Intent-driven power management
Various systems and methods for implementing intent-driven power management are described herein. A system includes: a power monitoring unit to collect real-time telemetry of a processor on a compute node; and a power level controller to: receive a power intent for execution of an application on the compute node; configure a power level of the processor of the compute node based on the power intent, the processor to execute the application; set an initial execution priority of the application on the compute node based on the power intent; and modify the initial execution priority based on the power intent and the real-time telemetry of the compute node.
1 . A system for implementing intent-driven power management, comprising:
a power monitoring unit to collect real-time telemetry of a processor on a compute node; and
a power level controller to:
receive, from an external orchestrator system, a power intent for execution of an application on the compute node, wherein the power intent is described as a declarative statement included in a service level agreement (SLA);
configure a power level of the processor of the compute node based on the power intent, the processor to execute the application;
set an initial execution priority of the application on the compute node based on the power intent;
modify the initial execution priority based on the power intent and the real-time telemetry of the compute node;
receive processor utilization metrics from the compute node;
receive processor power metrics from the compute node;
receive instructions per cycle metrics from the compute node;
receive a delta for a service level agreement for each node that the application is executing on, the delta for the service level agreement produced based on a comparison between an end-to-end service level agreement telemetry and an intent-based service level agreement;
receive a delta for power consumption for each node that the application is executing on, the delta for power consumption produced based on a comparison between the power intent and processor power metrics; and
configure the power level of the processor based on the delta for a service level agreement and the delta for power consumption, wherein to configure the power level of the processor, the power level controller is configured to implement a trained power model that determines an amount of corrective movement to align the end-to-end service level agreement telemetry with the intent-based service level agreement.
2 . The system of claim 1 , wherein the power level includes a power limit 1 (PL1) setting.
3 . The system of claim 1 , wherein the power level includes a power limit 2 (PL2) setting.
4 . The system of claim 1 , wherein the initial execution priority includes a preset priority.
5 . The system of claim 1 , wherein the real-time telemetry of the processor includes per core frequency, per core power state residency, uncore frequency, or overall power consumption.
6 . The system of claim 1 , wherein the power level controller is to:
configure the power level of the processor based on the processor utilization metrics, the processor power metrics, and the instructions per cycle metrics.
7 . The system of claim 6 , wherein the processor utilization metrics include an exponential weighted moving average of processor utilization.
8 . The system of claim 6 , wherein the processor power metrics include an exponential weighted moving average of processor power consumption.
9 . The system of claim 6 , wherein the instructions per cycle metrics include an exponential weighted moving average of instructions per cycle the application executes on the processor.
10 . A method for implementing intent-driven power management, comprising:
collecting real-time telemetry of a processor on a compute node;
receiving, from an external orchestrator system, a power intent for execution of an application on the compute node, wherein the power intent is described as a declarative statement included in a service level agreement (SLA);
configuring a power level of the processor of the compute node based on the power intent, the processor to execute the application;
setting an initial execution priority of the application on the compute node based on the power intent;
modifying the initial execution priority based on the power intent and the real-time telemetry of the compute node;
receiving processor utilization metrics from the compute node;
receiving processor power metrics from the compute node;
receiving instructions per cycle metrics from the compute node;
receiving a delta for a service level agreement for each node that the application is executing on, the delta for the service level agreement produced based on a comparison between an end-to-end service level agreement telemetry and an intent-based service level agreement;
receiving a delta for power consumption for each node that the application is executing on, the delta for power consumption produced based on a comparison between the power intent and processor power metrics; and
configuring the power level of the processor based on the delta for a service level agreement and the delta for power consumption, by implementing a trained power model that determines an amount of corrective movement to align the end-to-end service level agreement telemetry with the intent-based service level agreement.
11 . The method of claim 10 , including:
configuring the power level of the processor based on the processor utilization metrics, the processor power metrics, and the instructions per cycle metrics.
12 . The method of claim 11 , wherein the processor utilization metrics include an exponential weighted moving average of processor utilization.
13 . At least one non-transitory machine-readable medium including instructions for implementing intent-driven power management, which when executed by a machine, cause the machine to perform operations comprising:
collecting real-time telemetry of a processor on a compute node;
receiving, from an external orchestrator system, a power intent for execution of an application on the compute node, wherein the power intent is described as a declarative statement included in a service level agreement (SLA);
configuring a power level of the processor of the compute node based on the power intent, the processor to execute the application;
setting an initial execution priority of the application on the compute node based on the power intent;
modifying the initial execution priority based on the power intent and the real-time telemetry of the compute node;
receiving processor utilization metrics from the compute node;
receiving processor power metrics from the compute node;
receiving instructions per cycle metrics from the compute node;
receiving a delta for a service level agreement for each node that the application is executing on, the delta for the service level agreement produced based on a comparison between an end-to-end service level agreement telemetry and an intent-based service level agreement;
receiving a delta for power consumption for each node that the application is executing on, the delta for power consumption produced based on a comparison between the power intent and processor power metrics; and
configuring the power level of the processor based on the delta for a service level agreement and the delta for power consumption, wherein to configure the power level of the processor, by implementing a trained power model that determines an amount of corrective movement to align the end-to-end service level agreement telemetry with the intent-based service level agreement.
14 . The non-transitory machine-readable medium of claim 13 , wherein the power level includes a power limit 1 (PL1) setting.
15 . The non-transitory machine-readable medium of claim 13 , wherein the power level includes a power limit 1 (PL2) setting.
16 . The non-transitory machine-readable medium of claim 13 , wherein the initial execution priority includes a preset priority.
17 . The non-transitory machine-readable medium of claim 13 , wherein the real-time telemetry of the processor includes per core frequency, per core power state residency, uncore frequency, or overall power consumption.
18 . The non-transitory machine-readable medium of claim 13 , including:
configuring the power level of the processor based on the processor utilization metrics, the processor power metrics, and the instructions per cycle metrics.
19 . The non-transitory machine-readable medium of claim 18 , wherein the processor utilization metrics include an exponential weighted moving average of processor utilization.
20 . The non-transitory machine-readable medium of claim 18 , wherein the processor power metrics include an exponential weighted moving average of processor power consumption.
21 . The non-transitory machine-readable medium of claim 18 , wherein the instructions per cycle metrics include an exponential weighted moving average of instructions per cycle the application executes on the processor.