Systems and methods for reducing carbon emissions from machine learning computational tasks

Abstract

A computer system is provided, including a scheduling computing device configured to receive computing task data defining a computing task to be performed, retrieve site data corresponding to each of a plurality of data processing computing devices, and select, based on the computing task data and the site data, i) a first computing algorithm for performing the computing task, ii) a first data processing computing device among the plurality of data processing computing devices, and iii) at least one time period for executing the first computing algorithm by the first data processing computing device, wherein the first computing algorithm, the first data processing computing device, and the at least one time period are selected to facilitate a reduction in carbon dioxide emissions associated with the execution of the computing algorithm, and further include a scheduling computing device configured to instruct the first data processing computing device to execute the first computing algorithm during the at least one time period.

Description

[Technical Field]

[0001] The field of the invention relates generally to computer systems that perform computational learning tasks, such as machine learning (ML), artificial intelligence (AI), computational fluid dynamics simulations, and other computational and learning tasks, and more particularly to systems and methods for reducing carbon dioxide emissions resulting from the performance of computational learning tasks. [Background technology]

[0002] Computational learning tasks, such as ML and AI, are widely used in many applications. However, increasing training data sizes and model sizes, for example, necessitate increasing computational resources required to train ML models. Because ML is an energy-intensive computational task and computational resources are often powered by carbon dioxide-emitting sources, performing ML tasks can generate significant amounts (e.g., greater than 500 tons) of carbon dioxide emissions and / or other undesirable emissions. Therefore, a system that can reduce the carbon dioxide emissions associated with performing ML tasks while maintaining desired performance characteristics (e.g., accuracy) of ML models is desirable. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] U.S. Patent Application Publication No. 2022-0261685 Summary of the Invention

[0004] In one aspect, a computer system is provided. The computer system includes a plurality of data processing computing devices. Each of the plurality of data processing computing devices is configured to execute one or more computational algorithms. The computer system further includes a scheduling computing device including a memory and a processor in communication with the memory and the plurality of data processing computing devices. The processor is configured to receive computational task data defining a computational task to be performed. The processor is further configured to retrieve site data corresponding to each of the plurality of data processing computing devices. The site data specifies an expected carbon dioxide emission associated with use of each of the plurality of data processing computing devices. The processor is further configured to select, based on the computational task data and the site data, i) a first computational algorithm for executing the computational task, ii) a first data processing computing device of the plurality of data processing computing devices, and iii) at least one time period for executing the first computational algorithm by the first data processing computing device, wherein the first computational algorithm, the first data processing computing device, and the at least one time period are selected to promote a reduction in a carbon dioxide emission associated with execution of the computational algorithm. The processor is further configured to instruct the first data processing computing device to execute the first computing algorithm within at least one period of time.

[0005] In another aspect, a method is provided. The method is performed by a scheduling computing device including a memory and a processor in communication with the memory and a plurality of data processing computing devices. Each of the data processing computing devices is configured to execute one or more computational algorithms. The method includes receiving, by the scheduling computing device, computational task data defining a computational task to be executed. The method further includes retrieving, by the scheduling computing device, site data corresponding to each of the plurality of data processing computing devices. The site data specifies an expected carbon dioxide emission associated with use of each of the plurality of data processing computing devices. The method further includes selecting, by the scheduling computing device, based on the computational task data and the site data: i) a first computational algorithm for executing the computational task; ii) a first data processing computing device of the plurality of data processing computing devices; and iii) at least one time period for executing the first computational algorithm by the first data processing computing device, wherein the first computational algorithm, the first data processing computing device, and the at least one time period are selected to promote a reduction in a carbon dioxide emission associated with execution of the computational algorithm. The method further includes the step of instructing, by the scheduling computing device, the first data processing computing device to execute the first computing algorithm within at least one time period.

[0006] In another aspect, a scheduling computing device is provided. The scheduling computing device includes a memory, a processor in communication with the memory, and a plurality of data processing computing devices. Each of the plurality of data processing computing devices is configured to execute one or more computational algorithms. The processor is configured to receive computational task data defining a computational task to be executed. The processor is further configured to retrieve site data corresponding to each of the plurality of data processing computing devices. The site data specifies an expected carbon dioxide emission associated with use of each of the plurality of data processing computing devices. The processor is further configured to select, based on the computational task data and the site data, i) a first computational algorithm for executing the computational task, ii) a first data processing computing device of the plurality of data processing computing devices, and iii) at least one time period for executing the first computational algorithm by the first data processing computing device, wherein the first computational algorithm, the first data processing computing device, and the at least one time period are selected to promote a reduction in carbon dioxide emission associated with execution of the computational algorithm. The processor is further configured to instruct the first data processing computing device to execute the first computational algorithm within the at least one time period.

[0007] These and other features, aspects and advantages of the present disclosure will become better understood by considering the following detailed description in conjunction with the accompanying drawings, in which like characters represent like parts throughout. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an exemplary computer system for performing ML tasks. [Figure 2A] 1 is a graph showing an example of a site's renewable energy mix over time. [Figure 2B] 10 is a graph showing another example of a site's renewable energy mix over time. [Figure 2C] 10 is a graph showing yet another example of a site's renewable energy mix over time. [Figure 3] FIG. 10 is a diagram illustrating an example of a schedule for allocation of ML tasks. [Figure 4A] 1 is a graph illustrating an example of available power for a site over time. [Figure 4B] 10 is a graph illustrating another example of available power at a site over time. [Figure 4C] 10 is a graph illustrating yet another example of available power at a site over time. [Figure 5] 1 is a flowchart of an example method for scheduling the execution of an ML task. DETAILED DESCRIPTION OF THE INVENTION

[0009] In the following specification and claims, reference will be made to a number of terms that shall be defined to have the following meanings.

[0010] The singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise.

[0011] As used herein throughout this specification and claims, approximation language may be applied to modify any quantitative expression that may vary permissibly without resulting in a change in the basic function involved. Thus, values ​​modified with terms such as "about," "substantially," "approximately," etc. are not limited to the exact value specified. In at least some instances, approximation language may correspond to the precision of an instrument for measuring the value. Here, and throughout this specification and claims, range limitations are combinable and / or interchangeable, and unless the context and language dictate otherwise, such ranges are identified and include all subranges encompassed therein.

[0012] In an exemplary embodiment, a computer system for scheduling and executing ML tasks includes a plurality of data processing computing devices, each configured to execute one or more machine learning (ML) algorithms. The data processing computing devices may be located at a plurality of different sites, and each site may receive power from some combination of renewable and non-renewable power sources (e.g., carbon-emitting power sources). The proportion of power received from renewable or non-carbon-emitting power sources may fluctuate or change over time.

[0013] The computer system further includes a scheduling computing device including a memory and a processor in communication with the memory and the plurality of data processing computing devices. The processor is configured to receive ML task data defining ML tasks to be performed and to retrieve site data corresponding to each of the plurality of data processing computing devices. The task data may include training data used to train ML models, a specific task to be performed (regression, classification, clustering), and / or a specification list of the ML models required. The site data specifies an expected carbon footprint associated with use of each of the plurality of data processing computing devices at a given time.

[0014] The scheduling computing device is further configured to select, based on the ML task data and the site data, i) an ML algorithm for executing the ML task, ii) a data processing computing device or location for executing the ML task, and iii) at least one time period for executing the first ML algorithm by the selected data processing computing device. The ML algorithm, data processing computing device, and time period are selected to reduce carbon dioxide emissions associated with executing the ML algorithm. For example, the ML algorithm and the number of iterations for executing the ML algorithm can be selected to maximize the accuracy of the generated ML model per unit energy used or per unit carbon dioxide emissions generated. The location and time period for executing the ML algorithm can be selected to maximize the proportion of renewableally sourced electricity during the execution of the ML algorithm, thereby reducing the resulting carbon dioxide emissions.

[0015] 1 is a block diagram illustrating an exemplary computer system 100. The computer system 100 includes multiple data processing computing devices 102, each of which may be configured to execute one or more ML algorithms. The computer system further includes a scheduling computing device 104 including a memory 106, a processor 108, and an input / output (I / O) interface 110. The I / O interface 110 is configured to enable the scheduling computing device 104 to communicate with the data processing computing devices, for example, via a local and / or cloud computing network. In some embodiments, the I / O interface 110 further enables data to be input and / or output to the scheduling computing device, for example, via a local user interface and / or via another computing device in communication with the scheduling computing device 104.

[0016] The data processing computing devices 102 are located at multiple different locations or sites. In some embodiments, the data processing computing devices may be associated with or form part of a data center or other similar system. Each location or site may receive power from some combination of renewable or non-renewable resources. The combination at each site may vary over time. For example, the amount of renewable or non-renewable power available at a particular site may depend on the time of day, weather (e.g., solar and / or wind energy), and other factors that change over time.

[0017] The scheduling computing device 104 is configured to store data, referred to herein as “site data,” that describes the data processing computing devices 102 and their respective sites. The site data may include information about the data processing computing devices 102, such as, for example, hardware processing and storage capabilities. The site data may further include information that enables the scheduling computing device 104 to predict the type of power source (e.g., renewable or non-renewable) available to power the data processing computing device 102 for a given time period. Thus, as described in further detail below, for a given computational or ML task, the scheduling computing device 104 can predict the amount of energy required by the data processing computing device 102 to complete the ML task and, if at least a portion of that energy is provided by non-renewable resources, the carbon dioxide emissions associated with completing the ML task. The site data may be updated (e.g., periodically or continuously) based on various input sources, such as the data processing computing device 102, other computing devices associated with the computer system 100, and / or external data sources such as the Internet.

[0018] Scheduling computing device 104 is configured to receive task data relating to or defining an ML task to be executed by computer system 100. The task data may include, for example, input and output data types, input datasets, desired or required accuracy of the ML model, deadlines for completing the ML task, and / or other parameters or requirements associated with the ML task. In some embodiments, the task data further includes quality specifications (e.g., expected accuracy), datasets (e.g., including size, preprocessing, and / or feature selection), expected energy requirements, and / or expected computation time. In some embodiments, the task data further defines a priority (e.g., high, medium, and / or low) of the ML task, based on which the ML task may be prioritized relative to other ML tasks during scheduling.

[0019] Based on the task data, the scheduling computing device 104 is configured to select an ML algorithm for performing the ML task. Some examples of ML algorithms that may be performed include Bidirectional Encoder Representation from Transformers (BERT), Generative Pre-trained Transformer 2 (GPT-2), Generative Pre-trained Transformer 3 (GPT-3), Elmo, Meena, transformers, and / or other models. In some embodiments, additional algorithm specifications and / or quantification parameters, such as hyperparameters (e.g., learning rate) and / or structural parameters (e.g., number of hidden layers), are determined. In some embodiments, the scheduling computing device 104 is further configured to determine the number of iterations of the selected ML algorithm that need to be performed to achieve a desired accuracy or to operate optimally. To select the ML algorithm, the scheduling computing device 104 is configured to calculate an estimated energy requirement for each of a group of candidate ML algorithms, taking into account the number of iterations each candidate ML algorithm needs to perform. The scheduling computing device 104 is configured to select an ML algorithm from among the candidate ML algorithms based on the estimated energy requirements. For example, the scheduling computing device 104 may select the ML algorithm that requires the least energy or the ML algorithm that has the highest accuracy per unit of energy and / or per unit of carbon dioxide footprint. As described in more detail below, in some embodiments, the ML algorithm is selected in combination with the location and time at which the ML algorithm is run to optimize energy usage and / or carbon dioxide footprint, for example.

[0020] The accuracy of a model can be measured by many means, including loss functions such as mean squared error or mean absolute error that can be generalized on a validation data set (not part of training), or the classification error rate for classification problems. Accuracy for unsupervised learning problems such as clustering can be determined through metrics such as the Jensen-Shannan distance, or other techniques that determine the reproducibility of the learned solution on data drawn from the source distribution.

[0021] The scheduling computing device 104 is further configured to select a data processing computing device 102 for executing the ML task. A particular data processing computing device 102 may be selected based on one or more of the ML task data, the selected ML algorithm, site data associated with the data processing computing device 102, and the predicted energy usage and / or carbon dioxide emissions associated with executing the selected ML algorithm on the data processing computing device 102. For example, in some embodiments, the scheduling computing device 104 calculates the energy requirements and / or carbon dioxide emissions for executing the ML algorithm on each of a group of candidate data processing computing devices 102 and selects a data processing computing device 102 from among the group of candidate data processing computing devices 102 based on the predicted energy requirements and / or carbon dioxide emissions of each candidate data processing computing device 102. For example, the scheduling computing device 104 may select the ML algorithm that requires the least energy and / or produces the smallest carbon dioxide emissions. As described in more detail below, in some embodiments, the data processing computing device 102 used to execute the ML algorithm is selected in combination with the ML algorithm and the time of day that the ML algorithm is executed to, for example, optimize energy usage and / or carbon dioxide emissions.

[0022] The scheduling computing device 104 is further configured to select at least one time period for executing the ML task. In some embodiments, the ML task may be executed in parallel or serially by multiple data processing computing devices 102, in consecutive or non-consecutive time periods. To determine when to execute the ML task, the scheduling computing device 104 considers the availability of renewable power over time. For example, the scheduling computing device 104 may select a time period in which more renewable power is expected to be available (e.g., based on the time of day, weather conditions, or other factors) to reduce the amount of non-renewable power used and carbon dioxide emissions generated. The scheduling computing device 104 may further consider the availability of each data processing computing device 102 based on already scheduled tasks and reschedule the tasks to optimize energy usage and / or carbon dioxide emissions. As described in more detail below, in some embodiments, the time period for executing the ML algorithm is selected in combination with the ML algorithm and the data processing computing device 102 on which the ML algorithm is executed, e.g., to optimize energy usage and / or carbon dioxide emissions.

[0023] For example, to select an ML algorithm, a number of iterations to run the ML algorithm, a data processing computing device 102 to run the ML algorithm, and / or a time period for running the ML algorithm, the scheduling computing device 104 is configured to run a constrained optimization problem that maximizes the accuracy of the ML algorithm per unit of energy expected to be used and / or per unit of carbon dioxide emissions expected to be generated. The constraints of the optimization problem may include factors such as the computing power of the data processing computing device 102, the expected availability of renewable electricity, and / or the current utilization of the data processing computing device 102 (e.g., for running other ML tasks). For a given task, further constraints may apply, such as, for example, the accuracy required for the ML model, a deadline for completing the ML task, the ML task having to be performed by only one data processing computing device 102 and / or site or by one or more specifically specified data processing computing devices 102 and / or sites, the ML task having to be performed within a preset period and / or within a single consecutive period, a maximum energy consumption or maximum carbon dioxide emission, whether the ML task can be performed in parallel by multiple data processing computing devices 102, and / or other constraints.

[0024] The scheduling computing device 104 is further configured to generate a schedule associated with the ML task. The schedule includes a selected algorithm, a selected data processing computing device 102, and at least one selected time period for executing the ML algorithm. In some embodiments, the ML algorithm may be executed partially at two or more sites and / or within separate or discontinuous time periods. The scheduling computing device 104 is configured to instruct the selected data processing computing device to execute the selected ML algorithm according to the schedule. In some embodiments, the scheduling computing device 104 may modify the schedule based on subsequent input of additional ML tasks, for example, to optimize each task to reduce energy usage and / or carbon dioxide emissions. In some embodiments, the ML tasks may be further scheduled to reduce or minimize start delays relative to assigned priorities (e.g., high, medium, or low priorities) associated with the ML tasks.

[0025] In some embodiments, the scheduling computing device 104 may be configured to preprocess input data for an ML model, or to have the data processing computing device 102 preprocess the input data for the ML model, to reduce the quantization bits of the input data. This preprocessing can reduce the energy required to run an ML algorithm and create an ML model without significantly affecting the quality or accuracy of the ML model's output. For example, in some implementations, input data can be preprocessed to reduce quantization levels, discarding unnecessary, and potentially unfounded, requirements on accuracy. Scaling and dimensionality reduction, such as through methods such as principal component analysis (PCA), can be used to effectively preprocess data to improve performance and reduce the energy cost of training.

[0026] In some embodiments, a particular site (e.g., corresponding to a particular data processing computing device 102) utilizes heat generated from the data center's cooling process to power the associated data processing computing device 102 (e.g., to power a combined cycle gas / steam power plant). In such embodiments, the scheduling computing device 104 can take this capability into account when selecting a data processing computing device 102 for executing an ML task.

[0027] In some embodiments, the scheduling computing device 104 is further configured to consider the stability of the power grid when selecting a data processing computing device 102 to complete each ML task in order to distribute the execution of the ML tasks across the computer system 100 in a manner that promotes grid stability. In such embodiments, the scheduling computing device 104 may further receive and store information about the power grid to determine which data processing computing device 102 to use to complete the ML tasks while maintaining power grid stability.

[0028] 2A, 2B, and 2C show a first graph 200, a second graph 202, and a third graph 204, respectively. Graphs 200, 202, and 204 each illustrate an example “renewable energy mix,” or percentage of available electricity that is renewableally sourced, for each site. Graphs 200, 202, and 204 each include a horizontal axis 206 representing four discrete time periods t1, t2, t3, and t4, which may correspond, for example, to portions of a day or other time periods. Graphs 200, 202, and 204 further include a vertical axis corresponding to the renewable energy mix, or the dimensionless ratio of the amount of renewable electricity to the total amount of electricity used at each site (e.g., by each data processing computing device 102). As shown in graphs 200, 202, and 204, the renewable energy mix for a given site can change over time. This may be due to factors such as the availability of renewable (e.g., solar and / or wind) electricity, other power demands on the system, and / or other such factors. As discussed above with respect to Figure 1, the scheduling computing device 104 is configured to take the expected renewable energy mix into account when selecting sites and times for executing ML tasks.

[0029] 3 is a chart 300 illustrating an exemplary schedule for allocating ML tasks over time among different data processing computing devices 102, as may be determined by a scheduling computing device 104. The chart 300 includes a vertical axis 302 representing three data processing computing devices 102, including a first data processing computing device 304, a second data processing computing device 306, and a third data processing computing device 308. The chart 300 further includes a horizontal axis 310 representing 24 discrete periods (e.g., times of day) during which some of the ML tasks may be executed.

[0030] Consider the example scenario shown in Table 1, which includes specific task data associated with three ML tasks (ML Task 0, ML Task 1, and ML Task 2).

[0031] [Table 1]

[0032] For each ML task, the expected energy requirement indicates the expected amount of energy needed to execute the ML task, the arrival time represents the time the ML task will be received, the computation time represents the expected number of hours needed to complete the computation of the ML task, and the start time tolerance represents an indication of the priority of the start time (e.g., the ML task must start within 2, 4, or 6 hours). Based on this task data, the scheduling computing device can schedule ML task 0, ML task 1, and ML task 2 among the first data processing computing device 304, the second data processing computing device 306, and the third data processing computing device 308, as shown in chart 300. As described above, the scheduling computing device 104 is configured to schedule the ML tasks to increase or maximize an achievable figure of merit (e.g., accuracy per unit energy and / or accuracy per unit carbon dioxide emission) based on the constraints of the task data.

[0033] 4A, 4B, and 4C show a first graph 400, a second graph 402, and a third graph 404, respectively. The first graph 400 represents energy availability and usage over time for the first data processing computing device 304, the second graph 402 represents energy availability and usage over time for the second data processing computing device 306, and the third graph 404 represents energy availability and usage over time for the third data processing computing device 308. Each of the first graph 400, the second graph 402, and the third graph 404 includes a vertical axis 408 that indicates power in kilowatts (kW). Each of the first graph 400, the second graph 402, and the third graph 404 further includes a horizontal axis 410 that indicates 24 discrete periods that correspond to the discrete periods represented by the horizontal axis 310 of the chart 300 shown in FIG. 3.

[0034] Each of the first graph 400, second graph 402, and third graph 404 further includes a total available power curve 412, an available non-carbon dioxide emitting power curve 414, and an available carbon dioxide emitting power curve 416. The total available power curve 412 represents the total available power of a given data processing computing device 102 for a period of time, the available non-carbon dioxide emitting power curve 414 represents the power available from non-carbon dioxide emitting resources for that period, and the available carbon dioxide emitting power curve 416 represents the power available from carbon dioxide emitting resources for that period. Thus, the value represented by the total available power curve 412 is the sum of the values ​​represented by the available non-carbon dioxide emitting power curve 414 and the available carbon dioxide emitting power curve 416.

[0035] As shown in graphs 400 and 402, when run according to the schedule shown in chart 300, the power 418 consumed during each time period by each data processing computing device 102 is less than the available carbon dioxide emitting power represented by available carbon dioxide emitting power curve 416. Therefore, in this scenario, ML task 0, ML task 1, and ML task 2 can be run without emitting any carbon dioxide.

[0036] 5 is a flowchart illustrating an example method 500 for scheduling ML tasks for execution using a computer system such as computer system 100 (shown in FIG. 1). The computer system includes a plurality of data processing computing devices (such as data processing computing device 102), each configured to execute one or more computational algorithms (e.g., ML and / or AI algorithms). The computer system further includes a scheduling computing device (such as scheduling computing device 104), which includes a memory (such as memory 106) and a processor (such as processor 108) in communication with the memory and the plurality of data processing computing devices.

[0037] The method 500 includes receiving 502 computational task data defining a computational task to be performed (e.g., an ML and / or AI task and / or a computationally expensive simulation). The method 500 further includes retrieving 504 site data corresponding to each of the plurality of data processing computing devices, the site data specifying an expected carbon footprint associated with use of each of the plurality of data processing computing devices.

[0038] The method 500 further includes step 506 of selecting, based on the computational task data and the site data, i) a first computational algorithm for executing the computational task, ii) a first data processing computing device among the plurality of data processing computing devices, and iii) at least one time period for executing the first computational algorithm by the first data processing computing device, wherein the first computational algorithm, the first data processing computing device, and the at least one time period are selected to reduce carbon dioxide emissions associated with executing the computational algorithm.

[0039] The method 500 further includes instructing 508 the first data processing computing device to execute the first ML algorithm during at least one period of time.

[0040] In some embodiments, the computational task is an ML task and the first computational algorithm is an ML algorithm.

[0041] In some embodiments, the method 500 further includes selecting the first ML algorithm based on a ratio of accuracy to unit carbon dioxide emission.

[0042] In some embodiments, the method 500 further includes determining a number of iterations to run the first ML algorithm based on a ratio of accuracy to unit carbon dioxide emission.

[0043] In some embodiments, the method 500 further includes determining, based on the site data, a proportion of the electricity supplied by renewable resources for each of the plurality of data processing computing devices for each of the plurality of time periods, and selecting a first data processing computing device and at least one time period based on the proportion determined for each of the plurality of data processing computing devices for each of the plurality of time periods.

[0044] In some embodiments, method 500 further includes receiving one or more user-defined constraints for selecting a first data processing computing device and at least one time period, and selecting the first data processing computing device and at least one time period further based on the one or more user-defined constraints. In some such embodiments, the one or more user-defined constraints stipulate that only one data processing computing device be used to execute the first ML algorithm. In some such embodiments, the one or more user-defined constraints stipulate that the first ML algorithm be executed during a single continuous time period.

[0045] In some embodiments, the method 500 further includes receiving second ML task data defining a second ML task to be executed, and further generating a schedule based on the second ML task data.

[0046] In some embodiments, the method 500 further includes pre-processing the input data of the first computational algorithm to reduce the number of quantization bits of the input data and / or the number of dimensions used for training.

[0047] Exemplary technical effects of the methods, systems, and apparatus described herein include at least one of: (a) reducing carbon dioxide emissions associated with the execution of an ML task by selecting an ML algorithm, a location for executing the ML task, and a duration for executing the ML task based on the expected contribution of renewable power sources to the execution of the ML task; and (b) reducing carbon dioxide emissions associated with the execution of an ML task by selecting an ML algorithm and the number of iterations in executing the ML algorithm based on the accuracy of the resulting ML model per expected unit of carbon dioxide emissions associated with the execution of the ML algorithm.

[0048] Exemplary embodiments of computer systems are provided herein. The systems and methods are not limited to the specific embodiments described herein; rather, system components and / or method steps can be utilized independently and separately from other components and / or steps described herein. For example, the methods may also be used in combination with other electronic systems and are not limited to practice with only the electronic systems and methods described herein. Rather, exemplary embodiments can be implemented and utilized in connection with many other electronic systems.

[0049] Some embodiments involve the use of one or more electronic or computing devices (e.g., scheduling computing device 104). Such devices typically include a processor, processing device, or controller, such as, for example, a general-purpose central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a reduced instruction set computer (RISC) processor, an application-specific integrated circuit (ASIC), a programmable logic controller (PLC), a field-programmable gate array (FPGA), a digital signal processing (DSP) device, and / or any other circuit or processing device capable of performing the functions described herein. The methods described herein may be encoded as executable instructions embodied in a computer-readable medium, including, but not limited to, a storage device and / or a memory device. Such instructions, when executed by a processing device, cause the processing device to perform at least a portion of the methods described herein. The above-described embodiments are merely examples and are thus not intended to limit in any way the definition and / or meaning of the terms processor and processing device.

[0050] Although specific features of various embodiments of the present disclosure may be shown in some drawings and not in others, this is for convenience only, and in accordance with the principles of the present disclosure, any feature of a drawing may be referenced and / or claimed in combination with any feature of any other drawing.

[0051] This written description uses examples to disclose the invention, including the best mode, and to enable any person skilled in the art to practice the invention, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the invention is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they contain equivalent structural elements that do not differ in material way from the literal language of the claims. [Explanation of symbols]

[0052] 100 Computer Systems 102 Data processing computing equipment 104 Scheduling computer 106 memory 108 processors 110 Input / Output (I / O) Interface 200 First Graph 202 Second graph 204 Third Graph 206 Horizontal Axis 300 Chart (showing an exemplary schedule) 302 Vertical Axis 304 First data processing computing device 306 Second data processing computing device 308 Third data processing computing device 310 Horizontal axis 400 First Graph 402 Second Graph 404 Third Graph 408 Vertical Axis 410 horizontal axis 412 Total Available Power Curve 414 Non-CO2 Emissions Power Curve 416 Carbon Dioxide Emission Power Curve 418 Electricity 500 ways

Claims

1. A computer system (100), a plurality of data processing computing devices (102), each of the plurality of data processing computing devices (102) configured to execute one or more computing algorithms; A scheduling computing device (104) including a memory (106) and a processor (108) in communication with the memory (106) and the plurality of data processing computing devices (102), the processor (108) comprising: receiving computational task data defining a computational task to be performed; retrieving site data corresponding to each of the plurality of data processing computing devices (102), the site data specifying the carbon dioxide emissions expected to be associated with the use of each of the plurality of data processing computing devices (102); selecting, based on the computing task data and the site data, i) a first computing algorithm for executing the computing task, ii) a first data processing computing device (102) among the plurality of data processing computing devices (102), and iii) at least one time period for executing the first computing algorithm by the first data processing computing device (102), wherein the first computing algorithm, the first data processing computing device (102), and the at least one time period are selected to facilitate a reduction in carbon dioxide emissions associated with execution of the computing algorithm; instructing said first data processing computing device (102) to execute said first computing algorithm during said at least one period of time; a scheduling computing device (104) further configured to A computer system (100) comprising:

2. 2. The computer system (100) of claim 1, wherein the computational task is a machine learning (ML) task and the first computational algorithm is an ML algorithm.

3. The computer system (100) of claim 1, wherein the processor (108) is further configured to select the first calculation algorithm based on a ratio of accuracy to unit carbon dioxide emissions.

4. 2. The computer system (100) of claim 1, wherein the processor (108) is further configured to determine a number of iterations to perform the first calculation algorithm based on a ratio of accuracy to unit carbon dioxide emissions.

5. The processor (108): determining a percentage of electricity supplied by renewable sources for each of the plurality of data processing computing devices (102) for each of a plurality of time periods based on the site data; selecting the first data processing computing device (102) and the at least one period based on the ratio determined for each of the plurality of data processing computing devices (102) for each of the plurality of periods; The computer system (100) of claim 1 configured to:

6. The processor (108): receiving one or more user-defined constraints for selecting the first data processing computing device (102) and the at least one time period; selecting the first data processing computing device (102) and the at least one time period further based on the one or more user-defined constraints; The computer system (100) of claim 1 further configured to:

7. 7. The computer system (100) of claim 6, wherein the one or more user-defined constraints stipulate that only one data processing computing device (102) is used to execute the first computational algorithm.

8. 7. The computer system (100) of claim 6, wherein the one or more user-defined constraints dictate that the first computational algorithm be executed during a single continuous period of time.

9. The processor (108): receiving second computational task data defining a second computational task to be performed; selecting the first data processing computing device (102) and the at least one time period further based on the second computing task data; The computer system (100) of claim 1 further configured to:

10. 2. The computer system of claim 1, wherein the processor is further configured to preprocess input data of the first computational algorithm to reduce a number of quantization bits of the input data and / or a number of dimensions used for training.

11. 1. A method (500) performed by a scheduling computing device (104) including a memory (106) and a processor (108) in communication with the memory (106) and a plurality of data processing computing devices (102), comprising: Each of the data processing computing devices (102) is configured to execute one or more computing algorithms, and the method (500) comprises: receiving (502) computational task data defining a computational task to be performed by said scheduling computing device (104); retrieving (504) by said scheduling computing device (104) site data corresponding to each of said plurality of data processing computing devices (102), said site data specifying an expected carbon dioxide emission associated with use of each of said plurality of data processing computing devices (102); selecting (506), by the scheduling computing device (104), based on the computing task data and the site data: i) a first computing algorithm for executing the computing task; ii) a first data processing computing device (102) of the plurality of data processing computing devices (102); and iii) at least one time period for executing the first computing algorithm by the first data processing computing device (102), wherein the first computing algorithm, the first data processing computing device (102), and the at least one time period are selected to facilitate a reduction in carbon dioxide emissions associated with execution of the computing algorithm; instructing (508) by said scheduling computing device (104) said first data processing computing device (102) to execute said first computing algorithm during said at least one period; A method (500) comprising:

12. 12. The method (500) of claim 11, wherein the computational task is a machine learning (ML) task and the first computational algorithm is an ML algorithm.

13. 12. The method (500) of claim 11, further comprising selecting, by the scheduling computing device (104), the first computing algorithm based on a ratio of accuracy to unit carbon dioxide emissions.

14. 12. The method (500) of claim 11, further comprising determining, by the scheduling computing device (104), a number of iterations to execute the first computing algorithm based on a ratio of accuracy to unit carbon dioxide emissions.

15. determining, by the scheduling computing device (104), based on the site data, a percentage of electricity supplied by renewable sources for each of the plurality of data processing computing devices (102) for each of a plurality of time periods; selecting, by the scheduling computing device (104), the first data processing computing device (102) and the at least one period based on the ratio determined for each of the plurality of data processing computing devices (102) for each of the plurality of periods; 12. The method (500) of claim 11, further comprising:

16. receiving, by said scheduling computing device (104), one or more user-defined constraints for selecting said first data processing computing device (102) and said at least one time period; selecting, by the scheduling computing device (104), the first data processing computing device (102) and the at least one time period further based on the one or more user-defined constraints; 12. The method (500) of claim 11, further comprising:

17. 17. The method (500) of claim 16, wherein the one or more user-defined constraints stipulate that only one data processing computing device (102) is used to execute the first computational algorithm.

18. 17. The method (500) of claim 16, wherein the one or more user-defined constraints stipulate that the first computational algorithm be executed during a single continuous period of time.

19. A scheduling computing device (104) comprising a memory (106) and a processor (108) in communication with the memory (106) and a plurality of data processing computing devices (102), each of the data processing computing devices (102) configured to execute one or more computational algorithms, the processor (108) comprising: receiving computational task data defining a computational task to be performed; retrieving site data corresponding to each of the plurality of data processing computing devices (102), the site data specifying the carbon dioxide emissions expected to be associated with the use of each of the plurality of data processing computing devices (102); selecting, based on the computing task data and the site data, i) a first computing algorithm for executing the computing task, ii) a first data processing computing device (102) among the plurality of data processing computing devices (102), and iii) at least one time period for executing the first computing algorithm by the first data processing computing device (102), wherein the first computing algorithm, the first data processing computing device (102), and the at least one time period are selected to facilitate a reduction in carbon dioxide emissions associated with execution of the computing algorithm; instructing said first data processing computing device (102) to execute said first computing algorithm during said at least one period of time; The scheduling computing device (104) is further configured to:

20. 20. The scheduling computing device (104) of claim 19, wherein the computing task is a machine learning (ML) task and the first computing algorithm is an ML algorithm.

Images