar to an artificial neural network in which automated guided machine learning approach establishes correspondences between execution modes and task/input feature definitions based upon historical information. An adaptive selection is performed on-the-fly based on an initial trial run.

First claim

What is claimed is:

1. A system comprising:
a plurality of central processing units;
a memory; and
a controller, embedded in a processing component, the controller coupled to the memory and coupled to the plurality of central processing units, wherein the controller is operable to direct selection of a sequential execution mode or one of a plurality of parallel execution modes with respect to said plurality of central processing units, wherein the controller is further operable to:
generate respective definition characteristics for each task of a plurality of tasks, each definition characteristic comprising a task feature set including, for each respective task, at least one of: a number of instructions in the task, a number of task parameters associated with the task, a time of execution of the task, an amount of memory usage by the task, and a number of instructions in the respective task; and an input feature set including, for each respective task, at least one of: a number of inputs to the task, a size of an input to the task, a type of an input to the task, and a dimension of input data to the task;
establish a plurality of definition pairs that map the definition characteristics to execution modes using a neural network trained by supervised machine learning based training, and
responsive to the generated definition characteristics for each task, execute adaptive selection of the execution mode to be used by the central processing units for each task on-the-fly while the central processing units are running the tasks.

2. The system of claim 1 , wherein the definition operations further include collecting information related to respective heuristics associated with the execution mode selections, the heuristics including a summation of weighted correspondence among the task feature sets, input feature sets, and the execution modes.

3. The system of claim 2 , wherein each of the task feature sets further includes a set of data collected from previous runs of other tasks and a present characteristic of a current task.

4. The system of claim 3 , wherein the set of data collected from the previous runs of the other tasks includes, for each task of the other tasks, at least one feature selected from: a number of instructions of the task; a number of task parameters; a time of execution for the task; and a memory usage of the task.

5. The system of claim 2 , wherein each of the input feature sets includes a set of data describing characteristics of a current input.

6. The system of claim 5 , wherein each of the input feature sets includes at least one of: an input length; an input type; and a dimension of input data.

7. The system of claim 2 , wherein the heuristics associated with each execution mode include a set of process steps that selects an execution mode for one of the tasks.

8. The system of claim 2 , wherein a portion of the information is collected as training samples from historical runs of tasks that are similar to the plurality of tasks.

9. The system of claim 1 wherein execution mode selection includes selecting from among the sequential execution mode and the plurality of parallel execution modes with respectively different parameters each of the respectively different parameters including a number of tasks to be executed at a time.

10. A method executed by a controller, said method comprising:
gathering, by the controller, information related to a plurality of central processing units controlled by said controller;
associating, by the controller, a plurality of definition characteristics with a respective plurality of tasks to define respective execution modes for the plurality of tasks by performing training including performing supervised machine learning, wherein each definition characteristic includes, for each respective task, at least one of: a number of instructions in the task, a number of task parameters associated with the task, a time of execution of the task, an amount of memory usage by the task, a number of inputs to the task, a size of an input to the task, a type of an input to the task, and a dimension of input data to the task; and the execution modes include a sequential execution mode and a plurality of parallel execution modes;
performing, by the controller, adaptive selection of an execution mode for each task from among the execution modes, wherein said adaptive selection is performed on-the-fly responsive to the definition characteristics; and
causing, by the controller, the plurality of central processing units to execute the plurality of tasks utilizing the selected execution modes.

11. The method of claim 10 wherein the supervised machine learning includes:
inputting task features and input data features, the input data features including the number of inputs for the task to provide a first part of paired training sample data;
calculating a current controller output;
calculating a difference between the current controller output and a target controller output, to provide the difference as a second part of the paired training sample data; and
upon determining that the calculated difference is greater than a threshold, propagating an error back to the calculating of the current controller output, otherwise proceeding with a next paired training sample data.

12. The method of claim 11 , wherein the proceeding with the next paired training sample data is performed for input sample data from a first input sample set to a last-input sample set associated with the task features and input data features.

13. The method of claim 10 , wherein the performing adaptive selection of an execution mode includes:
launching a trial version of one of the plurality of tasks using a predetermined execution mode and collecting features therefrom;
selecting one of the execution modes in a feed forward procedure responsive to calculations based on the collected features from the trial version of the one task; and
initiating execution of the one task using the selected execution modes.

14. The method of claim 10 , wherein said gathering of the information includes gathering information associated with heuristics associated with the execution mode selection, the heuristics including a summation of weighted correspondence among task feature sets, input feature sets, and the execution modes.

15. The method of claim 10 , wherein:
the definition characteristics further include information associated with at least one feature selected from a number of task parameters, a time of execution, and a memory usage measure and information associated with at least one feature selected from: an input length, an input type, and a dimension of an input data item.

16. A system comprising:
a controller;
a learning module which configures the controller to receive a plurality of definition characteristics for a respective plurality of tasks, each definition characteristic comprising a task feature set including, for each respective task, at least one of: a number of instructions in the task, a number of task parameters associated with the task, a time of execution of the task, and an amount of memory usage by the task; and an input feature set including, for each respective task, at least one of: a number of inputs to the task, a size of an input to the task, a type of an input to the task, and a dimension of input data to the task;
an execution module which configures the controller to run trials of the respective tasks using a predetermined execution mode and to select respective execution modes of a plurality of execution modes to be used for the tasks running on a plurality of cen

Assignees

• Futurewei Technologies Inc
• Futurewei Technologies Inc

Inventors

• Chen Liya
• Tian Chen
• Hu Ziang

Classifications

• G05B13/027 Primary

  using neural networks only · CPC title

• G06N99/005

  Physics · mapped topic

• G06N20/00

  Machine learning · CPC title