Abstract: In some examples, a controller dynamically configures a property associated with monitoring performed by an agent. The controller stores, in a repository, metadata relating to the agent. The controller receives, from the agent, first sensor data that excludes the metadata, and uses indexing information in the first sensor data to retrieve the metadata from the repository.

Abstract: A technique includes an agent executing on a plurality of nodes while a job is being concurrently executed by the plurality of nodes. The plurality of nodes is power-capped by an existing node power consumption budget. The technique includes managing power consumption of the plurality of nodes. The managing includes the agent determining a performance footprint that is associated with execution of the job; and the managing includes the agent determining a second node power consumption budget based on the performance footprint. The second node power consumption budget is different than the existing node power consumption budget. The managing includes the agent providing a power consumption request to a global power dispatcher to set a new node power consumption budget for the plurality of nodes.

Abstract: A method for optimizing operations of high-performance computing (HPC) systems includes collecting data associated with a plurality of workload performance profiling counters associated with a workload during runtime of the workload in an HPC system. Based on the collected data, the method includes using a machine-learning technique to classify the workload by determining a workload-specific fingerprint for the workload. The method includes identifying an optimization metric to optimize during running of the workload in the HPC system. The method includes determining an optimal setting for a plurality of tunable hardware execution parameters as measured against the optimization metric by varying at least a portion of the plurality of tunable hardware execution parameters.

Abstract: Systems and methods described herein make previously stranded power capacity (power that is provisioned for a data center according to a computing system's nameplate power consumption but is currently not useable) available to the data center. Systems described herein generate empirical power profiles that specify expected upper bounds for the power consumption levels that applications trigger. Using the upper bounds for application power-consumption levels, a computing system described herein can reliably release part of its provisioned nameplate power for other systems or data center consumers, reducing the amount of stranded power in a data center. The method described herein avoids performance penalties for most jobs by using sensor measurements made at a rapid rate explained herein to ensure that a system power cap based on running application's measured peak power consumption is reliable with reference to the power capacitance inherent in the computing system.

Abstract: One embodiment provides a system and method for predicting network power usage associated with workloads. During operation, the system configures a simulator to simulate operations of a plurality of network components, which comprises embedding one or more event counters in each simulated network component. A respective event counter is configured to count a number of network-power-related events. The system collects, based on values of the event counters, network-power-related performance data associated with one or more sample workloads applied to the simulator; and trains a machine-learning model with the collected network-power-related performance data and characteristics of the sample workloads as training data 1, thereby facilitating prediction of network-power-related performance associated with a to-be-evaluated workload.

Abstract: Systems and methods are provided for improving statistical and machine learning drift detection models that monitor computing health of a data center environment. For example, the system can receive streams of sensor data from a plurality of sensors in a data center; clean the streams of sensor data; generate, using a machine learning (ML) model, an anomaly score and a dynamic threshold value based on the cleaned streams of sensor data; determine, using the ML model and based on the anomaly score and the dynamic threshold value, a correctness indicator for a first sensor in the plurality of sensors; and using the correctness indicator, correct the first sensor.

Abstract: One embodiment provides a system and method for predicting network power usage associated with workloads. During operation, the system configures a simulator to simulate operations of a plurality of network components, which comprises embedding one or more event counters in each simulated network component. A respective event counter is configured to count a number of network-power-related events. The system collects, based on values of the event counters, network-power-related performance data associated with one or more sample workloads applied to the simulator; and trains a machine-learning model with the collected network-power-related performance data and characteristics of the sample workloads as training data 1, thereby facilitating prediction of network-power-related performance associated with a to-be-evaluated workload.

Abstract: In some examples, a controller dynamically configures a property associated with monitoring performed by an agent. The controller stores, in a repository, metadata relating to the agent. The controller receives, from the agent, first sensor data that excludes the metadata, and uses indexing information in the first sensor data to retrieve the metadata from the repository.

Abstract: A technique includes an agent executing on a plurality of nodes while a job is being concurrently executed by the plurality of nodes. The plurality of nodes is power-capped by an existing node power consumption budget. The technique includes managing power consumption of the plurality of nodes. The managing includes the agent determining a performance footprint that is associated with execution of the job; and the managing includes the agent determining a second node power consumption budget based on the performance footprint. The second node power consumption budget is different than the existing node power consumption budget. The managing includes the agent providing a power consumption request to a global power dispatcher to set a new node power consumption budget for the plurality of nodes.

Abstract: Systems and methods described herein make previously stranded power capacity (power that is provisioned for a data center according to a computing system's nameplate power consumption but is currently not useable) available to the data center. Systems described herein generate empirical power profiles that specify expected upper bounds for the power consumption levels that applications trigger. Using the upper bounds for application power-consumption levels, a computing system described herein can reliably release part of its provisioned nameplate power for other systems or data center consumers, reducing the amount of stranded power in a data center. The method described herein avoids performance penalties for most jobs by using sensor measurements made at a rapid rate explained herein to ensure that a system power cap based on running application's measured peak power consumption is reliable with reference to the power capacitance inherent in the computing system.

Abstract: Systems and methods described herein make previously stranded power capacity (power that is provisioned for a data center according to a computing system's nameplate power consumption but is currently not useable) available to the data center. Systems described herein generate empirical power profiles that specify expected upper bounds for the power consumption levels that applications trigger. Using the upper bounds for application power-consumption levels, a computing system described herein can reliably release part of its provisioned nameplate power for other systems or data center consumers, reducing the amount of stranded power in a data center. The method described herein avoids performance penalties for most jobs by using sensor measurements made at a rapid rate explained herein to ensure that a system power cap based on running application's measured peak power consumption is reliable with reference to the power capacitance inherent in the computing system.

Abstract: Systems and methods described herein make previously stranded power capacity (power that is provisioned for a data center according to a computing system's nameplate power consumption but is currently not useable) available to the data center. Systems described herein generate empirical power profiles that specify expected upper bounds for the power consumption levels that applications trigger. Using the upper bounds for application power-consumption levels, a computing system described herein can reliably release part of its provisioned nameplate power for other systems or data center consumers, reducing the amount of stranded power in a data center. The method described herein avoids performance penalties for most jobs by using sensor measurements made at a rapid rate explained herein to ensure that a system power cap based on running application's measured peak power consumption is reliable with reference to the power capacitance inherent in the computing system.