Ecovisor A Virtual Energy System for Carbon-Efﬁcient Applications Abel Souza Noman Bashir Jorge Murillo Walid Hanafy Qianlin Liang David Irwin and Prashant Shenoy

2025-05-04 0 0 2.79MB 15 页 10玖币

侵权投诉

Ecovisor: A Virtual Energy System for Carbon-Efﬁcient Applications*

Abel Souza, Noman Bashir, Jorge Murillo, Walid Hanafy, Qianlin Liang, David Irwin, and

Prashant Shenoy

University of Massachusetts Amherst

Abstract

Cloud platforms’ rapid growth is raising signiﬁcant

concerns about their carbon emissions. To reduce emis-

sions, future cloud platforms will need to increase their

reliance on renewable energy sources, such as solar and

wind, which have zero emissions but are highly unreli-

able. Unfortunately, today’s energy systems effectively

mask this unreliability in hardware, which prevents appli-

cations from optimizing their carbon-efﬁciency, or work

done per kilogram of carbon emitted. To address this

problem, we design an “ecovisor,” which virtualizes the

energy system and exposes software-deﬁned control of it

to applications. An ecovisor enables each application to

handle clean energy’s unreliability in software based on

its own speciﬁc requirements. We implement a small-scale

ecovisor prototype that virtualizes a physical energy sys-

tem to enable software-based application-level i) visibility

into variable grid carbon-intensity and renewable gener-

ation and ii) control of server power usage and battery

charging/discharging. We evaluate the ecovisor approach

by showing how multiple applications can concurrently

exercise their virtual energy system in different ways to

better optimize carbon-efﬁciency based on their speciﬁc

requirements compared to a general system-wide policy.

1. Introduction

Cloud platforms are growing exponentially, and have been

for some time, with a recent analysis estimating a 6

in-

crease in their capacity from 2010-2018, or roughly a

22.4% increase per year [

]. This hyperscale growth is

being driven by the continual development of new and

useful, but often computationally-intensive, applications,

particularly in artiﬁcial intelligence (AI) [

]. As they

have grown, cloud platforms have aggressively optimized

their energy-efﬁciency, e.g., by reducing their power us-

age effectiveness (PUE)

to near the optimal value of

[

], to mitigate large increases in their energy con-

sumption and cost.

However, further improving energy-efﬁciency is be-

coming increasingly challenging, as it is already highly

optimized. Thus, continued growth in cloud capacity

will likely result in much larger increases in energy con-

sumption moving forward. Of course, this energy growth

*Paper to appear at ASPLOS 2023.

1The PUE is the ratio of total data center power to server power.

is also increasing cloud platforms’ carbon and green-

house gas (GHG) emissions, which are causing climate

change [

]. The negative environmental effects of the

cloud’s hyperscaler growth have begun to receive signiﬁ-

cant attention. As a result, all the major cloud providers

have announced aggressive goals for reducing, and ul-

timately eliminating, their platforms’ carbon emissions

over the next decade, while acknowledging that many of

the technologies necessary to achieve these sustainability

goals have yet to be developed [1,20,31,54,65].

Reducing cloud platforms’ carbon emissions will re-

quire them to power their cloud and edge data centers

using cleaner “lower-carbon” energy sources. A distin-

guishing characteristic of clean energy is its unreliability:

it is intermittent and not available in unlimited quantities

at any single location all the time. Notably, clean energy’s

unreliability manifests itself in two distinct ways within

our current energy system: i) the unreliability of renew-

able power generation and ii) the volatility of grid power’s

carbon-intensity. In the former case, the power generated

by zero-carbon renewable energy sources, primarily solar

and wind, at any location is unreliable because it varies

based on changing environmental conditions. In the latter

case, the carbon-intensity of grid power — in kg

equivalent per watt (W) — is volatile because it varies

based on the carbon emissions of the different types of

generators the grid uses to satisfy its variable demand. As

we discuss, both forms of unreliability are important to

consider in reducing cloud platforms’ carbon emissions.

Compared to other industries, computing is uniquely

well-positioned to reduce its carbon emissions by tran-

sitioning to cleaner energy sources, despite their unreli-

ability, for numerous reasons. Most importantly, com-

putation often has signiﬁcant spatial, temporal, and per-

formance ﬂexibility, which enables execution location,

time, and intensity shifting to better align with the avail-

ability of low-carbon grid power and zero-carbon renew-

able power [

]. In addition, computation can

also leverage numerous software-based fault-tolerance

techniques, including checkpointing, replication, and re-

computation, to continue execution despite unexpected

variations in the availability of low-carbon energy, which

may require throttling or shutting down servers [64].

Unfortunately, today’s cloud applications cannot lever-

age the unique combination of advantages above to opti-

mize their carbon-efﬁciency, or work done per kilogram

arXiv:2210.04951v1 [cs.OS] 10 Oct 2022

(kg) of carbon (and other GHGs) emitted, because current

energy systems effectively mask clean energy’s unrelia-

bility from them in hardware. That is, energy systems

have traditionally exposed a reliability abstraction – the

abstraction of a reliable supply of power on demand up to

some maximum – to electrical devices, including servers,

via their electrical socket interface. In many cases, the

energy system now includes a connection to not only the

grid, but also an increasingly rich local energy system

that may include substantial energy storage, e.g., batter-

ies [

], and co-located renewable energy sources, e.g.,

wind and solar [

]. Since energy systems hide their

increasing complexity behind the reliability abstraction,

they provide applications no control of, or visibility into,

the characteristics of their energy supply, i.e., its con-

sumption, generation, or carbon emissions. Thus, appli-

cations cannot optimize carbon-efﬁciency by regulating

their power usage to respond to changes in grid power’s

carbon-intensity and renewable power’s availability.

To address the problem, this paper presents the design

and implementation of an ecovisor—a software system

that exposes software-deﬁned control of a virtual energy

system directly to applications. An ecovisor is akin to a

hypervisor but virtualizes the energy system of comput-

ing infrastructure instead of virtualizing the computing

resources of a single server. Importantly, an ecovisor en-

ables applications to handle clean energy’s unreliability

within their software stack based on their own speciﬁc

characteristics, performance requirements, and sustain-

ability goals by leveraging one or more dimensions of

software ﬂexibility and software-based fault-tolerance.

Ecovisors also enable applications to exercise software-

based control of their virtual energy system to mitigate

clean energy’s unreliability. Speciﬁcally, instead of tem-

porally or spatially shifting their computing workload,

applications can control their virtual battery to temporally

shift their clean energy usage—by storing renewable or

low-carbon grid energy when it is available for later use.

In some sense, our approach extends the end-to-end

principle [

] to the energy system by i) recognizing that

the energy system’s current reliability abstraction prevents

designing carbon-efﬁcient applications, and ii) address-

ing the problem by pushing control of the energy system

from hardware into software. Our approach is also in-

spired by the exokernel argument from operating systems

that advocates delegating resource management to appli-

cations [

]. Our ecovisor extends this approach by

delegating not only resource management to applications,

but also the energy (and carbon) that powers those re-

sources. Our hypothesis is that exposing software-deﬁned

visibility and control of a virtualized energy system en-

ables applications to better optimize carbon-efﬁciency

based on their speciﬁc characteristics and requirements

compared to a general system policy. In evaluating our

hypothesis, this paper makes the following contributions.

Virtualizing the Energy System.

We present our ecov-

isor design, which virtualizes a physical energy system

to enable software-based control of server power con-

sumption and battery charging/discharging, as well as

visibility into variable grid carbon-intensity and renew-

able generation. Our ecovisor exposes a software API to

applications that enable them to control their use of power

to respond to uncontrollable variations in grid energy’s

carbon-intensity and renewable energy’s availability.

Carbon-Efﬁciency Optimizations.

We present multi-

ple case studies showing how a range of different ap-

plications can use the ecovisor API to optimize their

carbon-efﬁciency. Our case studies highlight two im-

portant concepts including: i) different applications use

their virtual energy system in different ways to opti-

mize carbon-efﬁciency, and ii) application-speciﬁc poli-

cies can better optimize carbon-efﬁciency compared to

general one-size-ﬁts-all system policies. While optimiz-

ing energy-efﬁciency has been well-studied in comput-

ing, there has been little research on optimizing carbon-

efﬁciency, which is fundamentally different and the only

metric that really matters for addressing climate change.

Implementation and Evaluation.

We implement a

small-scale ecovisor prototype on a cluster of mi-

croservers that exposes a virtual grid connection, solar

array, and batteries to applications. We evaluate our pro-

totype’s ﬂexibility by concurrently executing the case

study applications above, and showing that optimizing

their carbon-efﬁciency on a shared infrastructure requires

application-speciﬁc policies. For example, an interactive

web service may use carbon budgeting to maintain a strict

latency SLO as carbon-intensity varies, while a parallel

batch job might instead adjust its degree of parallelism.

2. Motivation and Background

Motivation

. Sustainable computing focuses on the design

and operation of carbon-efﬁcient computing infrastructure

and applications. This paper focuses on reducing Scope 2

operational carbon (and other GHG) emissions from using

electricity [

], which represents a signiﬁcant fraction of

cloud platforms’ emissions. Optimizing Scope 1 direct

emissions and Scope 3 embodied emissions are outside

our scope. While these other classes of emissions are also

important, cloud platforms have few Scope 1 emissions,

and have no direct control over their Scope 3 emissions.

While cloud platforms have long focused on optimizing

energy-efﬁciency, optimizing carbon-efﬁciency is funda-

mentally different. To illustrate, consider that a highly

energy-efﬁcient system can be highly carbon-inefﬁcient if

its grid-supplied power derives from burning fossil fuels,

while a highly energy-inefﬁcient system can be highly

carbon-efﬁcient if its power derives solely from zero-

carbon renewable energy. As this trivial example shows,

a cloud platform’s carbon-efﬁciency depends, in part, on

the carbon-intensity of its energy supply, which varies

widely over time based on variations in both grid power’s

carbon-intensity and local renewable power’s availability.

Since modifying a cloud platform’s operations to adapt

to variations in carbon-intensity, e.g., by throttling work-

loads when carbon-intensity is high, is challenging, cloud

providers have largely focused on transparently reducing

their net carbon emissions using carbon offsets. Such off-

sets are an accounting mechanism that enables offsetting

the direct use of carbon-intensive energy by purchasing

zero-carbon renewable energy generated at another time

and location [

]. Carbon offsets are attractive be-

cause they do not require complex operational changes

to reduce net carbon emissions. Many prominent tech-

nology companies have eliminated their net carbon emis-

sions [

], which they often refer to as running

on “100% renewable energy.” Unfortunately, carbon off-

sets do not reduce direct carbon emissions, and become

increasingly less effective as carbon emissions decrease,

as there is less carbon left to offset. As a result, elimi-

nating absolute carbon emissions will ultimately require

cloud platforms to change their operations to reduce their

direct carbon emissions by better aligning their computing

load with when and where low-carbon energy is available.

Reducing direct carbon emissions is challenging largely

because it introduces a new constraint that requires

voluntarily making difﬁcult tradeoffs between perfor-

mance/availability, cost, and carbon emissions. In gen-

eral, modifying design and operations to reduce direct

carbon emissions decreases performance/availability and

also increases cost, as energy prices do not (yet) incor-

porate the cost of carbon’s negative externalities to the

environment. Importantly, the optimal tradeoff between

performance/availability, cost, and carbon emissions dif-

fers across applications and users. As we show in §5,

the policies for reducing the carbon emissions of delay-

tolerant batch applications are signiﬁcantly different from

those for interactive web services that must maintain a

strict latency Service Level Objective (SLO). More gen-

erally, though, cloud users, i.e., companies, have widely

different goals, strategies, and tolerances for reducing

carbon (at the expense of increased cost and lower perfor-

mance/availability), which cloud platforms do not know.

As a result, cloud platforms are not well-positioned to

manage carbon emissions at the system-level on behalf of

their users, which motivates our ecovisor’s approach of

exposing energy and carbon management to applications.

The motivation for our ecovisor’s application-level con-

trol of carbon is analogous to that for cloud auto-scaling:

all cloud platforms support elastic auto-scaling that en-

ables applications to horizontally or vertically scale their

resources in response to variations in their workload’s in-

tensity [

]. These auto-scaling policies are application-

speciﬁc for similar reasons as above, i.e., differing appli-

cation requirements and user tradeoffs between cost and

performance/availability. Our ecovisor’s API, discussed

in §3, enables similar “auto-scaling” but in response to

variations in grid power’s carbon-intensity and local re-

newable energy’s availability. A simple evolutionary path

to enabling such “carbon-scaling” using an ecovisor is

to augment existing cloud auto-scaling APIs. For exam-

ple, existing APIs, such as Amazon CloudWatch [

] and

Azure Monitor [

], already expose visibility into platform

resource usage, and could easily be extended to include

power and carbon information. In this case, cloud plat-

forms would “delegate” carbon-scaling to applications

just as they currently delegate auto-scaling resources.

While the ecovisor approach could apply to existing

cloud platforms, especially those hosted at datacenters

with substantial co-located renewables [

] and energy

storage [

], there is currently no ﬁnancial incentive to

reduce carbon. This is a social problem, not a technical

one. In the end, to halt climate change, government poli-

cies will likely be necessary to create strong incentives for

monitoring and reducing carbon emissions, either directly,

e.g., via carbon caps, or indirectly, e.g., via carbon pricing.

Nevertheless, cloud platforms have already begun to ex-

pose visibility into their carbon emissions [

], driven by

their customers’ increasing desire to measure and report

carbon emissions data. This combination of customer

demand and government policy is likely to incentivize

future cloud platforms to adopt ecovisor-like mechanisms

for measuring and controlling carbon emissions.

Background.

Our work assumes a datacenter’s physi-

cal energy system connects to up to three distinct power

sources: the electric grid, local batteries (or other forms

of energy storage), and local renewable generation, such

as solar or wind. The power supplied to servers (and

other computing equipment) is a mix of these three power

sources. Not all facilities will have connections to all

three power sources, and the capacity of each source may

vary. For example, many large cloud data centers may not

have local renewables, while smaller edge sites might not

require a grid connection, i.e., if they have enough local

renewables and battery capacity to be self-powered [34].

An ecovisor requires software-deﬁned monitoring and

control of both server power and the physical energy sys-

tem, i.e., power’s supply, demand, and carbon emissions.

Monitoring Power. An ecovisor must be capable of

monitoring each energy source’s power generation and

consumption. Energy system components commonly ex-

pose power monitoring via programmatic APIs. For ex-

ample, battery charge controllers, such as Tesla’s Pow-

erwall, support querying a battery’s energy level, and its

charge/discharge rate from the grid and solar [

], while

solar inverters support querying current and historical so-

lar power generation [

]. Our ecovisor builds on these

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Ecovisor:AVirtualEnergySystemforCarbon-EfcientApplications*AbelSouza,NomanBashir,JorgeMurillo,WalidHanafy,QianlinLiang,DavidIrwin,andPrashantShenoyUniversityofMassachusettsAmherstAbstractCloudplatforms'rapidgrowthisraisingsignicantconcernsabouttheircarbonemissions.Toreduceemis-sions,futurecloudpla...

展开>> 收起<<

Ecovisor A Virtual Energy System for Carbon-Efﬁcient Applications Abel Souza Noman Bashir Jorge Murillo Walid Hanafy Qianlin Liang David Irwin and Prashant Shenoy.pdf

共15页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Ecovisor A Virtual Energy System for Carbon-Efﬁcient Applications Abel Souza Noman Bashir Jorge Murillo Walid Hanafy Qianlin Liang David Irwin and Prashant Shenoy

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: