Breaking Barriers: From Black Box to Intent-Driven, User-Friendly

Power Management

Thijs Metsch

1 a

and Adrian Hoban

2 b

Intel Labs, Intel Corporation, Santa Clara, CA, U.S.A.

Network and Edge Group, Intel Corporation, Santa Clara, CA, U.S.A.

{thijs.metsch, adrian.hoban}@intel.com

Keywords:

Edge Computing, Cloud Computing, Power Management, Energy Efﬁciency, Intent-Driven-Systems.

Abstract:

Power management in cloud and edge computing platforms is challenging due to the need for domain-speciﬁc

knowledge to conﬁgure optimal settings. Additionally, the interfaces between application owners and resource

providers often lack user-friendliness, leaving efﬁciency potentials unrealized. This abstraction also hinders

the adoption of efﬁcient power management practices, as users often deploy applications without optimization

considerations. Efﬁcient energy management works best when user intentions are clearly speciﬁed. Without

this clarity, applications are treated as black boxes, complicating the process of setting appropriate throttling

limits. This paper presents an application intent-driven orchestration model that simpliﬁes power management

by allowing users to specify their objectives. Based on these intentions, we have extended Kubernetes to

autonomously conﬁgure system settings and activate power management features, enhancing ease of use.

Our model demonstrates the potential to reduce power consumption in a server ﬂeet within a range of ≈ 5-

55% for a sample AI application. When applied broadly, the research offers promising potential to address

both economic and environmental challenges. By adopting this model, applications can be more efﬁciently

orchestrated, utilizing advanced resource management techniques to mitigate the power usage surge that is in

part driven by applications such as AI and ML.

1 INTRODUCTION

Managing power efﬁciently in cloud and edge com-

puting is increasingly challenging due to the growing

power demands of applications such as AI and ML

based applications. As the need for environmentally

friendly deployments grows, understanding the roles

and motivations of different users/stakeholders – ap-

plication owners and resource providers – becomes

crucial. Current power management features often re-

quire domain-speciﬁc knowledge and platform-level

conﬁgurations, which are either not easily accessible

or deliberately restricted (e.g. for security reasons) to

application owners. This results in sub-optimal de-

ployments and inefﬁciencies in platform management

and leave untapped savings for the resource providers.

Overall, this hinders the adoption of the green edge-

cloud continuum.

Moreover, hardware features often automatically

boost to higher performance modes when detecting

https://orcid.org/0000-0003-3495-3646

https://orcid.org/0009-0001-4970-889X

activity, leading to increased power draw. Without ad-

equate context, these boosts can be inefﬁcient and un-

necessary. By providing context through application

intents, power management can become more precise

and efﬁcient, aligning performance boosts with ac-

tual needs. Additionally, system-wide power conﬁg-

urations generally result in higher power draw com-

pared to ﬁne-tuned, per-application conﬁgurations.

Resource providers should only ﬂexibly perform per-

application tuning, and space and time shift them if

they understand the application intents. This aligns

with the concept of ”tell me what you want, not how to

do it”, demonstrating how resource providers beneﬁt

more from understanding user intentions rather than

following a strict set of instructions.

This paper uses an Intent-Driven Orchestration

(IDO) model as a novel approach to address these

challenges. By allowing application owners to spec-

ify their intents through a set of objectives instead

of low-level resource requirements, intent-driven sys-

tems bridge the gap between user-friendly interfaces

and complex power management conﬁgurations. The

obtained experimental data indicates an enhanced ef-

Metsch, T. and Hoban, A.

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management.

DOI: 10.5220/0013150000003953

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2025), pages 24-35

ISBN: 978-989-758-751-1; ISSN: 2184-4968

ﬁciency in the computing continuum.

Intent-driven systems ensure that applications are

not treated as black boxes but are managed based on

intentions, enabling more effective power optimiza-

tion. Key beneﬁts include:

• Semantic Application Portability - Declaring

intents with targeted application objectives remain

consistent across platforms, unlike declaring spe-

ciﬁc resource requests, which can result in signif-

icantly different performance and power charac-

teristics depending on the platform context.

• Invariance - Intents allow resource providers to

dynamically modify system resource allocations

in response to changing contexts, such as avail-

able energy, optimizing overall system setup over

time as long as the intents can be met.

• Context - Intents provide context for resource

providers to maximize performance and energy

efﬁciency by understanding the application own-

ers’ objectives.

Throughout this paper, the implementation and

beneﬁts of IDO models for various applications are

demonstrated. By leveraging Intel

’s IDO extension

for Kubernetes

, we enable autonomous conﬁguration

of systems based on user-speciﬁed intents, simplify-

ing power management and enhancing ease of use.

The rest of the paper is structured as follows:

Section 2 introduces efﬁcient power management for

feature-aware advanced users. Section 3 showcases

the beneﬁts of an IDO model. Section 4 discusses the

results of providing intents to a orchestration stack,

followed by related work in Section 5 and conclusions

in Section 6.

2 PLATFORM LEVEL POWER

MANAGEMENT

Intel

Xeon

processors have an extensive set of

power management capabilities that were created to

automatically, or under user and systems administra-

tors preference, adjust the compute performance to

power consumption ratio dynamically in response to

the needs of the applications or the resource provider.

2.1 Core Sleep/Idle States (c-State)

The Advanced Conﬁguration and Power Interface

(ACPI) speciﬁcation

deﬁnes the processor power

https://github.com/intel/intent-driven-orchestration/

https://ueﬁ.org/htmlspecs/ACPI Spec 6 4 html/

state, known as its C-state. The C-state is sometimes

colloquially known as the processor “idle” state on a

per-core basis or on a CPU package basis. Core and

package C-states coordination is managed by the CPU

Power Control Unit. C-state values range from C0 to

Cn, where n is dependent on the speciﬁc processor.

When the core is active and executing instructions, it

is in the C0 state. Higher C-states indicate how deep

the CPU idle state is.

Higher C-states consume less energy when resi-

dent in that state but require longer latency times to

transition into the active C0 state. The BIOS/UEFI

conﬁguration can be conﬁgured to restrict how deeply

the cores and package can idle, e.g., it is possible to

restrict access to the deepest C-state.

In Linux, C-state management is implemented on

modern Intel

Xeon

processors with the intel idle

driver that is part of the CPU idle time management

subsystem. Linux categorizes a CPU core as being

idle if there are no tasks to run on it except for the

“idle” task. Note that C-state information is intro-

duced here for completeness and to draw a distinction

with P-State, however the experiments outlined in this

paper did not leverage C-state capabilities.

2.2 Core Frequency/Voltage State

(P-State)

Intel

Xeon

processors include the ability to alter

the processor operating frequency and voltage be-

tween high and low levels. The frequency and voltage

pairings are deﬁned in the ACPI speciﬁcations as the

processor Performance State (P-state). P-states are

SKU-speciﬁc settings ranging from the low-end of the

operating frequency and voltage pairings with mini-

mums deﬁned by Pn to the nominal operating con-

dition Base Frequency (P1), to the maximum single

core turbo (P01).

Intel

Xeon

processors with HW P-state Man-

agement (HWP) can manage P-state transitions auto-

matically with some tuning conﬁguration provided to

it by the operating system. With the Linux operat-

ing system the CPU Performance Scaling Subsystem

(CPUFreq)

is responsible for providing the OS-level

inputs to the processor P-State management capabil-

ity. The CPUFreq layer is composed of core foun-

dational capabilities, and scaling governors that con-

tain the algorithms for selecting P-States and scal-

ing drivers for communicating the desired P-States to

the processor. Users/administrators can conﬁgure the

CPU scaling with the sysfs ﬁlesystem and using utili-

https://www.kernel.org/doc/html/v6.7/admin-guide/

pm/cpufreq.html

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management

ties such as the CPUFreq Governor

. Note, the Intel

Xeon

processors’ internal logic will be the ﬁnal ad-

judicator of what P-State the processor should be in.

The intel pstate driver

is the CPU scaling driver

part of the CPUFreq subsystem for Intel

Xeon

pro-

cessors. When conﬁgured in the default active mode

with HWP enabled there are effectively two governor

modes exposed:

• A performance governor mode that promotes

maximum performance.

• A powersave governor mode that promotes maxi-

mum power savings with a reduction in peak per-

formance.

Intel

Xeon

processors can take as an input

to their P-state management an Energy Performance

Preference (EPP) setting. When in performance

mode, EPP is set to its maximum performance set-

ting (0). In powersave model, the EPP setting deter-

mines how aggressively the CPU pursues power sav-

ing conﬁgurations. Furthermore, uncore frequency

scaling is a power management mechanism in mod-

ern CPUs that adjusts the frequency of the uncore do-

main, which includes components like the memory

controller, last-level cache, and interconnects. Unlike

P-states and C-states, which operate on a per-core ba-

sis, uncore frequency scaling independently regulates

the performance of shared subsystems.

2.3 Power Manager for Kubernetes

The Intel

Power Manager for Kubernetes software

is a Kubernetes Operator that has been developed to

provide Kubernetes cluster users with a mechanism

to dynamically request adjustment of worker node

power management settings applied to cores allocated

to the Pods running the applications. The power

management-related settings can be applied to indi-

vidual cores or to groups of cores, and each may have

different policies applied. It is not required that ev-

ery core in the system be explicitly managed by this

Kubernetes power manager. When the Power Man-

ager is used to specify core power related policies, it

overrides the default settings. The container deploy-

ment model in scope is for containers running on bare

metal (i.e., host OS) environments.

The Power Manager for Kubernetes software has

two main components, the Conﬁguration Controller

https://www.kernel.org/doc/Documentation/cpu-freq/

governors.txt

https://www.kernel.org/doc/html/v5.15/admin-guide/

pm/intel pstate.html

https://github.com/intel/kubernetes-power-manager

and the Power Node Agent, which in turn has a de-

pendency on the Intel

Power Optimization Library

which in turn conﬁgures the intel pstate driver.

The Conﬁguration Controller deploys, sets, and

maintains the conﬁguration of the Power Node Agent.

By default, it applies four cluster-administrator or

user-modiﬁable PowerProﬁle settings to the Power

Node Agent. This enables Pods to select between

a performance, balance-performance, balance-power

proﬁle, and a default proﬁle, to be conﬁgured on cores

allocated to the Pod. Note, in the experiments in this

paper, Proﬁles A-D were deﬁned as per 1.

The Power Manager for Kubernetes organizes

CPU cores into logical ”pools”, including a shared

pool, which represents the Kubernetes shared CPU

pool, and a default pool, a subset of system-reserved

cores within the shared pool. Cores in the default pool

are excluded from power setting conﬁgurations, while

those in the shared pool are assigned a power-saving

proﬁle.

The Power Manager for Kubernetes also supports

the notion of an ”exclusive” pool. This pool is used

to group cores allocated to Guaranteed QoS class of

Kubernetes Pods. When cores are allocated to Guar-

anteed Pods, they are moved from the Power Manager

Shared Pool into the Power Manager Exclusive Pool.

This mechanism supports the model where cores that

are not expected to be pinned to applications can be

conﬁgured to run in a low power model.

Note, per-core power management works in as-

sociation with CPU pinning deployment semantics.

The Kubernetes Resource Orchestration for 4th Gen

Intel

Xeon

Scalable Processors

technology guide

provides a more detailed discussion on this and other

resource allocation considerations.

2.4 Observability

One of the fundamental premises of improving on a

current state is the ability to measure it. This is partic-

ularly true for activities aiming to support more sus-

tainable computing at the edge & cloud.

The observability in focus in the context of this

paper is on telemetry aspects related to power con-

sumption at the full system (node) level as well as

at the application level. There are several tools and

techniques to measure power at both of these levels.

This paper does not aim to identify what might be the

best approach, merely use some of the tools and ap-

proaches consistently throughout the experiments and

https://github.com/intel/power-optimization-library

https://networkbuilders.intel.com/solutionslibrary/

kubernetes-resource-orchestration-for-4th-gen-intel-xeon-

scalable-processors-technology-guide

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

focus on the differential impact that the proposed ap-

proach has on power consumed in the environment.

For infrastructure, processor-speciﬁc and in-depth

data collection, tools such as telegraf

are available.

Telegraf has a plugin model that supports data col-

lection from many sources, including from Power-

stat

. Telegraf can subsequently be used to distribute

telemetry data from the node to data sinks.

Prometheus

is a Time Series Database with

systems monitoring and alerting capabilities.

Prometheus can be conﬁgured to pull data from

telemetry collection tools such as Telegraf via a

Telegraf Prometheus exporter plugin.

Grafana

is a monitoring and data visualiza-

tion tool. When the above tools are combined with

Grafana, a pipeline can be established to assess and

visualize the node-level power consumption.

In addition to node-level power consumption, un-

derstanding the per-application power consumption

can add to the understanding of how effective power

efﬁcient strategies are being from the application per-

spective. Scaphandre

is one tool that is focused on

per-app energy consumption metrics.

3 INTENT-DRIVEN

ORCHESTRATION

To switch to intent-driven systems, Intel

’s IDO ex-

tension for the Kubernetes control plane is key, en-

abling it to understand and process application-level

intents through Custom Resource Deﬁnitions (CRDs)

as discussed in (Metsch et al., 2023). By enabling

Kubernetes CRDs, the intents can be deﬁned as ob-

jects through the Kubernetes API. This allows appli-

cation owners to specify their intents through a set

of objectives rather than low-level resource require-

ments, simplifying the interface and enhancing user-

friendliness.

By allowing users to express their Intent with a set

of objectives and associated Key Performance Indica-

tors (KPIs), the IDO extensions automate the trans-

lation of these intents into actionable conﬁgurations,

facilitating more efﬁcient power management. The

KPIProﬁle enable users to deﬁne how the orchestra-

tion stack can observe their application’s behavior.

Fig. 1 shows the object model that details how in-

tents, objectives and KPIs relate. Objectives in the

https://github.com/inﬂuxdata/telegraf

https://manpages.org/powerstat/8

https://prometheus.io/

https://grafana.com/

https://github.com/hubblo-org/scaphandre

Figure 1: Object model detailing relationship between in-

tents, objectives and KPIs.

IDO model are unit-less objects categorized by a type

such as latency, throughput, availability, and power,

enabling their speciﬁcation across a wide range of ap-

plications, such as - but not limited to - AI applica-

tions, video processing, micro-services, web services,

and High-Throughput Computing (HTC).

The Kubernetes extension integrates a planning

component within Kubernetes, this planner is respon-

sible for continuously determining the necessary con-

ﬁgurations and setup needed to meet the speciﬁed in-

tents. It hence determine what & how to optimally

manage the overall system. It works in tandem with

the scheduler, which manages when & where to place

applications. By leveraging the insights provided by

the planner, the scheduler can make more informed

decisions, optimizing cluster level resource alloca-

tion and workload distribution. This collaboration

between the planner and scheduler ensures that the

system setup aligns with applications owners and re-

source provider’s objectives, trading of both perfor-

mance and efﬁciency.

To make informed decisions, the planner utilizes

an A* planning algorithm which leverages a set of

heuristics and utility functions that evaluate different

actions based on their potential impact on the speci-

ﬁed KPIs. These utility functions allow the planner to

incorporate the intentions of resource providers into

its decision-making. The planner can support var-

ious actions, including vertical and horizontal scal-

ing, and the conﬁguration of platform features such as

Intel

Resource Director Technology (Intel

RDT)

and Intel

Speed Select Technology (Intel

SST)

This ensures that the system can be dynamically ad-

justed to meet a set of performance and power man-

agement goals leveraging platform features under the

hood, catering for the needs of cloud-native applica-

https://www.intel.com/content/www/us/en/

architecture-and-technology/resource-director-technology.

html

https://www.intel.com/content/www/us/en/support/

articles/000095518/processors.html

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management

Figure 2: System overview of Kubernetes extensions en-

abling intent-driven orchestration.

tions. The A* planning algorithm may be replaced by

other algorithms and is a subject for further study.

A key feature of the IDO model is its pluggable ar-

chitecture, which employs plugins known as actuators

as shown in Fig. 2. Each actuator is responsible for a

speciﬁc function within the system and can use either

static lookup tables or more complex AI/ML mod-

els to determine the best orchestration actions based

on current conditions and user intents. This ﬂexi-

ble architecture allows the system to adapt to varying

workloads and objectives, continually optimizing for

performance and power consumption. The actuators’

ability to dynamically adjust conﬁgurations based on

near-real-time data ensures that the system remains

efﬁcient and responsive to changing demands. This

system operates in the orchestration time-domain, i.e.

the nature of how quickly the control loop may be

closed is a function of the orchestration software that

underpins the system. Timeliness should be consid-

ered with seconds/minutes/hours granularity.

For addressing efﬁcient power management, a

new actuator has been developed. This actuator can

adjust power settings based on power and perfor-

mance objectives, optimizing power usage while bal-

ancing it with application performance. Additionally,

it can be conﬁgured to prioritize power or carbon re-

duction intents, aligning with broader environmental

goals. Next to this power management actuator a CPU

rightsizing actuator has been used in tandem. The fol-

lowing sub-sections describe the key models used to

enable the concepts introduced in this paper.

3.1 Dynamic CPU Rightsizing

Dynamic CPU allocation rightsizing is a key in en-

ergy management, aiming to dynamically adjust CPU

resources allocated to applications. The primary goal

is to optimize energy usage while trading perfor-

mance by reducing the resources allocated at run-

time. This approach contrasts with static alloca-

tion methods, which often lead to resource over-

provisioning and consequently energy wastage. By

tailoring CPU allocation to current application needs,

dynamic rightsizing conserves energy while avoiding

resource overuse. Maximizing the utilization of avail-

able resources and ensuring that those resources asso-

ciated with an application are setup according to the

intents.

To enable the dynamic resource allocations,

a curve ﬁtting model that predicts resource

requirements-based performance metrics is used.

The equation 1 showcases the model used for

latency-related metrics. Whereby the goal is to learn

the values of the parameters p

, p

that describe

the characteristics of the application in the current

context. This model, applicable to a wide range of

applications across various domains and allows the

planner to anticipate CPU needs and adjust resources

accordingly. Our approach leverages pre-training,

on-the-ﬂy learning, and proactive adaptation to

continuously reﬁne the model. Pre-training with

historical data establishes a baseline understanding

of application behaviors, while on-the-ﬂy learning

incorporates near-real-time performance data to

ensure the model remains accurate as applications

evolve.

latency = p

∗ e

−p

∗cpu

units

+ p

(1)

Despite the gains of these techniques, the training

process is lightweight, ensuring that the energy sav-

ings from dynamic CPU rightsizing are not offset by

the costs of model maintenance. By dynamically ad-

justing CPU resources based on predictive modeling

and adaptive learning, our approach achieves energy

savings while maintaining application performance as

will be discussed in section 4.1.

3.2 Dynamic Power Management

Associating the most efﬁcient power proﬁle with

an application is crucial for effective energy man-

agement. The power proﬁle selection must con-

sider both the application’s intents and the current

system status, including the available energy sup-

ply. This approach facilitates optimal energy use, ad-

dressing both economic and environmental concerns.

For instance, during periods of high sustainably-

sourced energy availability, a performance-oriented

power proﬁle might be appropriate, whereas during

constrained sustainably-sourced energy availability, a

power-saving proﬁle can be adopted to conserve en-

ergy (Intel Network Builders, 2024).

To facilitate this dynamic power management,

a RandomForestRegressor model as described in

(Geurts et al., 2006) is used. This model is trained

to learn the relationships between selecting different

performance power proﬁles, allocated resources, as

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

well as the power consumption associated with each

proﬁle. By leveraging historical data coming from

the observability stack, the model learns how various

workload and system conditions impact power usage

and performance. Based on this model the planner

can make informed decisions about which power pro-

ﬁle to apply, optimizing both performance and power

efﬁciency. The use of a RandomForestRegressor is

particularly advantageous due to its robustness and

ability to provide insights.

The implementation currently relies on pre-

trained models. These models are initially trained in

a controlled development environment where exper-

iments with the applications are carried out to deter-

mine the effect power proﬁle selections have to gather

comprehensive training data. Once the models have

achieved satisfactory accuracy and reliability, they are

deployed to the production environment. This trans-

fer ensures that the models can operate in various sce-

narios with minimal adjustments, providing accurate

and efﬁcient power management. By selecting power

proﬁles on a per-application basis rather than system-

wide, the power management strategy can be tailored

to the speciﬁc needs of each application. This granu-

larity allows for further energy savings and efﬁciency

gains, as each application can operate under the most

suitable power conditions for its intents as will be dis-

cussed in section 4.2.

4 EXPERIMENTAL RESULTS

The experiments utilized a system featuring an Intel

Xeon

Processor. These processors are designed for

cloud & edge deployments, although the methodolo-

gies outlined in this paper are also applicable to other

deployment scenarios. The system ran on Ubuntu

22.04 LTS, with Kubernetes 1.27 deployed to man-

age the extensions described in Section 2. The con-

ﬁgured power proﬁles are detailed in Table 1, cho-

sen based on the CPU’s characteristics and power

curve behavior. Proﬁles A and B were selected for

their power-saving capabilities, while proﬁles C and

D were chosen for their performance attributes. To

enable a power saving mentality all cores are put in

Proﬁle A by default. Notably the proﬁles were se-

lected and conﬁgured with the CPUs characteristics

in mind by the resources owner, relieving the appli-

cation owners of the need to have (domain) speciﬁc

knowledge about their setup & power ratings.

To demonstrate the methodologies, various appli-

cations were employed to validate the effectiveness of

the features. The following applications were used:

Table 1: Power proﬁles conﬁgurations.

Name Min Max EPP

proﬁle A 800Mhz 1600Mhz power

proﬁle B 800Mhz 1800Mhz balance power

proﬁle C 800Mhz 2400Mhz

balance

performance

proﬁle D 800Mhz 3500Mhz performance

• FaaS (Function-as-a-Service) This application

performs mathematical calculations triggered by

HTTP events. Performance-related KPIs were in-

strumented using the Linkerd

service mesh, pro-

viding P99, P95, and P50 latencies. The key ob-

jective used for these experiments was the P99 la-

tency.

• OVMS (OpenVINO

Model Server)

This

application performs object detection on video

frames, representing a typical edge use case aimed

at conserving network bandwidth by processing

data locally. The application was modiﬁed to ex-

pose the processing time per video frame as a his-

togram via a Prometheus client

. The key objec-

tive for these experiments was the P99 latency.

• AI Mistral LLM This application functions as an

AI chatbot responding to incoming requests using

a Mistral 7B LLM

. It uses a Prometheus client

to expose the time required to generate a token.

The key objective for these experiments was the

average token creation time, as the ﬁrst response

token typically takes longer to generate, affecting

higher percentile latencies.

The aim of the selected application was to mimic

a scenario in which multiple applications from poten-

tially different tenants are run on the node. In this

environment it is the goal to optimize the power us-

age of each node, with the overarching goal to achieve

cluster level power optimization. All values were nor-

malized for comparison purposes, ranging from 0 to

the maximum observed value of a data series. This

normalization facilitates easier data interpretation and

demonstrates that the methodology can be applied

across a diverse set of use cases and scenarios.

4.1 Dynamic CPU Rightsizing

Fig. 3 illustrates the resulting models for the three ap-

plications when applying the methodology described

in Section 3.1. While the FaaS application can oper-

https://linkerd.io/

https://docs.openvino.ai/2023.3/ovms what is

openvino model server.html

https://prometheus.io/docs/instrumenting/clientlibs/

https://mistral.ai/news/announcing-mistral-7b/

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management

ate with fewer CPU cores, the AI Mistral LLM and

OVMS applications require a minimum number of

cores to function.

The resource allocation curves for each applica-

tion vary due to their unique implementations and

characteristics. However, the planner can utilize these

models to efﬁciently allocate the appropriate number

of resources for each application in a speciﬁc context.

The common characteristic from the charts in Fig.

3 is that from the perspective of the KPI being as-

sessed, there is a point beyond which adding ad-

ditional CPU cores provides limited improvements

in the KPI, i.e. the curve levels off. This point

varies by application but is dynamically determined

by the learning algorithms in the IDO implementa-

tion. While the underlying curve-ﬁtting model is rel-

atively simple, it accommodates a wide range of ap-

plications. Notably, although these curves correspond

to a speciﬁc power proﬁle, their shapes remain largely

consistent across proﬁles, differing primarily in am-

plitude.

4.2 Dynamic Power Management

Fig. 4 presents the resulting models after applying the

methodologies described in Section 3.2. It highlights

the impact of power proﬁle selection on the perfor-

mance of each application. Where 1.0 is shown in

these graphs represents the worst performing latency.

The best performing latency is closest to zero. The

Function and OVMS applications require a certain

amount of CPU resource to be allocated to achieve

lower latency objectives. This is also true for the AI

Mistral LLM application, however here the selection

of the power proﬁle plays a bigger role. To achieve

lower latencies at least the proﬁles C or D need to

be selected. This variation is due to the speciﬁc im-

plementation and hardware utilization of each appli-

cation, demonstrating that different applications have

unique characteristics that can be efﬁciently learned.

Notably, we found that AI applications are highly

ﬂexible and can be efﬁciently managed within the

context of power and energy constraints.

4.3 Enhancing Usability

The following paragraphs describe how the models

shown in section 4.1 and 4.2 can be used by the

IDO planner as described in section 3. Overall, this

demonstrates the ease of use of platform features for

the application owners – as they only work with their

intents and associated targeted objectives – while pro-

viding the resource providers with additional context

to efﬁciently manage their compute platform.

Making Decision Based on Performance Related

Objectives. Fig. 7 shows a screenshot of the dash-

board for the IDO planner. The table shows the event

timestamps, the current state of the application (de-

ﬁned by its objectives), the desired state, and the re-

sulting plan. In this example, the FaaS application is

under control of the planner, targeting a P99 objec-

tive. When the target objective shifts from 100ms to

200ms, the system efﬁciently frees up resources and

selects a more efﬁcient power proﬁle.

Making Decision Based on Power Related Objec-

tives. Fig. 8 shows a screenshot of the dashboard

for the IDO planner when applying power objectives.

In this example, the system is purely driven by power

objectives. When the target objective shifts from an

average of 10W to 5W while running the application,

the system selects a more power-efﬁcient proﬁle. This

capability is a crucial step toward enabling carbon-

based objectives in the future.

4.4 Experimental Gains

Power proﬁle selection impacts not only the perfor-

mance of individual applications but also the power

consumption estimations as shown in the previous

section 4.3. Fig. 5 demonstrate how the model de-

scribed in section 3.2 can also be used predict the

power draw for a given application based on its power

proﬁle and CPU resource allocation.

For applications that fully utilize compute re-

sources, high power draw is observed when more

cores are allocated, which are allowed to operate at

higher frequencies. Next to the core frequencies that

can be adjusted using the P-states and C-states, the

uncore components (e.g., the memory controller, in-

terconnects, etc.) can also be managed independently.

Since uncore components are shared by all applica-

tions, their activity signiﬁcantly impacts each appli-

cation’s overall power draw estimation.

Current power estimation models attribute power

usage based primarily on utilization of the cores asso-

ciated by the processes. However, the power draw of

shared components can result in elevated power draw

irrespective of which application is responsible, a fac-

tor not adequately captured in most power estimation

tools. Consequently, further research is required to

reﬁne power estimation for improved accuracy. Also,

more ﬁne-grained per sub-component frequency scal-

ing methodologies will enable further gains. For ap-

plications that fully utilize compute resources and are

memory-bound—such as the AI Mistral LLM evalu-

ated here (as can be conﬁrmed by analyzing the In-

structions per Cycle (IPC) and memory bandwidth us-

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

(a) Function as a Service App (b) OVMS App (c) AI Mistral LLM App

Figure 3: The CPU rightsizing models illustrate the impact of CPU resource allocations on latency-related objectives for

various applications, including: a) P99 tail latency for a FaaS-style deployment, b) P99 compute latency for processing video

frames, and c) average token creation time for an AI LLM model. The forecast based on the curve-ﬁtting are shown in blue

for the performance proﬁle D.

(a) Function as a Service App (b) OVMS App (c) AI Mistral LLM App

Figure 4: The selection of the right performance proﬁle has an effect on the performance of the application.: a) P99 tail latency

for a FaaS-style deployment, b) P99 compute latency for processing video frames, and c) average token creation time for an

AI LLM model. Note, a value of 1.0 is the normalized worst (highest) latency reading measured.

age nearing the CPUs’ theoretical maximums) – the

current power model is effective and allows for intent-

driven orchestration using power objectives.

These ﬁndings align with those presented in

(Pereira et al., 2017) and in general call for more

efﬁcient software. While the results are speciﬁc to

the platform and applications used, an intent-driven

model enhances semantic application portability and

can learn and adapt to different contexts.

Finally, Fig. 6 shows the experimental gains for

the AI Mistral LLM application in more detail. In this

set of charts, the effects that two of the system-wide

CPU frequency scaling governors have when applied

to the processor for this application is contrasted to an

per-application intent-driven model. The normaliza-

tion point was chosen as that produced by the default

system-wide lower power consuming conﬁguration in

the powersave governor. For these experiments, the

node level utilization was observed to be ≈ 35% as

multiple applications were deployed.

The effects of the two system-wide conﬁgurations

are broadly aligned with their monikers. I.e., the per-

formance CPU frequency governor delivers the best

application KPI which is the lowest average token

creation time while consuming the most node-level

and application-level power. The powersave CPU fre-

quency governor offered a considerably lower power

draw by the node and the application, but with a worse

(i.e. higher) average token creation time.

The system-wide conﬁguration choice implied by

selecting between the two native CPU frequency

governors has a signiﬁcant implication for the re-

source provider and the application owner. Purchas-

ing a high-performance processor and conﬁguring it

to run with the powersave CPU frequency governor is

shown in the experiment with the AI LLM application

to considerably limit the KPI potential. Such a conﬁg-

uration could result in the application owner to require

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management

Figure 5: The selection of the right performance proﬁle has

an effect on the power estimation per-application - e.g. for

an AI LLM model. Note, a value of 1.0 represents the nor-

malized highest power consumption measured.

actions such as invoking earlier scale-out strategies to

make up the performance short fall thereby undermin-

ing the intention of restricting the power draw of the

node. Nonetheless it is a simple, effective mechanism

to limit power consumption.

The behavior of the power proﬁles when applied

to the application is broadly aligned with expecta-

tions. Proﬁle A offered the worst KPI performance

(highest latency) and best power consumption (lowest

power draw), with Proﬁle D the opposite. The follow-

ing points are particularly interesting:

• At one end of the scale a considerably better (≈

55%) overall power consumption (lowest power

draw) could be achieved with Proﬁle A when

compared to the system wide powersave CPU fre-

quency governor.

• At the other end of the scale an equivalent appli-

cation KPI, the best average token creation time

(lowest time) was achieved with Proﬁle D as with

using the performance governor while also having

a slightly lower (≈ 5%) power draw.

• Proﬁle C offers a ≈ 30% performance gain and

Proﬁle D a gain of ≈ 40% compared to a system-

wide powersave, while both outperforming the

latter. Both proﬁles offer a better power efﬁciency,

with ≈ 5% and ≈ 15% power savings respec-

tively. Notably the difference between their per-

formance characteristic is just ≈ 5% while the

power savings are within the range of ≈ 10%, of-

fering options for trading-off efﬁciency gains.

• Through an IDO model – leveraging an per-

application power management capability – simi-

lar performance can be achieved, while this model

enable showcase power efﬁciency gains contrast-

ing it to system-wide conﬁguration options across

the board.

Similar to the results observed with the AI Mistral

LLM application, both the Function and OVMS appli-

cations demonstrate potential power savings of up to

≈ 50%. This level of efﬁciency is largely attributed

to their smaller CPU footprint, which allows most of

the system to remain in a power-saving mode.

This observation underscores the importance of

consolidating applications with compatible power

proﬁles (based on their intents) on servers to optimize

power efﬁciency for resource providers. It is impor-

tant to note, however, that these results may vary de-

pending on the speciﬁc system conﬁguration and ap-

plications utilized.

The experiment shows that IDO offers an alter-

native approach that absolves a system administrator

from having to choose between two system-wide set-

tings that either prioritizes power savings while lim-

iting performance or drives performance while forc-

ing a high-power consumption conﬁguration. The

dynamic nature by which IDO selects different pro-

ﬁles for the application allows for application intent

to drive the minimum power consumed based on re-

specting a KPI. When the application has a lower KPI

requirement or is simply not busy enough to cause it

to miss a KPI, IDO can be used to ensure the minimal

power needed to meet the KPI is consumed.

5 RELATED WORK

The European Telecommunications Standards Insti-

tute (ETSI) has highlighted the necessity for declar-

ative management in evolving telecommunications

networks, emphasizing the deﬁnition of desired out-

comes without specifying declarative resource asks.

This approach simpliﬁes network operations, reduces

operational costs, and accelerates service deployment,

making intent-driven applications crucial for modern

management (Cai et al., 2023). This aligns with the

goals of the IDO model as presented in this paper.

The increasing demand for compute capacity,

driven in part by the rapid growth of AI/ML appli-

cations, necessitates rethinking of power and energy

management strategies (Lin et al., 2024). The growth

in use of these applications may place signiﬁcant

stress on power grids, emphasizing the need for user-

friendly energy management solutions. The method-

ology proposed in this paper is a pivotal piece of this

puzzle, enabling users to embrace hardware abstrac-

tion while seamlessly utilize platform-level power

management features.

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

(a) Average token creation time (b) Node level power draw (c) Application power estimation

Figure 6: Effect of intent-driven power management on an AI LLM application.

Power capping and frequency management have

proven effective in controlling the power draw of AI-

intensive applications (Patel et al., 2024). The utiliza-

tion of platform features has generally been demon-

strated to enhance performance and power efﬁciency

across a wide range of applications (Veitch et al.,

2024). This paper builds on that foundation by sim-

plifying the adoption of these features. The hetero-

geneous nature of edge locations adds complexity to

managing the cloud-edge continuum, calling for more

autonomous management and abstraction. Research

has shown that abstracting the complexity of edge in-

frastructure and using service-level objectives to de-

ﬁne service requirements yields signiﬁcant beneﬁts

(Guim et al., 2022). These ﬁndings form the foun-

dation for the follow-up work presented here.

Efﬁcient orchestration and resource management

has been tackled from multiple angles (Metsch et al.,

2015). Previous work has addressed individual prob-

lems such as auto-scaling (Roy et al., 2011), place-

ment (Bobroff et al., 2007), and overbooking to max-

imize utilization (Tom

as and Tordsson, 2013). How-

ever, these solutions focus primarily on efﬁcient re-

source usage rather than ease of use for energy efﬁ-

ciency. The work presented is complementary, poten-

tially enhancing the methodology proposed here.

Coordination between compute demand in data

centers and (local) power grids capacities has shown

to be effective in improving carbon emissions (Lin

and Chien, 2023). These techniques require resource

providers to plan their capacity ahead and commu-

nicate this to the power grid. Efﬁcient planning is

crucial, necessitating an understanding of the inten-

tions and priorities of application owners. The intent-

driven approach discussed in this paper provides the

necessary context to resource providers.

6 CONCLUSIONS

This paper presents an IDO model that simpliﬁes

power management for application owners while pro-

viding resource providers with essential context for

efﬁcient energy and resource optimization. By en-

abling the conﬁguration of power proﬁles on a per-

application basis based on user intents, this approach

offers signiﬁcant beneﬁts towards greener solutions.

Making power-saving modes the default in data cen-

ters and at the edge is an essential step forward. Con-

textual information provided by the intents is criti-

cal, as hardware cannot effectively throttle and op-

timize without it, leaving signiﬁcant efﬁciency gains

untapped. By accepting minor performance trade-offs

can result in substantial power savings.

Our experimental results demonstrate that when

intents and their objectives allow for it, power sav-

ings of ≈ 5-55% for this sample AI application can

be achieved compared to baseline settings. We ex-

pect these gains to only increase given higher core

count systems entering the market. The IDO model

facilitates allocating and conﬁguring the necessary re-

sources for an application to meet the intents while

maintaining greater power efﬁciency compared to

conﬁguring the system as a whole.

The intent-driven model beneﬁts application own-

ers by simplifying the use of power management fea-

tures, reducing the need for deep domain-speciﬁc

knowledge. This approach bridges the gap be-

tween application owners’ needs and what resource

providers can offer, as platform features are often not

exposed to application owners. Simultaneously, it

beneﬁts resource providers by enabling more efﬁcient

resource optimization as the intents provide context,

resulting in better overall system performance and en-

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management

ergy savings.

Future work is planned to address per-application

power estimation, as current solutions lack the de-

sired accuracy. Additionally, we aim to explore car-

bon objectives built on top of power objectives. One

limitation of the current solution is the need for core

pinning, which could be replaced with a more ﬂexi-

ble model in the future. Better-informed resource al-

location can help the scheduler make more efﬁcient

decisions based on inputs from the planner, enhanc-

ing ﬂeet-wide performance and energy efﬁciency. For

example, consolidating workloads with similar power

proﬁle needs can be more efﬁcient, avoiding single

application’s high-performance demands increase the

overall power draw of a node.

By integrating these advancements, the IDO

model provides a step towards managing the increas-

ing compute demands driven by AI and ML appli-

cations, ensuring that power and energy management

strategies remain effective and user-friendly.

ACKNOWLEDGEMENTS

This paper was in-part supported by European

Union’s Horizon Europe research and innovation pro-

gramme under grant agreement No. 101135576. The

authors would like to sincerely thank Francesc Guim,

Denisio Togashi, Karol Weber and Marlow Weston

for their contributions to the IDO capability and the

methodologies discussed in this paper.

REFERENCES

Bobroff, N., Kochut, A., and Beaty, K. (2007). Dynamic

placement of virtual machines for managing sla viola-

tions. In 2007 10th IFIP/IEEE International Sympo-

sium on Integrated Network Management, pages 119–

128.

Cai, X., Deng, H., Deng, L., Elsawaf, A., Gao, S.,

Martin de Nicolas, A., Nakajima, Y., Pieczerak,

J., Triay, J., Wang, X., Xie, B., and Zafar, H.

(2023). Evolving nfv towards the next decade.

https://www.etsi.org/images/ﬁles/ETSIWhitePapers/

ETSI-WP-54-Evolving NFV towards the next

decade.pdf. [Accessed 10-07-2024].

Geurts, P., Ernst, D., and Wehenkel, L. (2006). Extremely

randomized trees. Mach. Learn., 63(1):3–42.

Guim, F., Metsch, T., Moustafa, H., Verrall, T., Carrera, D.,

Cadenelli, N., Chen, J., Doria, D., Ghadie, C., and

Prats, R. G. (2022). Autonomous lifecycle manage-

ment for resource-efﬁcient workload orchestration for

green edge computing. IEEE Transactions on Green

Communications and Networking, 6(1):571–582.

Intel Network Builders (2024). Elastic and energy pro-

portional edge computing infrastructure solution brief.

Intel.

Lin, L. and Chien, A. A. (2023). Adapting datacenter capac-

ity for greener datacenters and grid. In Proceedings of

the 14th ACM International Conference on Future En-

ergy Systems, e-Energy ’23, page 200–213, New York,

NY, USA. Association for Computing Machinery.

Lin, L., Wijayawardana, R., Rao, V., Nguyen, H.,

GNIBGA, E. W., and Chien, A. A. (2024). Exploding

ai power use: an opportunity to rethink grid planning

and management. In Proceedings of the 15th ACM

International Conference on Future and Sustainable

Energy Systems, e-Energy ’24, page 434–441, New

York, NY, USA. Association for Computing Machin-

ery.

Metsch, T., Ibidunmoye, O., Bayon-Molino, V., Butler,

J., Hern

andez-Rodriguez, F., and Elmroth, E. (2015).

Apex lake: A framework for enabling smart orches-

tration. In Proceedings of the Industrial Track of the

16th International Middleware Conference, Middle-

ware Industry ’15, pages 1–7, New York, NY, USA.

Association for Computing Machinery.

Metsch, T., Viktorsson, M., Hoban, A., Vitali, M., Iyer, R.,

and Elmroth, E. (2023). Intent-driven orchestration:

Enforcing service level objectives for cloud native de-

ployments. SN Computer Science, 4(3).

Patel, P., Choukse, E., Zhang, C., Goiri, I. n., Warrier, B.,

Mahalingam, N., and Bianchini, R. (2024). Charac-

terizing power management opportunities for llms in

the cloud. In Proceedings of the 29th ACM Interna-

tional Conference on Architectural Support for Pro-

gramming Languages and Operating Systems, Volume

3, ASPLOS ’24, page 207–222, New York, NY, USA.

Association for Computing Machinery.

Pereira, R., Couto, M., Ribeiro, F., Rua, R., Cunha, J.,

Fernandes, J. a. P., and Saraiva, J. a. (2017). En-

ergy efﬁciency across programming languages: how

do energy, time, and memory relate? In Proceedings

of the 10th ACM SIGPLAN International Conference

on Software Language Engineering, SLE 2017, page

256–267, New York, NY, USA. Association for Com-

puting Machinery.

Roy, N., Dubey, A., and Gokhale, A. (2011). Efﬁcient

autoscaling in the cloud using predictive models for

workload forecasting. In 2011 IEEE 4th International

Conference on Cloud Computing, pages 500–507.

Tom

as, L. and Tordsson, J. (2013). Improving cloud infras-

tructure utilization through overbooking. In Proceed-

ings of the 2013 ACM Cloud and Autonomic Comput-

ing Conference, CAC ’13, pages 1–10, New York, NY,

USA. Association for Computing Machinery.

Veitch, P., MacNamara, C., and Browne, J. J. (2024). Un-

core frequency tuning for energy efﬁciency in mixed

priority cloud native edge. In 2024 Joint European

Conference on Networks and Communications & 6G

Summit (EuCNC/6G Summit), pages 925–930.

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

APPENDIX

The Figs. 7 and 8 present screenshots of the dashboard for the IDO planner. The tables in the screenshots display

the timestamps, the current state and desired state (as deﬁned by their objectives), and the actions taken. As time

progresses, the outcomes of the decisions are reﬂected in the upper rows of the table.

Figure 7: Dashboard showcasing the intent-driven orchestration planner’s decisions based on performance related objectives.

Figure 8: Dashboard showcasing the intent-driven orchestration planner’s decisions based on power related objectives.

Breaking Barriers: From Black Box to Intent-Driven, User-Friendly Power Management