Data center job scheduling using machine learning — Nvidia Corporation (US12206748B2)

ABSTRACT

Abstract

A method includes receiving, using a processing device, a first condition associated with an operation at a data center, where the operation at the data center pertains to a first location at the data center, the first location corresponding to a first parameter value. The method further includes providing the first condition as an input to a machine learning model. The method also includes performing one or more reinforcement learning techniques using the machine learning model to cause the machine learning model to output an indication of a final location associated with the operation, where the final location corresponds to a final parameter value that is closer to a target than the first parameter value corresponding to the first location at the data center.

Description

TECHNICAL FIELD

At least one embodiment pertains to a use of machine learning to perform and facilitate job scheduling in a data center. For example, at least one embodiment pertains to technology for data center job scheduling utilizing reinforcement learning, according to various novel techniques described herein.

BACKGROUND

Data centers can store and process data for various purposes. Typically, the data center will include multiple nodes, each node having multiple graphics processing units (GPUs). When the data center receives a job (e.g., operation), the data center will utilize a job scheduler to select a location within the data center for the job. For example, the job scheduler will look for available nodes or GPUs and select a location accordingly—e.g., the job scheduler will look for three (3) available GPUs if a job received at the data center indicates three (3) GPUs are required. Some conventional solutions consider topology information when selecting the location. However, conventional solutions fail to consider thermal conditions or energy conditions when selecting the location in the data center for the job. Accordingly, temperature at various locations in the data center can be high and require long cooling times or the data center can consume additional power.

BRIEF DESCRIPTION OF DRAWINGS

Various embodiments in accordance with the present disclosure will be described with reference to the drawings, in which:

FIGS. 1 A and 1 B are example system utilizing machine learning in job scheduling, in accordance with at least some embodiments;

FIG. 2 illustrates an example system architecture for reinforcement learning, in accordance with at least some embodiments;

FIG. 3 illustrates a flow diagram of an example method for data center job scheduling using machine learning, in accordance with at least some embodiments;

FIG. 4 illustrates a flow diagram of an example method for data center job scheduling using machine learning, in accordance with at least some embodiments;

FIG. 5 A illustrates an inference and/or training logic, in accordance with at least some embodiments.

FIG. 5 B illustrates an inference and/or training logic, in accordance with at least some embodiments.

FIG. 6 illustrates an example data center system, in accordance with at least some embodiments.

FIG. 7 illustrates a computer system, in accordance with at least some embodiments.

FIG. 8 illustrates a computer system, in accordance with at least some embodiments.

FIG. 9 illustrates at least portions of a graphics processor, in accordance with at least some embodiments.

FIG. 10 illustrates at least portions of a graphic processor, in accordance with at least some embodiments.

FIG. 11 illustrates an example data flow diagram for an advanced computing pipeline, in accordance with at least some embodiments.

FIG. 12 illustrates a system diagram for an example system for training, adapting, instantiating and deploying machine learning models in an advanced computing pipeline, in accordance with at least some embodiment.

FIGS. 13 A and 13 B illustrate a data flow diagram for a process to train a machine learning model, as well as client-server architecture to enhance annotation tools with pre-trained annotation models, in accordance with at least some embodiments.

DETAILED DESCRIPTION

Data centers can store and process data for various purposes. For example, data centers can be used for networks that facilitate object (e.g., pedestrian) detection for an automotive where the automotive is trained based on a large input data set from training data stored at the data center. In other examples, the data center can store output data from medical instruments which can later be used for machine learning inferences. In these and other examples, the data center can process and store large amounts of data using multiple nodes, each node including multiple processing units (e.g., GPUs) to store and process data. Typically, the data center can utilize a job scheduler to select locations within the data center when the data center receives a job request. For example, the data center can receive a job request that takes four (4) graphics processing units (GPUs) to execute. The data center can use the job scheduler to select or find four available GPUs within the data center to execute the job request. The job scheduler can place job requests in a queue if there are not enough resources to execute the respective job request.

In conventional solutions, the job scheduler is primarily looking for available nodes and does not consider other factors. For example, the job scheduler can select the first available node, irrespective of other jobs or operations being executed in the data center. In data centers, the temperature can be maintained for each GPU, each node, and for the data center as a whole. Typically, the data center can cool a data center rack (e.g., a structure housing multiple nodes, servers, cables, networking devices, and other data center computing equipment) uniformly. Accordingly, the data center or node temperature is dependent on a job request that is generating the most heat. The job request that generates a maximum amount of heat can also dictate a cooling rate at a data center rack. If the job scheduler allocates a respective job request to the same data center rack, the temperature can increase and take longer to cool off—e.g., the data center rack can be unavailable for additional job requests for a longer duration as the cooling period takes longer. Similarly, if the job scheduler allocates the respective job request to nodes that are adjacent to other nodes already executing a job request, the temperature can increase for both nodes and take longer to cool off Allocating based on just the availability can also negatively impact water cooling, heat dissipation, air flow, a thermal equilibrium of a node, electricity costs, etc. Therefore the data center can perform inefficiently and increase costs.

Advantageously, aspects of the present disclosure can address the deficiencies above and other challenges by scheduling jobs in a data center using reinforcement learning with a machine learning model. For example, a processing device (e.g., job scheduler or executing agent) can provide a machine learning model with a condition (e.g., a thermal condition) and a job request. The processing device can receive or determine the condition based on user input or other requirements of the data center. The processing device can receive a condition representing a thermal or energy constraint or target for the data center—e.g., the processing device can receive a condition that indicates to reduce a temperature of the data center as a whole as much as possible when executing a respective job request. The processing device can receive a condition that represents at least one of a cooling condition for a node or GPU of the data center, a total energy consumption of the data center, water cooling conditions, heat dissipation conditions, air flow conditions, temperature conditions, location (e.g., location of adjacent operations, or location of all operations currently being executed at the data center) conditions, thermal equilibrium conditions, electricity consumption conditions, electricity costs conditions, a time (e.g., time of day or a duration associated with performing the job request) condition, etc. The processing device can also receive information regarding the location of other operations being executed across the data center. The machine learning model can determine (e.g., predict) a location within the data center that satisfies (or is the closest to satisfying) a target parameter associated with the condition. For example, the machine learning model can determine a location within the data center that will reduce energy consumption or energy costs the most. In other examples, the machine learning model can determine a location within the data center that will cause the least temperature increase or least cooling time. In that, the job scheduler can use the machine learning model to select locations for a job request that reduces a carbon footprint of the data center—e.g., the job scheduler can factor in conditions other than availability when selecting the location for the job request.

The processing device can generate a reward (e.g., parameter) that indicates a net change in a respective condition for a respective job—e.g., the processing device can generate a reward that indicates a net difference in temperature from executing the job request at a location rather than a second location. The reward can be used to update a training of the machine learning model—e.g., the machine learning model can undergo reinforcement learning based on the received reward. For example, the machine learning model can receive the reward, determine the selected location decreases energy consumption, and the processing device can update the machine learning model accordingly. The updated machine learning model can then predict a different location to execute the job request at, receive a second reward, update accordingly, and so forth. This process can be repeated, and the reinforcement learning of the location selection using the machine learning model can continue until a determination is made that no further improvements are being output (e.g., there are no other locations that will be closer to the target than a previous location predicted by the machine learning model)—e.g., until the job scheduler is optimized to select a location given the condition and job request. Additionally, the processing device can select the location based on the output of the machine learning model, determine an actual parameter (e.g., reward), and update the machine learning model if the actual results differ from the machine learning model prediction.

As the data center can receive any number of job requests, each having a different condition, the job scheduler can select one or more reinforcement learning techniques for each unique condition received—e.g., the job scheduler can select a first reinforcement learning technique for optimizing thermal conditions and a second reinforcement learning for a optimizing energy consumption. Accordingly, the data center can be optimized for any number of conditions as requested by a user. For example, the job scheduler can receive a condition associated with a predetermined water temperature, where the water can be used for other tasks such as melting snow on roads leading to the data center. In such examples, the processing device can select one or more reinforcement learning models associated with selecting a node in the data center that satisfies the predetermined water temperature.

By scheduling jobs using reinforcement learning (e.g., machine learning), the data center can conform to any number of conditions provided by a user—e.g., the job scheduler is not limited to selecting a location based on just availability. Additionally, the data center can perform better when temperature and energy consumption are considered and optimized for while selecting locations for jobs. For example, the job scheduler can avoid selecting adjacent nodes for job requests to reduce temperature and cooling times, allowing the data center to process more job requests faster.

FIG. 1 A is a block diagram of a system 100 implementing machine learning in job scheduling, according to at least one embodiment. The system 100 can include a data center 110 coupled to a network 103 . In some embodiments, the system 100 can include a client device 124 coupled with the network 103 .

The data center 110 can include a rack 112 of one or more computing systems 114 ( 1 )- 114 (N), where N is a positive integer equal to or greater than zero. Each computing system 114 can include a computing device 116 and a service processor 120 . In at least one embodiment, the computing device 116 can be considered a node. In other embodiments, multiple computing devices 116 can be considered a node—e.g., a node can include one or more computing devices 116 . In some embodiments, the computing device 116 can be an example of a graphics processing unit (GPU) or central processing unit (CPU). Although one computing device 116 is shown for each computing system 116 , it should be noted each computing system 114 can include any number of computing devices 116 greater than one (1). In at least one embodiment, the service processor 120 is a baseboard management controller (BMC). The BMC can be part of an IPMI type interface and can be located on a circuit board (e.g., motherboard) of the computing device 116 being monitored. The BMC can include one or more sensors that are operatively coupled to the computing device

TECHNICAL FIELD

At least one embodiment pertains to a use of machine learning to perform and facilitate job scheduling in a data center. For example, at least one embodiment pertains to technology for data center job scheduling utilizing reinforcement learning, according to various novel techniques described herein.

BACKGROUND

Data centers can store and process data for various purposes. Typically, the data center will include multiple nodes, each node having multiple graphics processing units (GPUs). When the data center receives a job (e.g., operation), the data center will utilize a job scheduler to select a location within the data center for the job. For example, the job scheduler will look for available nodes or GPUs and select a location accordingly—e.g., the job scheduler will look for three (3) available GPUs if a job received at the data center indicates three (3) GPUs are required. Some conventional solutions consider topology information when selecting the location. However, conventional solutions fail to consider thermal conditions or energy conditions when selecting the location in the data center for the job. Accordingly, temperature at various locations in the data center can be high and require long cooling times or the data center can consume additional power.

BRIEF DESCRIPTION OF DRAWINGS

Various embodiments in accordance with the present disclosure will be described with reference to the drawings, in which:

FIGS. 1 A and 1 B are example system utilizing machine learning in job scheduling, in accordance with at least some embodiments;

FIG. 2 illustrates an example system architecture for reinforcement learning, in accordance with at least some embodiments;

FIG. 3 illustrates a flow diagram of an example method for data center job scheduling using machine learning, in accordance with at least some embodiments;

FIG. 4 illustrates a flow diagram of an example method for data center job scheduling using machine learning, in accordance with at least some embodiments;

FIG. 5 A illustrates an inference and/or training logic, in accordance with at least some embodiments.

FIG. 5 B illustrates an inference and/or training logic, in accordance with at least some embodiments.

FIG. 6 illustrates an example data center system, in accordance with at least some embodiments.

FIG. 7 illustrates a computer system, in accordance with at least some embodiments.

FIG. 8 illustrates a computer system, in accordance with at least some embodiments.

FIG. 9 illustrates at least portions of a graphics processor, in accordance with at least some embodiments.

FIG. 10 illustrates at least portions of a graphic processor, in accordance with at least some embodiments.

FIG. 11 illustrates an example data flow diagram for an advanced computing pipeline, in accordance with at least some embodiments.

FIG. 12 illustrates a system diagram for an example system for training, adapting, instantiating and deploying machine learning models in an advanced computing pipeline, in accordance with at least some embodiment.

FIGS. 13 A and 13 B illustrate a data flow diagram for a process to train a machine learning model, as well as client-server architecture to enhance annotation tools with pre-trained annotation models, in accordance with at least some embodiments.

DETAILED DESCRIPTION

Data centers can store and process data for various purposes. For example, data centers can be used for networks that facilitate object (e.g., pedestrian) detection for an automotive where the automotive is trained based on a large input data set from training data stored at the data center. In other examples, the data center can store output data from medical instruments which can later be used for machine learning inferences. In these and other examples, the data center can process and store large amounts of data using multiple nodes, each node including multiple processing units (e.g., GPUs) to store and process data. Typically, the data center can utilize a job scheduler to select locations within the data center when the data center receives a job request. For example, the data center can receive a job request that takes four (4) graphics processing units (GPUs) to execute. The data center can use the job scheduler to select or find four available GPUs within the data center to execute the job request. The job scheduler can place job requests in a queue if there are not enough resources to execute the respective job request.

In conventional solutions, the job scheduler is primarily looking for available nodes and does not consider other factors. For example, the job scheduler can select the first available node, irrespective of other jobs or operations being executed in the data center. In data centers, the temperature can be maintained for each GPU, each node, and for the data center as a whole. Typically, the data center can cool a data center rack (e.g., a structure housing multiple nodes, servers, cables, networking devices, and other data center computing equipment) uniformly. Accordingly, the data center or node temperature is dependent on a job request that is generating the most heat. The job request that generates a maximum amount of heat can also dictate a cooling rate at a data center rack. If the job scheduler allocates a respective job request to the same data center rack, the temperature can increase and take longer to cool off—e.g., the data center rack can be unavailable for additional job requests for a longer duration as the cooling period takes longer. Similarly, if the job scheduler allocates the respective job request to nodes that are adjacent to other nodes already executing a job request, the temperature can increase for both nodes and take longer to cool off Allocating based on just the availability can also negatively impact water cooling, heat dissipation, air flow, a thermal equilibrium of a node, electricity costs, etc. Therefore the data center can perform inefficiently and increase costs.

Advantageously, aspects of the present disclosure can address the deficiencies above and other challenges by scheduling jobs in a data center using reinforcement learning with a machine learning model. For example, a processing device (e.g., job scheduler or executing agent) can provide a machine learning model with a condition (e.g., a thermal condition) and a job request. The processing device can receive or determine the condition based on user input or other requirements of the data center. The processing device can receive a condition representing a thermal or energy constraint or target for the data center—e.g., the processing device can receive a condition that indicates to reduce a temperature of the data center as a whole as much as possible when executing a respective job request. The processing device can receive a condition that represents at least one of a cooling condition for a node or GPU of the data center, a total energy consumption of the data center, water cooling conditions, heat dissipation conditions, air flow conditions, temperature conditions, location (e.g., location of adjacent operations, or location of all operations currently being executed at the data center) conditions, thermal equilibrium conditions, electricity consumption conditions, electricity costs conditions, a time (e.g., time of day or a duration associated with performing the job request) condition, etc. The processing device can also receive information regarding the location of other operations being executed across the data center. The machine learning model can determine (e.g., predict) a location within the data center that satisfies (or is the closest to satisfying) a target parameter associated with the condition. For example, the machine learning model can determine a location within the data center that will reduce energy consumption or energy costs the most. In other examples, the machine learning model can determine a location within the data center that will cause the least temperature increase or least cooling time. In that, the job scheduler can use the machine learning model to select locations for a job request that reduces a carbon footprint of the data center—e.g., the job scheduler can factor in conditions other than availability when selecting the location for the job request.

The processing device can generate a reward (e.g., parameter) that indicates a net change in a respective condition for a respective job—e.g., the processing device can generate a reward that indicates a net difference in temperature from executing the job request at a location rather than a second location. The reward can be used to update a training of the machine learning model—e.g., the machine learning model can undergo reinforcement learning based on the received reward. For example, the machine learning model can receive the reward, determine the selected location decreases energy consumption, and the processing device can update the machine learning model accordingly. The updated machine learning model can then predict a different location to execute the job request at, receive a second reward, update accordingly, and so forth. This process can be repeated, and the reinforcement learning of the location selection using the machine learning model can continue until a determination is made that no further improvements are being output (e.g., there are no other locations that will be closer to the target than a previous location predicted by the machine learning model)—e.g., until the job scheduler is optimized to select a location given the condition and job request. Additionally, the processing device can select the location based on the output of the machine learning model, determine an actual parameter (e.g., reward), and update the machine learning model if the actual results differ from the machine learning model prediction.

As the data center can receive any number of job requests, each having a different condition, the job scheduler can select one or more reinforcement learning techniques for each unique condition received—e.g., the job scheduler can select a first reinforcement learning technique for optimizing thermal conditions and a second reinforcement learning for a optimizing energy consumption. Accordingly, the data center can be optimized for any number of conditions as requested by a user. For example, the job scheduler can receive a condition associated with a predetermined water temperature, where the water can be used for other tasks such as melting snow on roads leading to the data center. In such examples, the processing device can select one or more reinforcement learning models associated with selecting a node in the data center that satisfies the predetermined water temperature.

By scheduling jobs using reinforcement learning (e.g., machine learning), the data center can conform to any number of conditions provided by a user—e.g., the job scheduler is not limited to selecting a location based on just availability. Additionally, the data center can perform better when temperature and energy consumption are considered and optimized for while selecting locations for jobs. For example, the job scheduler can avoid selecting adjacent nodes for job requests to reduce temperature and cooling times, allowing the data center to process more job requests faster.

FIG. 1 A is a block diagram of a system 100 implementing machine learning in job scheduling, according to at least one embodiment. The system 100 can include a data center 110 coupled to a network 103 . In some embodiments, the system 100 can include a client device 124 coupled with the network 103 .

The data center 110 can include a rack 112 of one or more computing systems 114 ( 1 )- 114 (N), where N is a positive integer equal to or greater than zero. Each computing system 114 can include a computing device 116 and a service processor 120 . In at least one embodiment, the computing device 116 can be considered a node. In other embodiments, multiple computing devices 116 can be considered a node—e.g., a node can include one or more computing devices 116 . In some embodiments, the computing device 116 can be an example of a graphics processing unit (GPU) or central processing unit (CPU). Although one computing device 116 is shown for each computing system 116 , it should be noted each computing system 114 can include any number of computing devices 116 greater than one (1). In at least one embodiment, the service processor 120 is a baseboard management controller (BMC). The BMC can be part of an IPMI type interface and can be located on a circuit board (e.g., motherboard) of the computing device 116 being monitored. The BMC can include one or more sensors that are operatively coupled to the computing device 116 or integrated within the computing device 116 . The sensors of a BMC measure internal physical variables such as temperature, humidity, power-supply voltage, fan speeds, communications parameters, and operating system (OS) functions. The BMC can provide a way to manage a computer that may be powered off or otherwise unresponsive. The service processor 120 provides out-of-band functionality by collecting the power consumption data of the computing device 116 independently from the computing device's CPU, firmware, and OS. The service processor 120 can provide the power consumption data via a network connection 122 independent from a primary network connection 118 of the computing device 116 . The service processor 120 can use the network connection 122 to the hardware itself rather than the OS or login shell to manage the computing device 116 , even if the computing device 116 is powered off or otherwise unresponsive. Although one rack 112 is illustrated, the data center 110 can include any number of racks 112 equal or greater than one (1).

In at least one embodiment, the rack 112 can be coupled with or include a data center manager 128 —e.g., the data center manager 128 can be coupled with multiple racks 112 or each rack 112 can include a data center manager 128 . In some embodiments, the data center manager 128 can manage computing device 116 of the rack 112 and computing systems 114 . In some embodiments, the data center manager 128 can include a service processor 130 and be connected to the network via network connection 132 . In at least one embodiment, the data center manager 128 includes a scheduler 160 with a machine learning (ML) model 104 —e.g., the data center manager 128 includes a ML supported scheduler 160 .

FIG. 1 B is a block diagram of a data center manager 128 in a data center 110 implementing machine learning to schedule operations or jobs, according to at least one embodiment. The data center manager 128 can include an interface 150 , a network manager 155 , a scheduler 160 , and a data store 165 . In at least one embodiment, the data center manager 128 is configured to manage the racks 112 —e.g., schedule operations. In some embodiments, the data center manager 128 can detect and respond to events occurring at the racks 112 —e.g., starting a new cluster for an incoming operation or detecting a current cluster or rack 112 is failing.

In at least one embodiment, the interface 150 can enable the data center manager 128 to communicate commands, operations, information, etc., to the racks 112 . For example the interface 150 can enable the data center manager 128 to schedule operations at the racks 112 . In at least one embodiment, the data center manager 128 can monitor and detect when nodes (e.g., computing devices 116 ) are down or creates clusters or pods to execute incoming operations. In some embodiments, the interface 150 is an example of an application programming interface (API) server. The interface can manage the network manager 155 , the scheduler 160 , and the data store 165 .

In at least one embodiment, network manager 155 can link the data center manager 128 and the racks 112 to a cloud network or cloud environment. For example, the network manager 155 can link a cluster to a cloud provider API. In some embodiments, the network manager 155 can check nodes in the cloud network, set up routes in the cloud network, or create, update, and delete cloud network load balancers.

In some embodiments, scheduler 160 is configured to schedule operations at nodes or computing device 116 of the racks 112 . In at least one embodiment, the scheduler 160 includes code, a component, or a configurable plug-in that indicates how to allocate resources—e.g., indicates a basis for selecting nodes for performing a job. In some embodiments, the scheduler 160 includes a processing device 102 that executes a machine learning (ML) model 104 trained to schedule jobs according to thermal conditions provided. For example, the processing device 102 can execute an ML model 104 to schedule operations in a manner that reduces energy consumption, reduces a temperature of a respective node, rack 112 , or the data center 110 or increases cooling efficiency. That is, the ML model 104 can schedule operations that optimize the thermal conditions of the data center 110 . For example, the scheduler 160 can receive an operation that will execute at three (3) nodes of the data center 110 . The ML model 104 can be trained to select three (3) available nodes that will execute the operation with the least amount of energy consumption or cause the least amount of temperature increase in the data center 110 . In some embodiments, the ML model 104 can be trained to optimize for at least one of the following thermal conditions; water cooling, heat dissipation, or air flow direction—e.g., where airflow direction describes whether a respective compute nodes are using air that is recirculated datacenter air or filtered air from outside the datacenter. Additional thermal conditions and constraints are discussed with reference to FIGS. 2 - 4 . In at least one embodiment, the ML model 104 can be trained by a set of training data that includes at least one of a water cooling parameter, heat dissipation parameter, air flow direction parameter, a type of operation to execute, an amount of nodes corresponding to executing the operation, and an amount of energy consumed. That is, the ML model 104 can be trained by mapping between the parameters and the thermal conditions and energy consumed or cooling provided. In one embodiment, the processing device 102 can calculate floating point operations per second (e.g., FLOPS) to measure computer performance using the parameters—e.g., the processing device 102 can calculate performance of the nodes given respective parameter values (e.g., given the node is operating with a first water cooling parameter (e.g., the node is cooling water at a respective rate). The ML model 104 can be trained to determine the energy consumed or cooling provided (e.g., for the node or for the data center) for the respective parameter—e.g., the energy consumed if the node is operating with the first water cooling parameter. In at least one embodiment, the processing device 102 can receive information regarding historical performance of a respective node—e.g., information regarding how much energy is consumed or a carbon footprint of utilizing the node. In at least one embodiment, the processing device 102 can receive information regarding historical performance of a respective GPU in a node. Additional details regarding the training is described with reference to FIGS. 2 - 4 . It should be noted that the ML model 104 can utilize any machine learning model—e.g., a deep (sequential) network, a single level of linear or non-linear operations, reinforcement learning, etc. For convenience, the remainder of the disclosure will refer to the implementation as a reinforcement learning model, even though some implementations might employ other types of learning machine models instead of, or in addition to, a reinforcement learning model. In some embodiments, the reinforcement learning model or algorithm can include a deep Q-network (DQN), a deep deterministic policy gradient (DDPG), Markov decision process (MDP), etc. In at least one embodiment, the ML model 104 is trained in a simulated environment as described with reference to FIG. 2 .

In at least one embodiment, ML model 104 can be trained to enable the scheduler 160 to make energy aware decisions. In at least one embodiment, the ML model 104 can be trained to use an outside air temperature (e.g., a temperature parameter) to schedule operations. For example, if the outside air temperature is less than 35 degrees Celsius (35° C.), that air could be filtered and directed to the compute nodes for cooling. In at least one embodiment, the node is configured to emit exhaust that is directed to outside the data center 110 . In such embodiments, the data center 110 consumes energy (e.g., there is an energy charge) to filter and blow the air from outside to a front of the racks 112 . In at least one embodiment, the data center 110 also consumes energy at the compute node based on the inlet air temperature provided. For example, as the input air temperature increases, racks 112 and nodes can consume additional energy due to increased fan power and additional leakage of the silicon at elevated temperatures. In at least one embodiment, the data center 110 can compare the energy consumed for filtering outside air with energy consumed to create 20 degrees Celsius (20° C.) Air with a traditional data center CRAH(Computer Room Air Handler) and compressor unit. For example, the data center 110 can consume additional power at a potentially lower temperature. Accordingly, the ML model 104 can be trained to select a node for an operation based on a performance of a respective node filtering outside air or using the CRAH.

In at least one embodiment, the ML model 104 can be trained to enable the scheduler 160 to select nodes for an operation to create or warm water temperatures. For example, the scheduler 160 can schedule more operations at a rack 112 if there is request or indication to have warm water for district heating applications. In at least one embodiment, water control is typically at a rack 112 or at a CDU (Coolant Distribution Unit) level. In at least one embodiment, it is difficult to isolate water flow at a compute node level. In some embodiments, the data center can generate higher output water temperature if all nodes in a rack 112 are active. Accordingly, if a scheduler 160 can optimize warm water for energy reuse, the scheduler 160 can schedule operations at one rack 112 to increase a number of active node at the rack 112 or CDU domain.

In at least one embodiment, the processing device 102 can train the ML model 104 using data from data store 165 . For example, the processing device 102 can utilize reinforcement learning techniques to train the ML model 104 . In at least one embodiment, the processing logic 102 can provide training data and a current indication of a current location (e.g., a current state) for performing a respective operation and train the ML model 104 to predict thermal and energy usage of performing the respective operation at a different location. In such embodiments, the processing logic 102 can also train the ML model to select an optimal location (e.g., a desired location) based on the thermal and energy conditions provided—e.g., the ML model 104 can be trained to predict an optimal location that will consume a least amount of energy to perform the respective operation. In some embodiments, the data store 165 can be outside the data center manager 128 —e.g., the data store 165 can be included in a rack 112 or be remote and coupled to the data center manager 128 via network 103 . The processing device 102 can analyze the thermal information data at data store 165 and utilize the ML model 104 to predict a node to execute an operation at that is the most optimal for a given thermal condition. The data center manager 128 can include a collection service for collecting the thermal information for one or more computing devices 116 or nodes in the data center 110 . In at least one embodiment, the collection service is used by the data center manager 128 for collecting thermal information for training the ML model 104 and for use by the ML model 104 after training. In some embodiments, the processing device 102 can receive the thermal information from the racks 112 . The processing device 102 or the interface 150 can store the thermal information in the data store 165 . By utilizing the ML model 104 , the scheduler 160 can schedule operations that optimize for thermal conditions and improve the efficiency of the data center 110 .

FIG. 2 illustrates a system 200 for performing reinforcement learning to predict a node most optimal to execute an operation at based on thermal conditions provided, according to at least one embodiment. In some embodiments, the system 200 includes an agent 202 (e.g., an actor) and an environment 204 —e.g., a simulated environment. The agent 202 trains one or more machine learning models (e.g., ML model(s) 104 ) which can be examples of a Q-network (e.g., or a deep Q-network (DQN)). The system 200 can be located in the processing device 102 as described with reference to FIG. 1 —e.g., within the scheduler 160 .

Reinforcement learning (RL) is a class of algorithms applicable to sequential decision making tasks. In some embodiments, the RL makes use of the Markov Decision Process (MDP) formalism where an agent 202 attempts to optimize a function in its environment 204 . An MDP can be completely described by a state space S (with states s∈S), an action space A (a∈A), a transition function T: S×A→S and a reward function R: S×A→

In an MDP, an episode evolves over discrete time steps t=0, 1, 2, . . . , n, where the agent 202 observes a state s t ( 206 ) and responds with an action a t ( 210 ) using a policy π(a t |s t ). The environment 204 provides to the agent 202 the next state s t+1 ˜T (s t , a t ) 212 and the reward r t =R(s t , a t ) 214 . The agent 202 is tasked with maximizing the return (cumulative future rewards) by learning an optimal policy π*.

A Q-network may be trained via a process referred to as Q-learning. Q learning is a reinforcement learning process that causes a Q-network to perform a sequence of actions that will eventually generate a maximum total reward. This total reward is also called the Q-value. A function for computing the Q value may be as follows:

Q ( s t ,a t )← r t +γmax a Q ( s t+1 ,a )

The above equation states that the Q-value yielded from being at state s t ( 206 ) and performing action a t ( 210 ) is the immediate reward r(s t , a t ) ( 214 ) plus the highest Q-value possible from the next state s t+1 ( 212 ), state where γ is the discount factor which controls the contribution of rewards further in the future. The recursive definition of Q-functions allows the expression of Q(s t+1 , a t+1 ) to be unrolled into future states, as follows:

Q ( s t ,a t )= r t +γr t+1 + . . . +γ n−1 r t+n−1 +γ n Q ( s t+n ,a t+n )

The Q-network learns to predict Q(s t , a t ) by performing the following update step:

Q ⁡ (

s t

,

a t

)

←

Q ⁡ (

s t

,

a t

)

+

α [

r

t + 1

+

γ

max a

Q ⁡ (

S

t + 1

, a

)

-

Q ⁡ (

s t

,

a t </

CLAIMS

Claims ( 20 )

What is claimed is:

1. A method comprising:

receiving, using a processing device, a first optimization condition associated with an operation at a data center, wherein the operation at the data center pertains to a first location at the data center, the first location corresponding to a first environmental parameter value;

providing the first optimization condition and the first environmental parameter value as an input to a reinforcement learning agent comprising a machine learning model;

obtaining an output of the reinforcement learning agent, the output comprising an indication of a final location at the data center, wherein the final location is associated with the operation, and wherein the final location corresponds to a final environmental parameter value that is closer to a target associated with the first optimization condition than the first environmental parameter value corresponding to the first location at the data center; and

providing a reward to the reinforcement learning agent, the reward corresponding to a comparison of the first environmental parameter value and the final environmental parameter value.

2. The method of claim 1 , further comprising:

re-training the machine learning model of the reinforcement learning agent based on the reward;

obtaining a second output of the reinforcement learning agent, the second output comprising an indication of a second location of the operation at the data center, wherein the second location corresponds to a second environmental parameter value; and

providing a second reward to the reinforcement learning agent, the second reward corresponding to a comparison of the first environmental parameter value and the second environmental parameter value.

3. The method of claim 2 , wherein re-training the machine learning model of the reinforcement learning agent based on the reward comprises:

determining a Q-value for the reinforcement learning agent based on the reward; and

re-training the machine learning model using the Q-value, wherein the machine learning model approximates a Q-value function of the reinforcement learning agent.

4. The method of claim 1 , further comprising:

executing the operation at the final location in response to the reinforcement learning agent outputting the final location.

5. The method of claim 4 , further comprising:

receiving, using the processing device, a second environmental parameter value associated with executing the operation at the final location;

determining that the second environmental parameter value is different than the final environmental parameter value in response to receiving the second environmental parameter value; and

updating the reward to correspond to a comparison of the first environmental parameter value and the second environmental parameter value.

6. The method of claim 1 ,

wherein the machine learning model is one of a plurality of machine learning models of the reinforcement learning agent, wherein each model of the plurality of machine learning models is trained in association with a respective optimization condition, and wherein the first optimization condition causes the reinforcement learning agent to select the machine learning model from the plurality of machine learning models to generate the output of the reinforcement learning agent.

7. The method of claim 1 , further comprising:

providing one or more additional optimization conditions of a plurality of optimization conditions to the reinforcement learning agent, wherein the reinforcement learning agent is trained to output the final location based on receiving the first optimization condition and the one or more additional optimization conditions of the plurality of optimization conditions.

8. The method of claim 1 , wherein the first optimization condition represents at least one of:

a cooling condition associated with the operation, total data center energy consumption condition, water cooling condition, heat dissipation condition, air flow condition, condition associated with temperature, condition associated with weather forecasts, condition associated with location of adjacent operations, condition associated with location of all operations currently executing at the data center, thermal equilibrium conditions, electricity consumption condition, condition associated with electricity costs, a condition associated with a time associated with the operation, condition associated with a type of operation, or a condition associated with an amount of energy associated with executing the operation.

9. The method of claim 1 , further comprising:

generating training data comprising an amount of energy associated with executing the operation and at least one of a cooling parameter associated with the operation, total data center energy consumption parameter, water cooling parameter, heat dissipation parameter, air flow parameter, parameter associated with temperature, parameter associated with weather forecasts, parameter associated with location of adjacent operations, parameter associated with location of all operations currently executing at the data center, thermal equilibrium parameter, electricity consumption parameter, parameter associated with electricity costs, a parameter associated with a time associated with the operation, data transfer associated with the operation, or type of operation; and

training the machine learning model of the reinforcement learning agent using the training data.

10. A system comprising:

a processing device to:

receive a first optimization condition associated with an operation at a data center, wherein the operation at the data center pertains to a first location at the data center, the first location corresponding to a first environmental parameter value;

provide the first optimization condition and the first environmental parameter value as an input to a reinforcement learning agent comprising a machine learning model;

obtaining an output of the reinforcement learning agent, the output comprising an indication of a final location at the data center, wherein the final location is associated with the operation, and wherein the final location corresponds to a final environmental parameter value that is closer to a target associated with the first optimization condition than the first environmental parameter value corresponding to the first location at the data center; and

providing a reward to the reinforcement learning agent, the reward corresponding to a comparison of the first environmental parameter value and the final environmental parameter value.

11. The system of claim 10 , wherein the processing device is further to:

re-train the machine learning model of the reinforcement learning agent based on the reward;

obtain a second output of the reinforcement learning agent, the second output comprising an indication of a second location of the operation at the data center, wherein the second location corresponds to a second environmental parameter value; and

providing a second reward to the reinforcement learning agent, the second reward corresponding to a comparison of the first environmental parameter value and the second environmental parameter value.

12. They system of claim 11 , wherein to re-train the machine learning model of the reinforcement learning agent based on the reward, the processing device is further to:

determine a Q-value for the reinforcement learning agent based on the reward; and

re-train the machine learning model using the Q-value, wherein the machine learning model approximates a Q-value function of the reinforcement learning agent.

13. The system of claim 10 , wherein the processing device is further to:

execute the operation at the final location in response to the reinforcement learning agent outputting the final location.

14. The system of claim 13 , wherein the processing device is further to:

receive a second environmental parameter value associated with executing the operation at the final location;

determine the second environmental parameter value is different than the final environmental parameter value in response to receiving the second environmental parameter value; and

update the reward to correspond to a comparison of the first environmental parameter value and the second environmental parameter value.

15. The system of claim 10 ,

wherein the machine learning model is one of a plurality of machine learning models of the reinforcement learning agent, wherein each model of the plurality of machine learning models is trained in association with a respective optimization condition, and wherein the first optimization condition causes the reinforcement learning agent to select the machine learning model from the plurality of machine learning models to generate the output of the reinforcement learning agent.

16. The system of claim 10 , wherein the processing device is to:

provide one or more additional optimization conditions of a plurality of optimization conditions to the reinforcement learning agent, wherein the reinforcement learning agent is trained to output the final location based on receiving the first optimization condition and the one or more additional optimization conditions of the plurality of optimization conditions.

17. The system of claim 10 , wherein the first optimization condition represents at least one of:

a cooling condition associated with the operation, total data center energy consumption condition, water cooling condition, heat dissipation condition, air flow condition, condition associated with temperature, condition associated with weather forecasts, condition associated with location of adjacent operations, condition associated with location of all operations currently executing at the data center, thermal equilibrium conditions, electricity consumption condition, condition associated with electricity costs, or a condition associated with a time associated with the operation, condition associated with a type of operation, or a condition associated with an amount of energy associated with executing the operation.

18. The system of claim 10 , wherein the processing device is further to:

generate training data comprising an amount of energy associated with executing the operation and at least one of a cooling parameter associated with the operation, total data center energy consumption parameter, water cooling parameter, heat dissipation parameter, air flow parameter, parameter associated with temperature, parameter associated with weather forecasts, parameter associated with location of adjacent operations, parameter associated with location of all operations currently executing at the data center, thermal equilibrium parameter, electricity consumption parameter, parameter associated with electricity costs, or a parameter associated with a time associated with the operation, data transfer associated with the operation, or type of operation; and

train the machine learning model of the reinforcement learning agent using the training data.

19. A non-transitory computer-readable medium storing instructions thereon, wherein the instructions, when executed by a processing device, cause the processing device to:

receive a first optimization condition associated with an operation at a data center, wherein the operation at the data center pertains to a first location at the data center, the first location corresponding to a first environmental parameter value;

provide the first optimization condition and the first environmental parameter value as an input to a reinforcement learning agent comprising a machine learning model;

obtain an output of the reinforcement learning agent, the output comprising an indication of a final location at the data center, wherein the final location is associated with the operation, and wherein the final location corresponds to a final environmental parameter value that is closer to a target associated with the first optimization condition than the first environmental parameter value corresponding to the first location at the data center; and

provide a reward to the reinforcement learning agent, the reward corresponding to a comparison of the first environmental parameter value and the final environmental parameter value.

20. The non-transitory computer-readable medium of claim 19 , wherein the first optimization condition represents at least one of:

a cooling condition associated with the operation, total data center energy consumption condition, water cooling condition, heat dissipation condition, air flow condition, condition associated with temperature, condition associated with weather forecasts, condition associated with location of adjacent operations, condition associated with location of all operations currently executing at the data center, thermal equilibrium conditions, electricity consumption condition, condition associated with electricity costs, or a condition associated with a time associated with the operation, condition associated with a type of operation, or a condition associated with an amount of energy associated with executing the operation.

US17/958,139

2022-09-30

2022-09-30

Data center job scheduling using machine learning

Active

US12206748B2

( en )

Priority Applications (1)

Application Number

Priority Date

Filing Date

Title

US17/958,139

US12206748B2

( en )

2022-09-30

2022-09-30

Data center job scheduling using machine learning

Applications Claiming Priority (1)

Application Number

Priority Date

Filing Date

Title

US17/958,139

US12206748B2

( en )

2022-09-30

2022-09-30

Data center job scheduling using machine learning

Publications (2)

Publication Number

Publication Date

US20240129380A1

US20240129380A1 ( en )

2024-04-18

US12206748B2

true

US12206748B2 ( en )

2025-01-21

Family

ID=90625908

Family Applications (1)

Application Number

Title

Priority Date

Filing Date

US17/958,139

Active

US12206748B2

( en )

2022-09-30

2022-09-30

Data center job scheduling using machine learning

Country Status (1)

Country

Link

US

( 1 )

US12206748B2

( en )

Families Citing this family (12)

* Cited by examiner, † Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

CN112508398B

( en )

*

2020-12-04

2022-09-16

北京邮电大学

Dynamic production scheduling method and device based on deep reinforcement learning and electronic equipment

US12518326B2

( en )

*

2022-10-04

2026-01-06

Bank Of America Corporation

Energy optimization platform using artificial intelligence for container routing

US20240143873A1

( en )

*

2022-10-26

2024-05-02

Dell Products L.P.

Method and System to Determine Optimal Rack Space Utilization

US20240378421A1

( en )

*

2023-05-09

2024-11-14

Optum, Inc.

Deep learning architectures for reducing incidental truncation bias, and systems and methods of use

US12632925B2

( en )

*

2023-06-30

2026-05-19

International Business Machines Corporation

Parameter-efficient and resolution-robust network architectures for image-to-image translation

US20250088452A1

( en )

*

2023-09-11

2025-03-13

Juniper Networks, Inc.

Energy aware weighted ecmp

US12699783B2

( en )

2024-01-24

2026-08-04

Dell Products L.P.

Encrypting data at rest and data in motion with trustworthy energy awareness

US12597947B2

( en )

2024-01-24

2026-04-07

Dell Products L.P.

Data compression with trustworthy energy awareness

US20250238156A1

( en )

*

2024-01-24

2025-07-24

Dell Products L.P.

Data migration and asynchronous replication with trustworthy energy awareness

US12705178B2

( en )

2024-01-24

2026-08-11

Dell Products L.P.

Data prefetching operations with trustworthy energy awareness

CN119847722A

( en )

*

2024-05-14

2025-04-18

中关村泛联移动通信技术创新应用研究院

Calculation segmentation map adjustment method, device, equipment, storage medium and program product

CN119544717A

( en )

*

2024-11-26

2025-02-28

西安交通大学

A data center cluster online scheduling method and system

Citations (28)

* Cited by examiner, † Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US20050228618A1

( en )

*

2004-04-09

2005-10-13

Patel Chandrakant D

Workload placement among data centers based on thermal efficiency

US7447920B2

( en )

*

2004-08-31

2008-11-04

Hewlett-Packard Development Company, L.P.

Workload placement based on thermal considerations

US7644051B1

( en )

*

2006-07-28

2010-01-05

Hewlett-Packard Development Company, L.P.

Management of data centers using a model

US8489745B2

( en )

*

2010-02-26

2013-07-16

International Business Machines Corporation

Optimizing power consumption by dynamic workload adjustment

US20140180978A1

( en )

*

2012-12-21

2014-06-26

InsideSales.com, Inc.

Instance weighted learning machine learning model

US9069103B2

( en )

*

2010-12-17

2015-06-30

Microsoft Technology Licensing, Llc

Localized weather prediction through utilization of cameras

US20160087909A1

( en )

*

2014-09-18

2016-03-24

International Business Machines Corporation

Scheduling cost efficient datacenter load distribution

US9348394B2

( en )

*

2010-09-14

2016-05-24

Microsoft Technology Licensing, Llc

Managing computational workloads of computing apparatuses powered by renewable resources

US10216606B1

( en )

*

2016-09-30

2019-02-26

Nautilus Data Technologies, Inc.

Data center management systems and methods for compute density efficiency measurements

US10234833B2

( en )

*

2016-07-22

2019-03-19

Intel Corporation

Technologies for predicting power usage of a data center

US10243879B2

( en )

*

2015-03-13

2019-03-26

Microsoft Technology Licensing, Llc

Intelligent placement within a data center

US10474497B1

( en )

*

2018-11-14

2019-11-12

Capital One Services, Llc

Computing node job assignment using multiple schedulers

US20200037473A1

( en )

*

2018-07-25

2020-01-30

Vmware, Inc.

Methods and apparatus to control power delivery based on predicted power utilization in a data center

US20200133702A1

( en )

*

2018-10-30

2020-04-30

Hewlett Packard Enterprise Development Lp

Virtual workload migrations

US10643121B2

( en )

*

2017-01-19

2020-05-05

Deepmind Technologies Limited

Optimizing data center controls using neural networks

US10791168B1

( en )

*

2018-05-21

2020-09-29

Rafay Systems, Inc.

Traffic aware network workload management system

US10841236B1

( en )

*

2018-03-30

2020-11-17

Electronic Arts Inc.

Distributed computer task management of interrelated network computing tasks

US10956230B2

( en )

*

2018-10-01

2021-03-23

Vmware, Inc.

Workload placement with forecast

US20220012611A1

( en )

*

2018-11-21

2022-01-13

Telefonaktiebolaget Lm Ericsson (Publ)

Method and machine learning manager for handling prediction of service characteristics

US11269395B2

( en )

*

2019-04-25

2022-03-08

Intel Corporation

Technologies for providing adaptive power management in an accelerator sled

US11340951B2

( en )

*

2019-10-23

2022-05-24

Microsoft Technology Licensing, Llc

Estimating an amount of resources that a program is expected to consume in a data center

US11424993B1

( en )

*

2017-05-30

2022-08-23

Amazon Technologies, Inc.

Artificial intelligence system for network traffic flow based detection of service usage policy violations

CN114970358A

( en )

*

2022-06-06

2022-08-30

深圳市证通电子股份有限公司

Data center energy efficiency optimization method and system based on reinforcement learning

US11461210B2

( en )

*

2019-06-26

2022-10-04

Kyndryl, Inc.

Real-time calculation of data center power usage effectiveness

US20230017632A1

( en )

*

2021-07-12

2023-01-19

Vapor IO Inc.

Reducing the environmental impact of distributed computing

US11609784B2

( en )

*

2018-04-18

2023-03-21

Intel Corporation

Method for distributing a computational process, workload distribution device and system for distributing a computational process

US11644882B2

( en )

*

2021-06-02

2023-05-09

Hewlett Packard Enterprise Development Lp

System and method for predicting power usage of network components

US11875191B1

( en )

*

2020-03-02

2024-01-16

Amazon Technologies, Inc.

Energy-optimizing placement of resources in data centers

2022

2022-09-30

US

US17/958,139

patent/US12206748B2/en

active

Active

Patent Citations (29)

* Cited by examiner, † Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US20050228618A1

( en )

*

2004-04-09

2005-10-13

Patel Chandrakant D

Workload placement among data centers based on thermal efficiency

US7447920B2

( en )

*

2004-08-31

2008-11-04

Hewlett-Packard Development Company, L.P.

Workload placement based on thermal considerations

US7644051B1

( en )

*

2006-07-28

2010-01-05

Hewlett-Packard Development Company, L.P.

Management of data centers using a model

US8489745B2

( en )

*

2010-02-26

2013-07-16

International Business Machines Corporation

Optimizing power consumption by dynamic workload adjustment

US9348394B2

( en )

*

2010-09-14

2016-05-24

Microsoft Technology Licensing, Llc

Managing computational workloads of computing apparatuses powered by renewable resources

US9069103B2

( en )

*

2010-12-17

2015-06-30

Microsoft Technology Licensing, Llc

Localized weather prediction through utilization of cameras

US20140180978A1

( en )

*

2012-12-21

2014-06-26

InsideSales.com, Inc.

Instance weighted learning machine learning model

US8788439B2

( en )

*

2012-12-21

2014-07-22

InsideSales.com, Inc.

Instance weighted learning machine learning model

US20160087909A1

( en )

*

2014-09-18

2016-03-24

International Business Machines Corporation

Scheduling cost efficient datacenter load distribution

US10243879B2

( en )

*

2015-03-13

2019-03-26

Microsoft Technology Licensing, Llc

Intelligent placement within a data center

US10234833B2

( en )

*

2016-07-22

2019-03-19

Intel Corporation

Technologies for predicting power usage of a data center

US10216606B1

( en )

*

2016-09-30

2019-02-26

Nautilus Data Technologies, Inc.

Data center management systems and methods for compute density efficiency measurements

US10643121B2

( en )

*

2017-01-19

2020-05-05

Deepmind Technologies Limited

Optimizing data center controls using neural networks

US11424993B1

( en )

*

2017-05-30

2022-08-23

Amazon Technologies, Inc.

Artificial intelligence system for network traffic flow based detection of service usage policy violations

US10841236B1

( en )

*

2018-03-30

2020-11-17

Electronic Arts Inc.

Distributed computer task management of interrelated network computing tasks

US11609784B2

( en )

*

2018-04-18

2023-03-21

Intel Corporation

Method for distributing a computational process, workload distribution device and system for distributing a computational process

US10791168B1

( en )

*

2018-05-21

2020-09-29

Rafay Systems, Inc.

Traffic aware network workload management system

US20200037473A1

( en )

*

2018-07-25

2020-01-30

Vmware, Inc.

Methods and apparatus to control power delivery based on predicted power utilization in a data center

US10956230B2

( en )

*

2018-10-01

2021-03-23

Vmware, Inc.

Workload placement with forecast

US20200133702A1

( en )

*

2018-10-30

2020-04-30

Hewlett Packard Enterprise Development Lp

Virtual workload migrations

US10474497B1

( en )

*

2018-11-14

2019-11-12

Capital One Services, Llc

Computing node job assignment using multiple schedulers

US20220012611A1

( en )

*

2018-11-21

2022-01-13

Telefonaktiebolaget Lm Ericsson (Publ)

Method and machine learning manager for handling prediction of service characteristics

US11269395B2

( en )

*

2019-04-25

2022-03-08

Intel Corporation

Technologies for providing adaptive power management in an accelerator sled

US11461210B2

( en )

*

2019-06-26

2022-10-04

Kyndryl, Inc.

Real-time calculation of data center power usage effectiveness

US11340951B2

( en )

*

2019-10-23

2022-05-24

Microsoft Technology Licensing, Llc

Estimating an amount of resources that a program is expected to consume in a data center

US11875191B1

( en )

*

2020-03-02

2024-01-16

Amazon Technologies, Inc.

Energy-optimizing placement of resources in data centers

US11644882B2

( en )

*

2021-06-02

2023-05-09

Hewlett Packard Enterprise Development Lp

System and method for predicting power usage of network components

US20230017632A1

( en )

*

2021-07-12

2023-01-19

Vapor IO Inc.

Reducing the environmental impact of distributed computing

CN114970358A

( en )

*

2022-06-06

2022-08-30

深圳市证通电子股份有限公司

Data center energy efficiency optimization method and system based on reinforcement learning

Also Published As

Publication number

Publication date

US20240129380A1

( en )

2024-04-18

Similar Documents

Publication

Publication Date

Title

US20240129380A1

( en )

2024-04-18

Data center job scheduling using machine learning

US12541570B2

( en )

2026-02-03

Unsupervised domain adaptation with neural networks

US11995883B2

( en )

2024-05-28

Scene graph generation for unlabeled data

US20240161396A1

( en )

2024-05-16

Unsupervised learning of scene structure for synthetic data generation

US20210089921A1

( en )

2021-03-25

Transfer learning for neural networks

EP4329255A2

( en )

2024-02-28

Live updating of machine learning models

WO2020236596A1

( en )

2020-11-26

Motion prediction using one or more neural networks

US12387028B2

( en )

2025-08-12

Data path circuit design using reinforcement learning

US20240232039A1

( en )

2024-07-11

Application execution allocation using machine learning

US20220391667A1

( en )

2022-12-08

Processing ultrahyperbolic representations using neural networks

US12530480B2

( en )

2026-01-20

Role-based large language model to enable security and accuracy

US12125277B1

( en )

2024-10-22

Real-time persistent object tracking for intelligent video analytics systems

US20240406058A1

( en )

2024-12-05

Network fabric link maintenance systems and methods

US12254554B2

( en )

2025-03-18

Offloading shader program compilation

US12217326B2

( en )

2025-02-04

Identifying application buffers for post-processing and re-use in secondary applications

US20240303988A1

( en )

2024-09-12

Dynamically composable object tracker configuration for intelligent video analytics systems

US12436897B2

( en )

2025-10-07

Cache management using eviction priority based on memory reuse

US12626157B2

( en )

2026-05-12

Identifying idle-cores in data centers using machine-learning (ML)

US20230342618A1

( en )

2023-10-26

Identifying idle processors using non-intrusive techniques

US11972281B2

( en )

2024-04-30

Just in time compilation using link time optimization

US12461087B2

( en )

2025-11-04

Fluid inspection using machine learning

US20260141056A1

( en )

2026-05-21

Audit logging across nodes in multi-node computing environments

US20250130838A1

( en )

2025-04-24

Online fault detection in safety critical systems using isolation and process migration

Legal Events

Date

Code

Title

Description

2022-09-30

FEPP

Fee payment procedure

Free format text : ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

2023-03-30

AS

Assignment

Owner name : NVIDIA CORPORATION, CALIFORNIA

Free format text : ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GANJU, SIDDHA;MENTOVICH, ELAD;BALINT, MICHAEL;AND OTHERS;REEL/FRAME:063167/0879

Effective date : 20220929

2024-04-10

STPP

Information on status: patent application and granting procedure in general

Free format text : NON FINAL ACTION MAILED

2024-07-18

STPP

Information on status: patent application and granting procedure in general

Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER

2024-09-18

STPP

Information on status: patent application and granting procedure in general

Free format text : NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS

2024-09-20

ZAAB

Notice of allowance mailed

Free format text : ORIGINAL CODE: MN/=.

2024-12-19

STPP

Information on status: patent application and granting procedure in general

Free format text : PUBLICATIONS -- ISSUE FEE PAYMENT VERIFIED

2025-01-01

STCF

Information on status: patent grant

Free format text : PATENTED CASE

Related documents

• A Scalable Digital Twin Framework for Energy Optimization in Data Centers#168279
• On the understandability of machine learning practices in deep learning and reinforcement learning based systems.#489494
• Artificial Intelligence for Energy Optimization in Data Centers#926607
• Resource Allocation Optimizing Resource Allocation in Data Centers and Networks using AI to Efficiently Distribute Bandwidth and Computing Power#107770
• KLineage: Recovering the Missing When of Kernel Optimization by Deoptimizing Experts#1141552