Systems and methods for power optimization of processors

Patent metadata

Field	Value
Publication number	US-10180828-B2
Application number	US-201514699876-A
Country	US
Kind code	B2
Filing date	Apr 29, 2015
Priority date	Apr 29, 2014
Publication date	Jan 15, 2019
Grant date	Jan 15, 2019

Abstract

A compilation system generates one or more energy windows in a program to be executed on a data processors such that power/energy consumption of the data processor can be adjusted in which window, so as to minimize the overall power/energy consumption of the data processor during the execution of the program. The size(s) of the energy window(s) and/or power option(s) in each window can be determined according to one or more parameters of the data processor and/or one or more characteristics of the energy window(s).

First claim

Accordingly, we claim:

1. A method for optimizing energy consumption of a data processor while executing a program, the method comprising performing by a compilation processor the steps of:
generating within the program a first window;
determining window type of the first window; and
inserting a power control operation in the first window based on, at least in part, energy consumed by operations in the first window, the energy consumed being computed using: at least one parameter of at least one component of the data processor, at least one characteristic of the first window, and the window type of the first window;
wherein the window type of the first window is computation bound window or memory bound window; and
when the window type of the first window is computation bound window the power control operation changes an operating voltage of a CPU of the data processor from a first value to a second value greater than the first value only if a parameter related to compute operations is not less than a function of a transition latency associated with transitioning operation of the CPU from a first CPU frequency to a second CPU frequency greater than the first CPU frequency; and
when the window type of the first window is memory bound window the power control operation changes an operating voltage of a CPU of the data processor from a second value to a first value less than the second value only if a parameter related to memory operations is not less than a function of a transition latency associated with transitioning operation of the data processor from a second CPU frequency to a first CPU frequency less than the second frequency.

2. The method of claim 1 , wherein generating the first window comprises generating a window of a particular size, the particular size being determined as a function of a transition latency associated with transitioning operation of a component of data processor from a first component-operating frequency to a second-component operating frequency different from the first component-operating frequency.

3. The method of claim 1 , wherein:
the first window comprises at least one group comprising a sequence of statements of the program; and
generating the first window comprises:
analyzing a representation of the program in a particular format; and
forming, using that format, the at least one group, at a granularity based on, at least in part, at least one parameter of at least one component of the data processor.

4. The method of claim 1 , wherein determining the window type of the first window comprises:
computing a number of computation operations associated with the first window;
computing a number of memory operations associated with the first window;
computing arithmetic intensity of the first window as a function of the number of computation operations and the number of memory operations; and
setting the window type to memory bound window if the arithmetic intensity is less than a threshold and otherwise setting the window type to computation bound window.

5. The method of claim 1 , wherein determining the window type of the first window comprises inserting in the first window:
an expression to compute at runtime a number of computation operations associated with the first window;
an expression to compute at runtime a number of memory operations associated with the first window;
an expression to compute at runtime an arithmetic intensity of the first window as a function of the number of computation operations and the number of memory operations; and
an expression to set at runtime the window type to memory bound window if the arithmetic intensity is less than a threshold and otherwise to set the window type to computation bound window.

6. The method of claim 1 , wherein the window type of the first window is selected from a group consisting of a computation bound window and a memory bound window.

7. The method of claim 1 , wherein the at least one component of the data processor is selected from a group consisting of: a central processing unit (CPU), a memory bank, a cache memory module, a memory bus, a memory controller, and an application specific accelerator.

8. The method of claim 1 , wherein the at least one parameter of the data processor is selected from a group consisting of: a transition latency associated with transitioning operation of at least one component of the data processor from a first frequency to a second frequency different from the first frequency; time to execute a compute operation; time to execute a memory operation; static power consumed by at least one component of the data processor; dynamic power consumed by at least one component of the data processor during a compute operation; and dynamic power consumed by at least one component of the data processor during a memory operation.

9. The method of claim 1 , wherein the at least one characteristic of the first window is selected from a group consisting of: a count of compute operations, a count of memory operations, an estimated compute time, an estimated number of compute cycles, an estimated number of data access cycles; and a memory footprint.

10. The method of claim 1 , wherein the parameter related to compute operations is selected from a group consisting of: a count of compute operations, an estimated number of cycles associated with the compute operations, a measured number of cycles associated with the compute operations, an estimated time required for the compute operations, a measured time required for the compute operations, an estimated energy required by the compute operations, a measured energy required by the compute operations, an estimated power required by the compute operations, and a measured power required by the compute operations.

11. The method of claim 1 , wherein:
the power control operation further changes an operating frequency of a memory bus of the data processor from a second memory bus frequency to a first memory bus frequency that is less than the second memory bus frequency.

12. The method of claim 1 , wherein the parameter related to memory operations is selected from a group consisting of: a count of memory operations, an estimated number of cycles associated with the memory operations, a measured number of cycles associated with the memory operations, an estimated time required for the memory operations, a measured time required for the memory operations, an estimated energy required by the memory operations, a measured energy required by the memory operations, an estimated power required by the memory operations, and a measured power required by the memory operations.

13. The method of claim 1 , wherein:
the power control operation further changes an operating frequency of a memory bus of the data processor from a first memory bus frequency to a second memory bus frequency that is greater than the first memory bus frequency.

14. The method of claim 1 , wherein:
the power control operation is based on, at least in part, an operation count associated with the first window, the operation count being determined at runtime; and
the power control operation modifies at runtime at least one attribute of at least one component of the data processor only if the operation count is greater than a threshold.

15. The method of claim 1 , further comprising performing by the compilation processor the steps of:
generating within the program a second window;
determining window type of the second window;
inserting a power control operation in the second window based on, at least in part, at least one of: at least one parameter of the data processor, at least one cha

Assignees

• Reservoir Labs Inc
• Significs And Elements Llc

Inventors

• Baskaran Muthu M
• Henretty Thomas
• Johnson Ann
• Konstantinidis Athanasios
• Langston M H
• Lethin Richard A
• Mcmahon Janice O
• Meister Benoit J
• Mountcastle Paul
• Pradelle Benoit

Classifications

• Y02D10/41

  Cross-Sectional Technologies · mapped topic

• G06F8/4432 Primary

  Reducing the energy consumption · CPC title

• Y02D10/00

  Energy efficient computing, e.g. low power processors, power management or thermal management · CPC title