FedZero: Leveraging Renewable Excess Energy in
Federated Learning
Philipp Wiesner
wiesner@tu-berlin.de
TU Berlin
Germany
Ramin Khalili
Huawei
Germany
Dennis Grinwald
dennis.grinwald@tu-berlin.de
TU Berlin
Germany
Pratik Agrawal
pratik.agrawal@tu-berlin.de
TU Berlin
Germany
Lauritz Thamsen
lauritz.thamsen@glasgow.ac.uk
University of Glasgow
United Kingdom
Odej Kao
odej.kao@tu-berlin.de
TU Berlin
Germany
ABSTRACT
Federated Learning (FL) is an emerging machine learning technique
that enables distributed model training across data silos or edge
devices without data sharing. Yet, FL inevitably introduces ineffi-
ciencies compared to centralized model training, which will further
increase the already high energy usage and associated carbon emis-
sions of machine learning in the future. One idea to reduce FL’s
carbon footprint is to schedule training jobs based on the availabil-
ity of renewable excess energy that can occur at certain times and
places in the grid. However, in the presence of such volatile and
unreliable resources, existing FL schedulers cannot always ensure
fast, efficient, and fair training.
We propose FedZero, an FL system that operates exclusively on
renewable excess energy and spare capacity of compute infrastruc-
ture to effectively reduce a training’s operational carbon emissions
to zero. Using energy and load forecasts, FedZero leverages the
spatio-temporal availability of excess resources by selecting clients
for fast convergence and fair participation. Our evaluation, based
on real solar and load traces, shows that FedZero converges sig-
nificantly faster than existing approaches under the mentioned
constraints while consuming less energy. Furthermore, it is robust
to forecasting errors and scalable to tens of thousands of clients.
CCS CONCEPTS
•Social and professional topics
→
Sustainability;•Comput-
ing methodologies →Distributed artificial intelligence.
KEYWORDS
sustainable computing, carbon efficiency, electricity curtailment,
federated learning, client selection, green AI
This work is licensed under a Creative Commons Attribution International
4.0 License.
E-Energy ’24, June 04–07, 2024, Singapore, Singapore
©2024 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0480-2/24/06
https://doi.org/10.1145/3632775.3639589
2015
2016
2017
2018
2019
2020
2021
2022
0
200
400
600
800
1000
1200
Curtailment (GWh)
Figure 1: Quarterly wind and solar curtailments by the Cali-
fornia ISO [
13
]. Leveraging this renewable excess energy in
FL can drastically reduce its operational carbon emissions.
ACM Reference Format:
Philipp Wiesner, Ramin Khalili, Dennis Grinwald, Pratik Agrawal, Lau-
ritz Thamsen, and Odej Kao. 2024. FedZero: Leveraging Renewable Ex-
cess Energy in Federated Learning. In The 15th ACM International Con-
ference on Future and Sustainable Energy Systems (E-Energy ’24), June 04–
07, 2024, Singapore, Singapore. ACM, New York, NY, USA, 13 pages. https:
//doi.org/10.1145/3632775.3639589
1 INTRODUCTION
The majority of today’s machine learning (ML) solutions perform
centralized learning, where all required training data are gathered
in a single location, usually an energy-efficient data center with
specialized hardware. Yet, in many practical use cases, it is not
feasible to collect data across a distributed system due to security
and privacy concerns or because large amounts of raw data cannot
be migrated from the deep edge to the cloud. Federated Learning
(FL) was introduced to address this issue by enabling distributed
training of ML models without transmitting training data over the
network [
37
]. In FL, we train a common ML model on clients that
cannot or do not want to share their data, by iteratively distributing
the model to a subset of them. Clients then train locally on their
own data and send back the updated models to the server, which
aggregates them before starting the next round.
Unfortunately, FL approaches require considerably more training
rounds than traditional ML and are often executed on infrastructure
that is less energy-efficient than centralized GPU clusters, resulting
in a significant increase in overall energy usage and associated
emissions [
39
,
43
,
63
]. Even without the application of FL, the train-
ing of large ML models is known to be an energy-hungry process
373
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
and has increasingly raised concerns in recent years [
17
,
52
,
61
].
As models keep growing in size and complexity, this problem is
expected to aggravate, which is why there are numerous efforts
towards more energy-efficient algorithms and hardware to reduce
the carbon footprint of AI. Yet, when focusing on reducing emis-
sions, “using renewable energy grids for training neural networks
is the single biggest change that can be made” [16, 42].
In this work, we study how the operational carbon emissions of
synchronous FL training can be reduced to zero by operating under
the hard constraint of only leveraging renewable excess energy
and spare computing capacity at cloud or edge resources. Excess
energy, also called stranded energy, occurs in electric grids when
more power is generated than demanded or when the grid does
not have sufficient capacity for transmission. If the oversupply
cannot be stored in batteries (which are expensive and only avail-
able in a limited capacity) or traded with neighboring grids (whose
excess energy patterns often correlate) the last resort is curtail-
ment, the deliberate reduction in production. Through curtailment,
the California Independent System Operator wasted more than
27 million megawatt-hours of utility-scale solar energy in 2022,
which is around 7 % of their entire solar production [
13
]. Due to
the increasing penetration of variable renewable energy sources,
the amount of curtailed energy is only expected to grow, as shown
in Figure 1. At the same time, many existing computing infrastruc-
tures are frequently underutilized or could be overclocked if the
occurrence of excess energy justifies reduced energy efficiency [
14
].
To make better use of these resources, carbon-aware computing,
i.e. considering the spatio-temporal availability of low-carbon en-
ergy during scheduling, has attracted much attention in recent
years [6, 19, 35, 44, 59, 60, 66].
FL is a promising workload for carbon-aware computing, as it
consists of energy-intensive batch jobs that are scheduled in geo-
distributed environments (to leverage spatial resource and energy
availability) without strict runtime requirements (to also leverage
temporal variations). However, as excess energy and the availability
of spare computing resources can be highly volatile, not explicitly
taking them into account during client selection can lead to signifi-
cantly longer training times due to stragglers: clients that perform
less local training than expected, or become entirely unavailable dur-
ing a training round. Furthermore, energy-agnostic selection strate-
gies can introduce biases by disproportionately selecting clients
that have a lot of excess resources available throughout the train-
ing. Yet, the idea of aligning FL scheduling with the availability of
renewable energy has so far only been studied theoretically and
under assumptions like independent and identically distributed
(iid) data and fixed “energy arrival" patterns that are not realistic
in practice and do not consider the above challenges [21].
To fill this gap, we propose FedZero, an FL system for hetero-
geneous and geo-distributed environments that utilizes forecasts
for renewable excess energy and spare computing capacity to en-
sure fast, efficient, and fair training under energy and resource
constraints. We summarize our contributions as follows:
•
We propose a system design for executing FL trainings ex-
clusively on renewable excess energy and spare computing
capacity which allows clients to share common energy bud-
gets at runtime.
•
We introduce a scalable client selection strategy that results
in fast convergence and fair client participation under vari-
able energy and resource constraints.
•
We evaluated our approach on different datasets, models,
and scenarios, to show that FedZero enables fast and energy-
efficient training while being robust to forecast errors.
•
We implemented FedZero and all baselines using Flower [
8
]
and Vessim [58] and made this code openly available1.
2 A CASE FOR FL ON EXCESS RESOURCES
FL was originally developed for use cases on mobile and edge de-
vices, where individual clients usually consume little energy. How-
ever, in recent years a variety of new application domains have
been explored to enable cross-device and cross-silo training, many
of which include clients with significant computing capabilities
and electricity demand. Examples of these novel FL settings in-
clude healthcare [
47
], the financial sector [
62
], remote sensing [
49
],
autonomous driving [
40
], and smart cities [
25
], which all consist
of complex models that require periodic re-training to adapt to
changing environments.
We argue that in environments where individual clients require
significant energy for participating in an FL training, it merits to
explicitly consider the availability of renewable excess energy in
client selection and during training to reduce carbon emissions.
2.1 Renewable Excess Energy
Due to the expanding deployment of variable renewable energy
sources such as solar and wind, it is becoming increasingly chal-
lenging to match power supply and demand at all times. If locally
occurring renewable excess energy cannot be passed on to neigh-
boring grids due to limited grid capacity and cannot be buffered in
some kind of energy storage, the only option left to operators is to
throttle supply. In this section, we describe the two main scenarios
in which renewable excess energy can occur.
The most direct way of operating IT infrastructure in a sustain-
able manner is through the use of on-site renewable energy, where
the energy source is located close to the datacenter or powerful
edge device. Within a microgrid, energy storage can buffer lim-
ited amounts of excess energy but it is expensive, entails losses,
and frequent charge cycles accelerate battery aging [
36
]. Moreover,
while more and more countries offer the possibility to sell energy
to the public grid, feed-in tariffs are usually well below purchase
prices [
46
]. Therefore, operators have a clear incentive to consume
all generated electricity directly.
The more common practice to achieve datacenters “powered
by 100 % renewable energy", as claimed by big cloud service provi-
ders [
3
,
20
,
38
], is through carbon accounting. Carbon accounting
allows operators to offset their consumption of carbon-intensive
grid energy through the purchase of renewable energy certificates.
However, today’s certificates allow that power production and
consumption can take place at vastly different times and locations,
which is why their utility for achieving science-based targets is
often questioned [
9
]. A prominent effort towards stricter carbon
accounting, often called 24/7 matching, are Google’s Time-based
Energy Attribute Certificates (T-EACs) [
54
] that are issued hourly
1https://github.com/dos-group/fedzero
374
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
and location-specific [
6
]. Similar ideas have been brought forward
by Microsoft [
41
] and Amazon [
27
]. During curtailment periods,
these certificates are expected to be very cheap and can be used for
the cost-effective execution of flexible workloads such as FL.
2.2 Computing on Spare Resources
As we do not want to promote the purchase of new hardware just
to enable the flexible scheduling of training jobs, FedZero aims
to schedule workloads exclusively on spare capacity of existing
hardware. We believe this is realistic in a wide range of scenarios
as many IT infrastructures are over-dimensioned to accommodate
peak loads. For example, public clouds, as well as emerging public
edge solutions, must maintain a relatively high proportion of spare
computing resources to sell their promise of infinite scalability.
Similarly, on-site infrastructure is often designed for peak loads
and can be severely underutilized outside these periods. A common
measure in cluster managers to increase the utilization during off-
peak hours is by deferring delay-tolerant workloads, for example
in the form of best-effort jobs [55].
Lately, several ideas have been developed to not only shift load
for increased resource utilization but for better aligning electricity
usage with grid carbon-intensity or excess energy [
35
,
44
,
59
]. More-
over, even infrastructure that is already utilized close to capacity
can be used to compute flexible workloads, if the use of other-
wise curtailed excess energy justifies the reduced energy efficiency
caused by overclocking and increased cooling [14].
3 PROBLEM STATEMENT
Our goal in this paper is to train a federated learning model with
no operational carbon emissions in an efficient and fair manner.
We aim to optimize for fast convergence and low overall energy
usage in a setting, where clients are only allowed to train on ex-
cess resources. We do so by cherry-picking clients that are likely
to have access to renewable excess energy and spare computing
capacity (or potential for overclocking) and by operating within
these constraints at runtime.
3.1 Challenges
FL on excess resources poses a number of new challenges that have
not yet been addressed by existing approaches.
Convergence speed and efficiency The availability of excess
resources can be highly variable. Thus, clients that have access to
excess energy and spare computing resources during client selec-
tion can run out of resources over the duration of a training round.
This leads to an increased number of stragglers that can severely
harm training performance. A common way to alleviate the impact
of stragglers in FL is to select more participants in each round than
actually needed, but to only wait for a number of early responses
before aggregating the results and starting a new round. The extent
of over-selection can be adapted to the environment but is usually
around 30 % in the related work [
10
,
30
,
31
]. While this makes FL
training more robust, it has the disadvantage of wasting computing
capacity and energy, since the work of some clients is discarded in
each round. Moreover, if multiple clients reside in the same power
domain, over-selection can actively harm training progress as more
clients share the same limited power source at runtime rather than
only attributing power to the most useful clients.
Common power budgets As clients can share the same source
of excess energy, we need to treat energy as a shared and limited
resource during client selection and at runtime. We use the term
power domain to describe the clustering of FL clients into groups
with access to the same source of renewable excess energy – either
because they are physically connected within a datacenter’s mi-
crogrid, or because their operation is covered through a common
budget of, for example, T-EACs (see Section 2.1). Power domains
are disjunct, meaning that one client can only be part of a single
power domain. Excess energy occurring within a power domain
must be shared by all clients within the domain. For example, in
this work, we study the behavior of FL training under resource
and energy constraints in two different solar-based scenarios: In
the first scenario, clients are spread across ten globally distributed
power domains (Figure 2a). In the second scenario, power domains
are in close geographic proximity (Figure 2b).
0 4h 8h 12h 16h 20h 24h
0
200
400
600
800
Excess power (W)
0 4h 8h 12h 16h 20h 24h
0
200
400
600
800
Excess power (W)
(a) Distributed power domains.
(b) Co-located power domains.
Figure 2: Excess power availability for different scenarios.
Fairness of participation Simply optimizing for available re-
sources in client selection will lead to a strong imbalance in favor of
clients who have a lot of excess resources available throughout the
training. In realistic settings, where the distribution of data varies
between clients, this can lead to significant biases toward certain
data in the training, which is unfair and ultimately harms model
performance. Hence, even if the availability of excess resources is
highly imbalanced, we want to ensure that all clients are able to
participate in a similar number of rounds.
Robustness and scalability To target the previous challenges,
we try to spread the client selection over different power domains
using forecasts of available excess energy and spare computing
resources – while still considering system and statistical utility.
However, forecasts usually come with a certain error, which is why
we require a solution that is robust to inaccurate forecasts.
Moreover, we need to ensure that any underlying optimization
comes with a low overhead and runtime complexity, as real-life FL
scenarios can comprise large numbers of clients.
3.2 Problem Formalization
We define
𝐶
as the set of clients distributed over a disjunct set
of power domains
𝑃
. Clients are characterized by their energy
efficiency
𝛿𝑐
and maximum computing capacity
𝑚𝑐
. For simplicity,
we do not consider other resources like memory in this work, but
they can be integrated the same way as
𝑚𝑐
. We divide time into
slots of duration
𝑡
, so an estimated training round duration
𝑑
is
always a multiple of
𝑡
. The duration of
𝑡
depends on the problem
375
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
Table 1: Overview of constants and variables
System-related constants
𝐶set of clients
𝑃set of power domains
𝐶𝑝set of clients in power domain 𝑝
𝑚𝑐
maximum capacity of client
𝑐
(batches/timestep)
𝛿𝑐energy efficiency of client 𝑐(energy/batch)
User-defined constants
𝑛number of selected clients per round
𝑑max maximum round duration in multiples of 𝑡
𝑚min
𝑐;𝑚max
𝑐
minimum/maximum number of batches client
𝑐
must participate per round
Input variables (updated each round)
𝑚spare
𝑐,𝑡 ∈ [0,𝑚𝑐]spare capacity forecast for client 𝑐at time 𝑡
𝑟𝑝,𝑡
excess energy forecast for power domain
𝑝
at
time 𝑡
𝜎𝑐fairness weighting of client 𝑐
Optimization variables (determined each round)
𝑑expected round duration
𝑏𝑐∈ {0,1}whether or not client 𝑐is selected
𝑚exp
𝑐,𝑡 ∈ [0,𝑚spare
𝑐,𝑡 ]
expected number of batches client
𝑐
will compute
at time
𝑡
considering energy and capacity con-
straints
setting but is usually in the order of one minute. We define the
training of a mini-batch, from now on called batch, as an atomic
operation to be performed within these time slots. Table 1 provides
an overview of all introduced variables and constants.
As common for FL in heterogenous environments [
10
,
34
], we
allow clients to train a variable amount of batches, but require
the configuration of a lower (
𝑚𝑚𝑖𝑛
𝑐
) and upper (
𝑚𝑚𝑎𝑥
𝑐
) bound per
client (for example, 1 to 5 local epochs). Furthermore, the server
should define the number of selected clients per round
𝑛
as well as a
maximum round duration
𝑑𝑚𝑎𝑥
after which results get aggregated,
even if not all clients did respond in time. We allow multiple clients
to share a common excess energy budget by clustering them into
power domains.
𝐶𝑝
describes the clients of a power domain
𝑝∈𝑃
.
Each power domain comprises a control plane, like an ecovisor [
51
],
which is responsible for attributing power to clients. Section 4.3
describes our client selection algorithm and optimization problem.
3.3 Boundaries
In this work, we do not explicitly consider energy storage or feeding
excess energy to the public grid, since these options are not always
available and have drawbacks compared to consuming excess en-
ergy directly [
60
]. Moreover, as we target larger-scale infrastructure
that is usually connected to the network via highly energy-efficient
optical fiber, we do not model the energy usage for data transmis-
sion. Lastly, we require the availability of excess resources during
training. Environments where relevant clients never have access
to renewable excess energy or spare computing capacity need to
default to a less radical approach and consider carbon-intensive
grid energy consumption at times.
4 SYSTEM DESIGN
An overview of FedZero’s protocol is depicted in Figure 3. The
training starts after a required amount of clients register themselves
with the server (Section 4.1,
➀
). At the beginning of each round, the
server requests forecasts on expected excess energy within power
domains and spare capacity at clients (Section 4.2,
➁
). FedZero then
selects
𝑛
clients for training, for which it expects the shortest round
duration under the given resource constraints (Section 4.3,
➂
).
This selection is performed based on the forecasts and information
on past participation or statistical utility of clients for ensuring
performance and fairness (Section 4.4). Next, the selected clients
train locally on spare capacity and, via continuous exchange with
the power domain controller, excess energy (Section 4.5,
➃
). Finally,
all participating clients send their updated model back to the server
which aggregates them and documents the participated batches
and local loss for future decisions (➄).
4.1 Client Registration
Before starting the training process, FedZero requires the following
information for each client:
(1)
The maximum computing capacity of a client is denoted as
𝑚𝑐
(batches/timestep) and can be derived from its FLOPS
(floating point operations per second), the model’s MACs
(multiply–accumulate operations), and the batch size. Alter-
natively, it can also be benchmarked before or during the
training. For variable capacity datacenters [
14
],
𝑚𝑐
should
describe the actual maximum with overclocking.
(2)
The energy efficiency is denoted as
𝛿𝑐
(energy/batch) and
can be obtained through measurements or derived from the
client’s system performance and power consumption charac-
teristics. Linear power modeling is a meaningful simplifica-
tion if we can assume power-proportional clients or sequen-
tial processing of workloads. If not,
𝛿𝑐
can also change from
round to round depending on the system utilization, which
is especially relevant in variable-capacity datacenters [14].
(3)
The control plane addresses define a client’s power domain, as
well as where to query load and excess energy forecasts from.
Load forecasts can be provided by the client itself or its clus-
ter manager/container orchestrator. Typical providers for
energy forecasts are electricity providers, microgrid control
systems, or ecovisors [51].
4.2 Forecasting Excess Energy and Load
To avoid picking clients with access to little or no resources during a
round, FedZero relies on multistep-ahead forecasts of excess energy
at power domains and spare capacity at clients.
Power production forecasts for variable renewable energy sources
like solar [
11
,
28
] and wind [
2
,
32
] are usually based on weather
models for mid- and long-term predictions as well as, in case of
solar, satellite data for short-term predictions that enable up to
5-minute resolution. For on-site installations, there exist a large
number of companies providing power production forecasts as a
service. In the case of time-based power purchase agreements, it is
the responsibility of the utility provider to inform their customers
of future energy budgets. For determining future excess energy, the
376
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
1
Round i
...
2 3 4
Preliminary
5
Round i + 1
...
Register
clients
Collect
forecasts
Select partici-
pating clients
Train locally on
excess resources
Aggregate
updates
Server
Power
Domain
Client
Power
Domain
Client
Client
Client
Training
Training
Training
Figure 3: At each training round, FedZero queries excess energy forecasts of power domains and load forecasts of individual
clients. Based on this information, it selects a set of clients for which it expects a short round duration at high statistical utility.
At runtime, clients have to periodically adjust their training performance to align with the actual available excess energy.
system furthermore needs to take load forecasts for IT infrastruc-
ture as well as co-located consumers into account. We define
𝑟𝑝,𝑡
to be the forecasted excess energy of power domain 𝑝at time 𝑡.
Load prediction is a widely researched field covering forecasts
related to application metrics, such as requests per second, as well
as the utilization of (virtualized) hardware resources like CPU, GPU,
or RAM. They usually entail time series forecasting models trained
on historical data but can also take additional context information
into account. As
𝑚𝑐
describes the maximum computing capacity
of a client in batches/timestep, we define
𝑚spare
𝑐,𝑡 ∈ [
0
,𝑚𝑐]
to be its
forecasted spare capacity at time 𝑡.
4.3 Client Selection
FedZero selects clients based on their forecasted energy and capac-
ity constraints as well as statistical utility.
Iterative search FedZero optimizes for system utility by selecting
clients that are expected to compute their
𝑚𝑚𝑖𝑛
𝑐
as fast as possi-
ble. We guarantee low computational overhead, by performing an
iterative search over possible round duration
𝑑
: For each round
duration, we solve a simple mixed-integer program (MIP) which
scales linearly with the number of clients and power domains (see
Section 5.5). For simplicity, the iterative search is described as an
incrementing for-loop in Algorithm 1. In practice, it can be imple-
mented as a binary search with complexity O(log𝑛).
On every iteration, Algorithm 1 heavily pre-filters entire power
domains (Line 6) and individual clients (Lines 11 and 8) that cannot
constitute valid solutions within the current
𝑑
, to further reduce
the runtime of the MIP. If no valid solution is found within the max-
imum round duration
𝑑𝑚𝑎𝑥
, the algorithm waits for conditions to
improve or it cloud resolve the situation by weakening constraints,
e.g. by considering grid energy. In this work, we only operate under
hard energy and capacity constraints.
Algorithm 1 Determine clients and round duration
1: 𝐶←set of clients
2: 𝑃←set of power domains
3: # search for shortest possible round duration
4: for 𝑑←1to 𝑑max do
5: # filter out power domains without excess energy
6: ¯
𝑃← {∀𝑝∈𝑃, ∀𝑡=1, ...,𝑑 :𝑟𝑝,𝑡 >0}
7:
# filter out clients that over-participated in the past (see Section 4.4)
8: ¯
𝐶← {∀𝑐∈𝐶:𝜎𝑐>0}
9:
# filter out clients without sufficient computing capacity or energy
10: for 𝑝∈¯
𝑃do
11: ¯
𝐶←¯
𝐶\ {∀𝑐∈𝐶𝑝:Í𝑑
𝑡=0min(𝑚spare
𝑐,𝑡 ,𝑟𝑝,𝑡
𝛿𝑐)<𝑚𝑚𝑖𝑛
𝑐}
12: # increase duration if there are not at least 𝑛valid clients
13: if |¯
𝐶|<𝑛then
14: continue
15: # select optimal clients
16: 𝑏←findOptimalClients(¯
𝐶, ¯
𝑃,𝑑)
17: if 𝑏is valid solution then
18: return 𝑏,𝑑
19: # wait, if no solution is found for 𝑑=𝑑𝑚𝑎𝑥
Optimization Problem For the MIP, we define two discrete
optimization variables per eligible client:
•𝑏𝑐∈ {0,1}equals 1 iff client 𝑐participates in the round
•𝑚exp
𝑐,𝑡 ∈ [
0
,𝑚spare
𝑐,𝑡 ]
denotes the expected number of batches
client 𝑐will compute at time 𝑡
377
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
The optimization problem is described as follows:
max
𝑏𝑐,𝑚exp
𝑐,𝑡 ∑︁
𝑐∈𝐶
𝑏𝑐·𝜎𝑐
𝑑
∑︁
𝑡=0
𝑚exp
𝑐,𝑡
s.t. 𝑏𝑐=1=⇒𝑚min
𝑐≤
𝑑
∑︁
𝑡=0
𝑚exp
𝑐,𝑡 ≤𝑚max
𝑐∀𝑐∈𝐶(1)
∑︁
𝑐∈𝐶𝑒
𝑚exp
𝑐,𝑡 ·𝛿𝑐≤𝑟𝑒,𝑡 ∀𝑒∈𝐸, 𝑡 =0, . . . ,𝑑 (2)
∑︁
𝑐∈𝐶
𝑏𝑐=𝑛(3)
We optimize for the maximum number of batches to be computed
within the input duration
𝑑
, weighted by each client’s statistical util-
ity
𝜎𝑐
. Equation (1) limits each selected client to compute between
𝑚𝑚𝑖𝑛
𝑐
and
𝑚𝑚𝑎𝑥
𝑐
batches. Equation (2) constrains all clients in a
power domain to not use more energy than available. Equation (3)
ensures that exactly 𝑛clients are selected per round.
Statistical utility We introduce a utility function
𝑓
:
𝐶→ {𝜎𝑐
:
∀𝑐∈𝐶}
which is invoked in every round and returns a weighting
that gives precedence to certain clients in the optimization problem.
This function can be based on the previous participation of clients,
an approximation of statistical client utility, or other user-defined
metrics, for example, to respect fairness constraints like group
parity.
The utility function applied in the remainder of this paper is
based on the statistical utility function proposed in Oort [30]:
𝜎𝑐=(|𝐵𝑐|√︃1
|𝐵𝑐|Í𝑘∈𝐵𝑐𝑙𝑜𝑠𝑠(𝑘)2,if 𝑝(𝑐) ≥ 1
1,otherwise
Oort approximates statistical client utility based on the number of
available training samples
𝐵𝑐
and the local training loss, which is
expected to correlate with the gradient norm.
4.4 Ensuring Fair Participation
When performing FL in heterogenous environments under excess
energy and capacity constraints, we actively have to take care of
avoiding biases towards powerful clients with lots of spare capacity
(
𝑚𝑐
) or clients within power domains with large amounts of excess
energy (
𝑟𝑒
). This problem is exacerbated by FedZero, which prefers
energy-efficient clients (
𝛿𝑐
) and – without further measures – tends
to select similar sets of clients in consecutive rounds, which can
harm the model’s generalization performance.
To mitigate the mentioned biases and reduce variance, we add
clients to a blocklist after they participate in a training round.
Blocked clients get assigned
𝜎𝑐=
0and are hence excluded from
future rounds. At the start of each round, clients can get released
from the blocklist with probability 𝑃(𝑐):
𝑃(𝑐)=((𝑝(𝑐) − 𝜔)−𝛼,if 𝑝(𝑐) − 𝜔>0
1,otherwise
where
𝑝(𝑐)
describes the number of rounds a client previously par-
ticipated and
𝛼
is a user-defined parameter that controls the speed
at which clients get released. A high
𝛼
will cause overparticipating
clients to remain longer on the blocklist, thereby reducing the set
of clients that FedZero can pick from. This can extend training time
but ensures fair participation. An
𝛼
close to 0 reduces the impact
of the blocklist. We consider
𝛼=
1for the remainder of this paper,
which turned out to provide the best balance between training
speed and performance in all evaluated experiments.
The parameter
𝜔
avoids decreasing release probabilities over
time and gets periodically updated to
𝜔=mean{𝑝(𝑐)
:
∀𝑐∈𝐶}
.
Users can choose a different
𝑃(𝑐)
for their use case, for example, to
improve group fairness or other custom metrics.
4.5 Executing Training Rounds
This section describes the local control loop executed by clients dur-
ing training rounds, see
➃
in Figure 3. Using the actually available
resources at runtime, each client tries to compute
𝑚𝑚𝑖𝑛
𝑐
batches
as fast as possible. Upon completion, it notifies the server but con-
tinues computation until
𝑚𝑚𝑎𝑥
𝑐
is reached. The server signals the
end of a training round and gathers all updated models once all
clients computed
𝑚𝑚𝑖𝑛
𝑐
, or once
𝑑𝑚𝑎𝑥
has passed. If a client does
not manage to compute at least
𝑚𝑚𝑖𝑛
𝑐
batches before
𝑑𝑚𝑎𝑥
, its work
is discarded to not impede the training progress, as commonly
performed in the literature [10].
Below, we discuss the two main challenges of the local control
loop: First, if multiple clients from the same power domain are par-
ticipating in the same round, they have to share a common energy
budget at runtime. Second, the actual available excess energy and
spare capacity are subject to short-term fluctuations and usually
differ from previously performed forecasts.
Sharing power at runtime If only one client within a power
domain is participating, it can make use of all available excess
energy at runtime. In this case, the capacity available for training is
defined as the minimum of the free capacity and the capacity that
can be powered using excess energy [60].
However, if two or more clients of a power domain participate
simultaneously and there is not enough energy for all of them, they
have to share a common energy budget at runtime, which has to
be coordinated by the power domain controller. To determine each
client’s share of power, we propose a simple two-step approach:
First, power is attributed to clients that have not yet reached their
minimum round participation
𝑚𝑚𝑖𝑛
𝑐
, weighted by how much energy
is still required to reach the threshold. If
𝑚𝑐𝑜𝑚𝑝
𝑐
describes the num-
ber of batches a client has already computed in the active round, this
can be written as:
𝐶𝑝→ {𝑚𝑖𝑛(
0
, 𝛿𝑐𝑚𝑚𝑖𝑛
𝑐−𝛿𝑐𝑚𝑐𝑜𝑚𝑝
𝑐)
:
∀𝑐∈𝐶𝑝}
.
Second, if there is still power left, it is attributed to all clients below
their maximum participation
𝑚𝑚𝑎𝑥
𝑐
, again weighted by the energy
required to reach this limit:
𝐶𝑝→ {𝑚𝑖𝑛(
0
, 𝛿𝑐𝑚𝑚𝑎𝑥
𝑐−𝛿𝑐𝑚𝑐𝑜𝑚𝑝
𝑐)
:
∀𝑐∈𝐶𝑝}
. As clients oblige capacity constraints and may not be
able to make use of their entire share of power, the actual distribu-
tion of power must be decided in constant consultation with clients.
Short-term variations As FedZero aims to not interfere with
other processes running on a client, the extent of local training
must be adapted over time. For example, to determine excess energy,
clients must periodically query their power domain controller, e.g.
ecovisor [
51
] and use this information to perform power capping
of tasks or entire containers [
18
,
33
]. Recently, works have been
378
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
proposed that do this in a more sophisticated manner than simply
throttling the training process’ access to resources. For example,
DISTREAL [
45
] handles the time-varying availability of resources
in FL by dynamically adjusting the computational complexity of
the trained neural network. This approach could be extended to
also consider power consumption, which is ultimately composed of
the utilization of computing resources. However, runtime behavior
in the presence of short-term resource and energy fluctuations is
currently not considered in our prototype.
5 EVALUATION
We implemented FedZero and all baselines using the FL framework
Flower [
8
] and the energy system simulator Vessim [
58
]. We ex-
tended Flower to enable discrete-event simulation over time series
datasets like excess energy availability and client load. It enables
us to perform experiments faster than in real-time by, for example,
skipping over time windows where the system is idle and waiting
for excess energy or spare capacity on clients. We use Gurobi
2
to solve the MIP. Our implementation and all datasets are openly
available (see Section 1).
5.1 Experimental Setup
To evaluate FedZero, we simulate the power usage characteristics
and performance of 100 FL clients using our Flower extension and
perform the training on four NVIDIA V100 and two RTX 5000 GPUs.
This allows us to evaluate our approach without training models
over multiple weeks and consuming megawatt hours of energy.
Table 2: Max energy consumption and training performance
of the three types of clients.
client max performance (samples per minute)
type energy DenseNet-121 EfficientNet-B1 LSTM KWT-1
small 70 W 110 118 276 87
mid 300 W 384 411 956 303
large 700 W 742 795 1856 586
Clients We model heterogeneity among clients by randomly
assigning them to one of three types (small,medium,large) that
are roughly based on the performance
3
and energy usage charac-
teristics of T4, V100, and A100 GPUs, respectively. However, we
downscaled their actual compute capabilities (samples per minute),
as shown in Table 2. We use 100 randomly selected machines from
the Alibaba GPU cluster trace dataset [
57
] to model client load
(gpu_wrk_util) and load forecasts (gpu_plan).
Scenarios For modeling power domains, we focus on on-site solar
energy generation in two scenarios based on real solar and solar
forecast data provided by Solcast
4
: A global scenario (ten globally
distributed cities from June 8-15, 2022) and a co-located scenario
(ten largest cities in Germany from July 15-22, 2022), both displayed
in Figure 2. The solar data is available in 5-minute resolution and
we assume a constant power supply for steps within this period.
2https://www.gurobi.com
3https://developer.nvidia.com/deep-learning-performance-training-inference
4https://solcast.com
power
production
Global Scenario Co-located Scenario
0 W
400 W
800 W
01234567
Time (days)
0
50
100
available
clients
01234567
Time (days)
comp. capacity
>= 60
40-60
20-40
< 20
Figure 4: Power production and client availability over the
course of both scenarios. While there are always some clients
available in the global scenario, in the co-located scenario
clients are always available around the same time.
Clients are randomly distributed over the ten power domains, which
each have a maximum output of 800 W. If there is little sun, or
multiple clients are selected within a domain, energy becomes a
limiting resource. The power and client availability is depicted
in Figure 4. The upper plot shows the energy availability within
energy domains, where each domain is represented by one line. The
lower plot depicts the availability of clients over time, color-coded
by how much of their total computational capacity is available for
training.
Datasets, models, parameters We evaluate our approach on
four datasets and models commonly used in FL evaluations.
•
CIFAR-100 [
29
] contains 60,000 32x32 color images across
100 classes. We model heterogeneous data by applying a
Dirichlet distribution with
𝛼=
0
.
5, similar to [
22
], which
skews the number of samples as well as the number of sam-
ples per class and client. We train
5
the convolutional model
DenseNet-121 [23] using FedProx [34] with 𝜇=0.1.
•
Tiny ImageNet contains 100,000 64×64 color images of 200
classes. We distribute samples to clients using the same
Dirichlet distribution as for CIFAR-100. We train
6
an Ef-
ficientNet-B1 [53] model using FedProx with 𝜇=0.1.
•
In the Sheakespare [
12
] dataset, each client represents one of
100 randomly selected speaking roles from a play. As in [
34
],
we train a two-layer LSTM
7
using FedProx with
𝜇=
0
.
001
to perform next character prediction.
•
Google Speech Commands contains more than 100,000 audio
samples of 30 different words. We randomly assigned speak-
ers to the 100 clients and train
8
the keyword transformer
model KWT-1 [7] for speech classification.
The number of samples computed per timestep was obtained
through benchmarking runs and is stated in Table 2. All simulations
use a timestep
𝑡=
1min and a max round duration
𝑑max =
60 min.
We select
𝑛=
10 clients each round which have to compute 1 to 5
local epochs, so
𝑚𝑚𝑖𝑛
and
𝑚𝑚𝑎𝑥
depend on the locally available
number of samples. Clients locally train on minibatches of size
10. We ran each experiment five times over the course of seven
5SDG optimizer , learning rate = 0.001, weight decay = 5e-4, momentum = 0.8
6Adam optimizer, learning rate = 0.001
7100 hidden units, 8D embedding layer, SDG optimizer, learning rate = 0.8, see [34]
8AdamW optimizer, learning rate = 0.001, weight decay = 0.1, see [7]
379
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
Table 3: Best accuracy and time/energy to reach the target accuracy of FedZero and the best-performing baselines.
Dataset & model Approach
Global scenario Co-located scenario
Target Best Time-to- Energy-to- Target Best Time-to- Energy-to-
accuracy accuracy accuracy accuracy accuracy accuracy accuracy accuracy
Random 1.3K
64.7 %
66.0 % 4.7 d 79.2 kWh
65.5 %
66.6 % 5.3 d 113.4 kWh
CIFAR-100 Oort 1.3K 66.4 % 4.5 d 103.8 kWh 66.4 % 5.4 d 138.7 kWh
DenseNet-121 Oort fc 65.8 % 5.3 d 102.4 kWh 66.1 % 6.4 d 126.7 kWh
FedZero 66.8 % 3.6 d 70.6 kWh 66.5 % 4.5 d 96.4 kWh
Random 1.3K
62.4 %
63.1 % 5.6 d 109.6 kWh
62.8 %
63.3 % 3.7 d 86.0 kWh
Tiny ImageNet Oort 1.3K 63.2 % 3.3 d 90.2 kWh 63.5 % 3.4 d 90.5 kWh
EfficientNet-B1 Oort fc 63.1 % 3.9 d 89.0 kWh 62.7 % - -
FedZero 63.6 % 2.9 d 67.1 kWh 63.6 % 3.4 d 75.8 kWh
Random 1.3K
50.4 %
50.7 % 4.6 d 97.9 kWh
50.9 %
51.5 % 4.5 d 90.0 kWh
Shakespeare Oort 1.3K 50.2 % - - 51.7 % 4.5 d 95.4 kWh
LSTM Oort fc 50.5 % 6.7 d 157.4 kWh 50.5 % - -
FedZero 53.1 % 1.8 d 40.0 kWh 53.1 % 2.3 d 42.8 kWh
Random 1.3K
83.6 %
85.2 % 4.8 d 103.5 kWh
82.8 %
85.1 % 4.3 d 80.8 kWh
Google Speech Oort 1.3K 86.9 % 3.6 d 99.0 kWh 86.4 % 3.4 d 85.0 kWh
KWT-1 Oort fc 87.0 % 3.7 d 86.2 kWh 84.9 % 3.7 d 76.6 kWh
FedZero 87.2 % 3.6 d 79.0 kWh 87.7 % 2.6 d 65.8 kWh
simulated days and report mean values.
Baselines We compare FedZero with existing approaches by
training six different baselines. First, we run all experiments using
Random client selection as well as the guided selection strategy
Oort [
30
]. We update each client’s system utility, an important factor
in Oort’s scheduling, based on the available energy and capacity in
every round. Both approaches can only select from clients, which
currently have access to excess energy and spare resources.
Second, we train the above baselines again but this time allow
them to select 1
.
3
𝑛
clients per round. Over-selection is commonly
employed in the related work [
10
,
30
] to counteract inefficiencies
caused by stragglers in unreliable environments. Once
𝑛
clients
have returned their results a new round starts. The baselines are
called Random 1.3n and Oort 1.3n.
Third, we want to demonstrate that access to forecasts alone is
not the decisive advantage over existing approaches. For this, we
train two baselines Random fc and Oort fc that only select 10 clients,
but have access to load and energy forecasts for filtering out clients
that are not expected to reach their 𝑚min
𝑐within 𝑑𝑚𝑎𝑥 .
Lastly, for each experiment, we define an Upper bound in con-
vergence speed and performance, by training a model that uses
random client selection but is not subject to any energy constraints
or existing load on clients (clients are still heterogeneous). This
baseline is not limited to renewable excess energy.
5.2 Performance Overview
Training performance Figure 5 displays the training progress
of FedZero and the baselines over the different experiments. We
can observe that FedZero consistently outperforms all baselines
in terms of top accuracy. While in some cases the performance of
Oort/Oort 1.3n/Oort fc is comparable (see Appendix A for details),
the gap is considerable in scenarios with heavy sample imbalance
like Shakespeare (2365
±
4674 samples per client;
min =
730;
max =
27950). This is due to the fact that none of the baselines considers
common power budgets during client selection. We found that this
50
60
CIFAR100
DenseNet-121
accuracy (%)
Global scenario
Upper bound
FedZero
Random
Oort
Random 1.3n
Oort 1.3n
Random fc
Oort fc
Co-located scenario
50
60
Tiny ImageNet
EfficientNet-B1
accuracy (%)
40
50
Shakespeare
LSTM
accuracy (%)
01234567
Training time (days)
60
80
Google Speech
KWT-1
accuracy (%)
01234567
Training time (days)
Figure 5: Training progress of all experiments.
380
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
problem is exacerbated for strategies that select clients based on
statistical utility, such as Oort: If a power domain does not have
access to excess energy for an extended period of time, the statistical
utility of its clients is usually high as they have not participated for
many rounds. Once excess energy is available, Oort heavily targets
these clients which leads to increased competition for energy at
runtime and, therefore, slower training progress.
For all datasets but CIFAR-100, FedZero reaches, or almost reaches,
the top accuracy of the Upper bound, suggesting that varying re-
source availability does not necessarily harm training performance,
but only increases the total training time.
Time-to-accuracy and energy-to-accuracy To further demon-
strate how FedZero improves FL under energy and capacity con-
straints, we define the top accuracy of the Random baseline as our
target accuracy for a specific experiment. Table 3 reports the time
and energy required for FedZero and the baselines to reach this tar-
get accuracy. The table only contains the best-performing baselines;
a full table with all results can be found in Appendix A.
FedZero has the lowest time-to-accuracy and energy-to-accuracy
across all experiments. On average, it reached the target accuracy
around 35 % faster in the global scenario and around 26 % faster
in the co-located scenario than Random 1.3n and Oort 1.3n, which
were among the fastest and most energy-efficient baselines. At the
same time, FedZero was using 36 % less energy on average in the
global scenario and 30 % less in the co-located scenario. Oort-based
baselines generally outperform the Random-based ones in terms
of top accuracy and convergence speed but at the cost of higher
energy usage. However, some Oort-based baselines do not reach the
target accuracy at all, due to the previously described inefficiencies
caused by selecting clients from the same power domain.
Round durations As FedZero knows about the system utility
and resource availability, it avoids combining clients with vastly dif-
ferent expected round durations. For example, in the global scenario
on CIFAR-100 it required 15.1
±
8.5 min per round. For comparison,
the Random baseline had an average round duration of 33.7
±
19.6
min, which was lowered to 22.7
±
17.7 min and 27.8
±
17.3 min by
Random 1.3n and Random fc, respectively. The round duration of
Oort-based baselines was 18.6±14.5 min on average.
The same applies to the co-located scenario, where all clients are
available around the same time. Here, Random-based baselines take
15.5
±
12.7 min and Oort-based baselines 12.0
±
13.0 min on average.
FedZero only requires 9.7
±
7.6 min, allowing it to perform consider-
ably more training rounds within the same time. This observation
is consistent across all experiments.
5.3 Fairness of Participation
When training under energy and capacity constraints, we inevitably
introduce biases towards clients that have lots of excess resources
available. To illustrate this, Figure 6a displays the average percent-
age of rounds in which clients have participated in the training for
the CIFAR-100 global scenario, grouped by power domain. As we
select 10 out of 100 clients per round, we ideally expect an average
client participation of 10 %. However, some power domains have
access to more excess energy than others, resulting in the Random
Table 4: CIFAR-100 performance on the global scenario under
imbalanced conditions (Berlin has unlimited resources).
Best accuracy Time-to-acc. Energy-to-acc.
Random 64.6 % 6.7 d 95.7 kWh
Oort 65.6 % 4.5 d 189.4 kWh
FedZero 66.9 % 3.5 d 83.4 kWh
0
10
20
Random
std=2.00 std=3.88
0
10
20
participation per domain (%)
Oort
std=1.95 std=9.96
37.89
Berlin
Cape Town
Hong Kong
Lagos
Mexico City
Mumbai
San Francisco
Stockholm
Sydney
São Paulo
Power Domains
0
10
20
FedZero
std=0.82
Berlin
Cape Town
Hong Kong
Lagos
Mexico City
Mumbai
San Francisco
Stockholm
Sydney
São Paulo
Power Domains
std=0.85
(a) Client participation
per power domain.
(b) Client participation with
unlimited resources for Berlin.
Figure 6: FedZero ensures fair participation of clients, even
under highly imbalanced conditions.
and Oort strategies to favor clients in these domains. We can ob-
serve that FedZero exhibits a much more balanced participation
within (marked by the error bar) as well as between power domains
(std as stated on each figure).
We conducted an additional set of experiments on the same
scenario, where the Berlin power domain has access to unlimited
excess energy and all clients within Berlin have unlimited comput-
ing resources available. The results of this experiment are displayed
in Figure 6b, where Berlin is colored in red. As clients in this power
domain are now always available for training, the Random baseline
almost doubles their participation from 11.0+-2.1 to 19.8+-2.6 %.
Even worse, Oort, which actively targets clients with high system
utility, more than triples the participation of Berlin clients from
12.0
±
2.9 to 37.9
±
1.3 %. Oort describes a mechanism for combining
its selection with a user-defined fairness metric. However, we found
that Oort must rely almost entirely on our fairness metric, hence,
disregard system and statistical utility, when attempting to achieve
fairness comparable to FedZero. Other than all baselines, which
introduce significant biases, FedZero only slightly leverages the
additional resources by increasing the mean participation of clients
in the domain from 10.2±0.3 to 11.3±0.3 %.
381
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
Table 4 displays the training performance of the three approaches
in the scenario where Berlin has unlimited resources. We observe
that Random used 19 % more energy and Oort even twice the energy
in the imbalanced scenario (Figure 6b) for reaching a comparable
accuracy as in the base scenario (Figure 6a). FedZero used only 4 %
more energy and reduced its time-to-accuracy.
Based on our results on fairness of participation, we expect
FedZero to also improve other fairness metrics such as accuracy
parity between clients. However, these metrics depend on vari-
ous additional factors like non-iid data distribution among power
domains, which is why we leave this analysis to future work.
5.4 Robustness Against Forecasting Errors
To investigate the impact of forecast quality on FedZero’s perfor-
mance, we performed further experiments based on the global sce-
nario on the Tiny ImageNet and Google Speech datasets. Figure 7
shows the training progress and distribution of round durations
for Tiny ImageNet. FedZero w/ error uses forecasts with realistic
errors as in all previous experiments, FedZero w/o error uses perfect
forecasts, and FedZero w/ error (no load) uses realistic errors for
excess energy but has no forecasts for spare capacity available, as
short term load might not always be predictable in every setting.
Note, that FedZero is not able to operate if there are no predictions
of excess energy at all, for example, due to communication loss.
01234567
Training time (days)
55
60
accuracy (%)
FedZero w/o error
FedZero w/ error
FedZero w/ error (no load)
Random 1.3n
Random fc
Oort 1.3n
Oort fc
0 20 40 60
Round durations (min)
0.00
0.05
0.10
distribution
Figure 7: Analysis of convergence behavior and round dura-
tions of FedZero under forecasts of different quality.
The three experiments based on FedZero show small differences
in convergence speed and energy usage. While FedZero w/ error
takes 2.8 d and 65.2 kWh to reach the target accuracy, using perfect
forecasts it requires 15.4 % less time and 15.2 % less energy. This
is because FedZero becomes better at avoiding stragglers through
the use of accurate predictions, resulting in shorter, hence, more
efficient rounds, as shown in the right of Fig 7. FedZero without
load forecasts takes 8.2 % more time to reach the target accuracy,
using 10.0 % more energy. This result is of course specific to how
we modeled load in our evaluation – the effect in other contexts
may be bigger or smaller. Still, we can see that all three experiments
converge to the same accuracy of 63.8
˙
%, while consistently exhibit-
ing better time-to-accuracy and energy-to-accuracy, showing that
FedZero can perform well even with suboptimal forecast quality.
We additionally performed this analysis on the Google Speech ex-
periment and got comparable results. For example, FedZero without
errors converged 5.2 % faster using 6.7 % less energy than FedZero
with realistic forecasts.
5.5 Overhead and Scalability
We analyze the overhead of FedZero’s client selection by profiling
the runtime of Algorithm 1, including the MIP, on an Apple M1 pro-
cessor. Each experiment was repeated 5 times and we report mean
values. Figure 8a shows the linear growth of runtime in regards to
the number of clients: Even at the biggest evaluated setting, 100k
clients distributed over 100k power domains searching over 1440
timesteps (24 hours in 1-minute resolution), the algorithm returns
within two minutes. For scenarios in the scale of the previous eval-
uation (100 clients, 10 power domains, 60 timesteps) it only takes
around 0.1 seconds to decide on a set of clients. We observe that
due to the
O(log𝑛)
runtime of the binary search, increasing the
timestep search space from 60 to 1440 (factor 24) only increases
the runtime by factor 1.8. Figure 8b shows the runtime of a single
MIP for different numbers of clients and power domains (note, that
the y-axis is linear in this figure). We observe, that the number of
power domains has little to no impact on the runtime for up to
10k domains. Increasing the number of power domains from 10k to
100k only increases the runtime from 15.4 to 20.1 seconds.
10 100 1000 10k 100k
number of clients and
power domains
0.01
0.1
1
10
100
runtime (seconds)
number of time steps
1
10
60
1440
10 100 1000 10k 100k
number of power domains
0
5
10
15
20
number of clients
1000 10k 100k
(a) Influence of timestep
search space.
(b) Influence of number of
clients and power domains.
Figure 8: FedZero overhead analysis.
In terms of communication overhead, FedZero requires excess
energy and load forecasts at the beginning of each round. Further-
more, clients periodically (in our evaluation minutely) sync with
their power domain to align their performance with the actual
available excess energy at training time. As both payloads are in
the order of kilobytes and we are assuming clients are connected
via fiber, this overhead is negligible.
6 RELATED WORK
Carbon-aware computing As carbon pricing mechanisms, such
as emission trading systems or carbon taxes, are starting to be
implemented around the globe [
5
], the IT industry is pushing to
increase the usage of low-carbon energy in datacenters. Carbon-
aware computing tries to reduce the emissions associated with
computing by shifting flexible workloads towards times [
19
,
44
,
59
] and locations [
66
,
67
] with clean energy. For example, Google
defers delay-tolerant workloads when power is associated with
high carbon intensity as a measure to reach their 24/7 carbon-free
target by 2030 [44].
While most research in carbon-aware computing aims at con-
suming cleaner energy from the public grid [
35
,
44
,
59
,
67
], recent
works also try to better exploit excess energy, similar to FedZero.
382
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
Cucumber [
60
] is an admission control policy that accepts low-
priority workloads on underutilized infrastructure, only if they
can be computed using excess energy. Similarly, Zheng et al. [
66
]
explore workload migration on underutilized data centers as a mea-
sure to reduce curtailment. The Zero-Carbon Cloud [
15
] already
targets the problem of curtailment at the level of infrastructure
planning, by placing data centers close to renewable energy sources.
Carbon footprint of ML The training of large ML models is
a highly relevant domain for carbon-awareness, due to the often
excessive energy requirements on the one hand, and certain flex-
ibility in scheduling on the other. Other than inference, which is
usually expected to happen at low latency, ML training jobs can
often be stopped and resumed, scaled up or down, or even migrated
between locations. Because of this, many papers have previously
addressed the carbon emissions of centralized ML [16, 17, 42, 52].
Qiu et al. [
43
] were the first to broadly study the energy con-
sumption and carbon footprint of FL and state that, "depending
on the configuration, FL can emit up to two orders of magnitude
more carbon than centralized machine learning." Further studies
investigate the carbon impact of hyperparameters such as con-
currency rate [
64
] or the cost of differential privacy [
48
]. Carbon
awareness in the context of FL has so far only been explored by
Güler and Yener [
21
] who define a model for intermittent energy
arrivals and propose a scheduler with provable convergence guar-
antees. However, their assumptions regarding energy arrivals are
highly simplified as they neither consider spare capacity on clients
nor non-iid data distributions. Moreover, other than FedZero, their
model does not allow multiple clients to share the same power
domain.
FL client selection Active (or guided) client selection in FL has
received significant attention in recent years, as researchers try to
improve the final accuracy, convergence speed, reliability, fairness,
or reduce communication overhead compared to random client
selection [
50
]. For example, Oort [
30
] exploits heterogeneous de-
vice capabilities and data characteristics by cherry-picking clients
with high statistical model efficiency as well as high system utility,
resulting in faster convergence and better final accuracy. Similarly,
other novel approaches like FedMarl [
65
] or Power-of-Choice [
24
]
utilize the local training loss of clients to bias the selection. FedZero
does not try to compete with but can be adapted to integrate with
other client selection strategies.
Few energy-aware client selection strategies exist, like EAFL [
4
],
that extend Oort’s utility function to additionally consider the bat-
tery level of clients. However, like most research addressing en-
ergy usage in FL [
56
,
63
], it aims to increase the operating time
of battery-constrained end devices and is not concerned with the
overall emissions associated with the training. FedZero describes
the first approach for FL training solely on excess energy and spare
computing capacity.
7 CONCLUSION
This paper proposes FedZero, a system design for fast, efficient,
and fair training of FL models using only renewable excess energy
and spare computing capacity. Our results show, that FedZero’s
client selection strategy converges significantly faster than all base-
lines under the mentioned resource constraints while ensuring fair
participation of clients, even under highly imbalanced conditions.
Moreover, our approach is robust against forecasting errors and
scalable to large-scale, globally distributed scenarios.
In future work, we want to investigate the impact of periodic
patterns in excess energy availability on training performance [
68
].
Furthermore, we plan to integrate FedZero with novel asynchro-
nous [
26
] or semi-synchronous [
1
] strategies, while explicitly taking
energy storage and grid carbon intensity into account.
ACKNOWLEDGMENTS
We sincerely thank Solcast for providing us with free access to their
solar forecast APIs. We furthermore want to thank the anonymous
reviewers of ICDCS ’23 and e-Energy ’24 for their helpful comments.
This research was supported by the German Academic Exchange
Service (DAAD) as ide3a and IFI as well as the German Ministry
for Education and Research (BMBF) as
BIFOLD
(grant 01IS18025A)
and Software Campus (grant 01IS17050).
REFERENCES
[1]
Ahmed M. Abdelmoniem, Atal Narayan Sahu, Marco Canini, and Suhaib A.
Fahmy. 2023. REFL: Resource-Efficient Federated Learning. In EuroSys. ACM.
https://doi.org/10.1145/3552326.3567485
[2]
David B. Alencar, Carolina de Mattos Affonso, Roberto C. L. Oliveira, Jorge
Laureano Moya Rodríguez, Jandecy Cabral Leite, and Jose Carlos R. Filho. 2017.
Different Models for Forecasting Wind Power Generation: Case Study. Energies
10 (2017). https://doi.org/10.3390/en10121976
[3] Amazon. 2022. Amazon’s 2022 Sustainability Report. (2022).
[4]
Amna Arouj and Ahmed M. Abdelmoniem. 2022. Towards Energy-Aware Fed-
erated Learning on Battery-Powered Clients. In Workshop on Data Privacy and
Federated Learning Technologies for Mobile Edge Network at ACM MobiCom.
[5]
World Bank. 2022. State and Trends of Carbon Pricing 2022. Technical Report.
Washington, DC: World Bank.
[6]
Noman Bashir, David Irwin, Prashant Shenoy, and Abel Souza. 2022. Sustainable
Computing - Without the Hot Air. In HotCarbon.
[7]
Axel Berg, Mark O’Connor, and Miguel Tairum Cruz. 2021. Keyword Transformer:
A Self-Attention Model for Keyword Spotting. In Proc. Interspeech 2021. 4249–4253.
https://doi.org/10.21437/Interspeech.2021-1286
[8]
Daniel J Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Titouan Parcollet, and
Nicholas D Lane. 2020. Flower: A Friendly Federated Learning Research Frame-
work. arXiv preprint arXiv:2007.14390 (2020).
[9]
Anders Bjørn, Shannon M. Lloyd, Matthew Brander, and H. Damon Matthews.
2022. Renewable energy certificates threaten the integrity of corporate science-
based targets. Nature Climate Change 12, 6 (2022).
[10]
Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex
Ingerman, Vladimir Ivanov, Chloé Kiddon, Jakub Konečný, Stefano Mazzoc-
chi, Brendan McMahan, Timon Van Overveldt, David Petrou, Daniel Ram-
age, and Jason Roselander. 2019. Towards Federated Learning at Scale:
System Design. In MLSys. https://proceedings.mlsys.org/paper/2019/file/
bd686fd640be98efaae0091fa301e613-Paper.pdf
[11]
Jamie M. Bright, Sven Killinger, David Lingfors, and Nicholas A. Engerer. 2018.
Improved satellite-derived PV power nowcasting using real-time power data
from reference PV systems. Solar Energy 168 (2018). https://doi.org/10.1016/j.
solener.2017.10.091
[12]
Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečn
`
y,
H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. 2019. LEAF: A
Benchmark for Federated Settings. In Workshop on Federated Learning for Data
Privacy and Confidentiality at NeurIPS.
[13]
California ISO. 2024. Managing oversupply. http://www.caiso.com/informed/
Pages/ManagingOversupply.aspx. accessed Jan. 2024.
[14]
Andrew Chien, Chaojie Zhang, Liuzixuan Lin, and Varsha Rao. 2022. Beyond
PUE: Flexible Datacenters Empowering the Cloud to Decarbonize. In HotCarbon.
[15]
Andrew A Chien, Chaojie Zhang, and Hai Duc Nguyen. 2019. Zero-carbon Cloud:
Research Challenges for Datacenters as Supply-following Loads. University of
Chicago, Tech. Rep. CS-TR-2019-08 (2019).
[16]
Payal Dhar. 2020. The carbon impact of artificial intelligence. Nature Machine
Intelligence 2 (2020), 423–425.
[17]
Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy
Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A. Smith, Nicole
383
E-Energy ’24, June 04–07, 2024, Singapore, Singapore Wiesner et al.
DeCario, and Will Buchanan. 2022. Measuring the Carbon Intensity of AI in
Cloud Instances. In ACM FAccT. https://doi.org/10.1145/3531146.3533234
[18]
Jonatan Enes, Guillaume Fieni, Roberto R. Expósito, Romain Rouvoy, and Juan
Touriño. 2020. Power Budgeting of Big Data Applications in Container-based
Clusters. In IEEE CLUSTER.
[19]
Gilbert Fridgen, Marc-Fabian Körner, Steffen Walters, and Martin Weibelzahl.
2021. Not All Doom and Gloom: How Energy-Intensive and Temporally Flexible
Data Center Applications May Actually Promote Renewable Energy Sources.
Business & Information Systems Engineering 63, 3 (2021).
[20] Google. 2022. 2022 Environmental Report. (2022).
[21]
Başak Güler and Aylin Yener. 2021. A Framework for Sustainable Federated
Learning. In 2021 19th International Symposium on Modeling and Optimization
in Mobile, Ad hoc, and Wireless Networks (WiOpt). https://doi.org/10.23919/
WiOpt52861.2021.9589930
[22]
Harry Hsu, Hang Qi, and Matthew Brown. 2019. Measuring the Effects of Non-
Identical Data Distribution for Federated Visual Classification. arXiv preprint
arXiv:1909.06335 (2019).
[23]
Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger.
2017. Densely Connected Convolutional Networks. In CVPR.
[24]
Yae Jee Cho, Jianyu Wang, and Gauri Joshi. 2022. Towards Understanding Biased
Client Selection in Federated Learning. In AISTATS.
[25]
Ji Chu Jiang, Burak Kantarci, Sema Oktug, and Tolga Soyata. 2020. Federated
Learning in Smart City Sensing: Challenges and Opportunities. Sensors 20, 21
(2020).
[26]
Zhifeng Jiang, Wei Wang, Baochun Li, and Bo Li. 2022. Pisces: Efficient Federated
Learning via Guided Asynchronous Training. In ACM Symposium on Cloud
Computing (SoCC). https://doi.org/10.1145/3542929.3563463
[27]
Lucas Joppa. 2021. Made to measure: Sustainability commitment
progress and updates. Microsoft. Retrieved Sept. 2023 from
https://blogs.microsoft.com/blog/2021/07/14/made-to-measure-sustainability-
commitment-progress-and-updates
[28]
Alexandra I. Khalyasmaa, Stanislav A. Eroshenko, T. Chakravarthy, Venu Gopal
Gasi, Sandeep Kumar Yadav Bollu, Raphael Caire, Sai Kumar Reddy Atluri,
and Suresh Karrolla. 2019. Prediction of Solar Power Generation Based on
Random Forest Regressor Model. In IEEE SIBIRCON. https://doi.org/10.1109/
SIBIRCON48586.2019.8958063
[29]
Alex Krizhevsky. 2009. Learning multiple layers of features from tiny images.
Technical Report.
[30]
Fan Lai, Xiangfeng Zhu, Harsha V. Madhyastha, and Mosharaf Chowdhury. 2021.
Oort: Efficient Federated Learning via Guided Participant Selection. In USENIX
OSDI. https://www.usenix.org/conference/osdi21/presentation/lai
[31]
Chenning Li, Xiao Zeng, Mi Zhang, and Zhichao Cao. 2022. PyramidFL: A Fine-
Grained Client Selection Framework for Efficient Federated Learning. In ACM
MobiCom. https://doi.org/10.1145/3495243.3517017
[32]
Qing’an Li, Chang Cai, Yasunari Kamada, Takao Maeda, Yuto Hiromori, Shuni
Zhou, and Jianzhong Xu. 2021. Prediction of power generation of two 30 kW
Horizontal Axis Wind Turbines with Gaussian model. Energy 231 (2021). https:
//doi.org/10.1016/j.energy.2021.121075
[33]
Shaohong Li, Xi Wang, Xiao Zhang, Vasileios Kontorinis, Sreekumar Kodakara,
David Lo, and Parthasarathy Ranganathan. 2020. Thunderbolt: Throughput-
Optimized, Quality-of-Service-Aware Power Capping at Scale. In USENIX OSDI.
https://www.usenix.org/conference/osdi20/presentation/li-shaohong
[34]
Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and
Virginia Smith. 2020. Federated Optimization in Heterogeneous Networks. In
MLSys.
[35]
Liuzixuan Lin, Victor M. Zavala, and Andrew Chien. 2021. Evaluating Coupling
Models for Cloud Datacenters and Power Grids. In ACM e-Energy. https://doi.
org/10.1145/3447555.3464868
[36]
Longjun Liu, Hongbin Sun, Chao Li, Tao Li, Jingmin Xin, and Nanning Zheng.
2017. Managing Battery Aging for High Energy Availability in Green Datacenters.
IEEE Transactions on Parallel and Distributed Systems 28, 12 (2017). https://doi.
org/10.1109/TPDS.2017.2712778
[37]
H. B. McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera
y Arcas. 2016. Communication-Efficient Learning of Deep Networks from De-
centralized Data. In AISTATS.
[38] Microsoft. 2022. 2022 Environmental Sustainability Report. (2022).
[39]
Rakshit Naidu, Harshita Diddee, Ajinkya K Mulay, Aleti Vardhan, Krithika
Ramesh, and Ahmed Zamzam. 2021. Towards Quantifying the Carbon Emissions
of Differentially Private Machine Learning. In Workshop on Socially Responsible
Machine Learning at ICML.
[40]
Anh Nguyen, Tuong Do, Minh Tran, Binh X. Nguyen, Chien Duong, Tu Phan,
Erman Tjiputra, and Quang D. Tran. 2022. Deep Federated Learning for Au-
tonomous Driving. In 2022 IEEE Intelligent Vehicles Symposium (IV).
[41]
Jake Oster. 2022. How we count carbon emissions from electricity matters.
Amazon. Retrieved Sept. 2023 from https://www.amazon.science/blog/how-we-
count-carbon-emissions-from-electricity-matters
[42]
David Patterson, Joseph Gonzalez, Urs Hölzle, Quoc Le, Chen Liang, Lluis-Miquel
Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. The
Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink.
Computer 55, 7 (2022). https://doi.org/10.1109/MC.2022.3148714
[43]
Xinchi Qiu, Titouan Parcollet, Javier Fernandez-Marques, Pedro Porto Buarque
de Gusmao, Daniel J. Beutel, Taner Topal, Akhil Mathur, and Nicholas D. Lane.
2021. A first look into the carbon footprint of federated learning. arXiv preprint
arXiv:2102.07627 (2021).
[44]
Ana Radovanovic, Ross Koningstein, Ian Schneider, Bokan Chen, Alexandre
Duarte, Binz Roy, Diyue Xiao, Maya Haridasan, Patrick Hung, Nick Care, Saurav
Talukdar, Eric Mullen, Kendal Smith, Mariellen Cottman, and Walfredo Cirne.
2022. Carbon-Aware Computing for Datacenters. IEEE Transactions on Power
Systems (2022).
[45]
Martin Rapp, Ramin Khalili, Kilian Pfeiffer, and Jörg Henkel. 2022. DISTREAL:
Distributed Resource-Aware Learning in Heterogeneous Systems. In AAAI.
[46] REN21. 2022. Renewables 2022 Global Status Report. (2022).
[47]
Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Milletarì, Holger R. Roth, Shadi
Albarqouni, Spyridon Bakas, Mathieu N. Galtier, Bennett A. Landman, Klaus
Maier-Hein, Sébastien Ourselin, Micah Sheller, Ronald M. Summers, Andrew
Trask, Daguang Xu, Maximilian Baust, and M. Jorge Cardoso. 2020. The future
of digital health with federated learning. npj Digital Medicine 3, 1 (2020).
[48]
René Schwermer, Ruben Mayer, and Hans-Arno Jacobsen. 2023. Energy vs Privacy:
Estimating the Ecological Impact of Federated Learning. In ACM e-Energy.
[49]
Jinhyun So, Kevin Hsieh, Behnaz Arzani, Shadi Noghabi, Salman Avestimehr, and
Ranveer Chandra. 2022. FedSpace: An Efficient Federated Learning Framework
at Satellites and Ground Stations. arXiv preprint arXiv:2202.01267 (2022).
[50]
Behnaz Soltani, Venus Haghighi, Adnan Mahmood, Quan Z. Sheng, and Lina
Yao. 2022. A Survey on Participant Selection for Federated Learning in Mobile
Networks. In Workshop on Mobility in the Evolving Internet Architecture (MobiArch)
at MobiCom.
[51]
Abel Souza, Noman Bashir, Jorge Murillo, Walid Hanafy, Qianlin Liang, David
Irwin, and Prashant Shenoy. 2023. Ecovisor: A Virtual Energy System for Carbon-
Efficient Applications. In ASPLOS.
[52]
Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2020. Energy and Policy
Considerations for Modern Deep Learning Research. In AAAI.
[53]
Mingxing Tan and Quoc Le. 2019. EfficientNet: Rethinking Model Scaling for
Convolutional Neural Networks. In ICML.
[54]
Maud Texier. 2021. A timely new approach to certifying clean energy. Google.
Retrieved Sept. 2023 from https://cloud.google.com/blog/topics/sustainability/t-
eacs-offer-new-approach-to-certifying-clean-energy
[55]
Muhammad Tirmazi, Adam Barker, Nan Deng, Md E. Haque, Zhijing Gene Qin,
Steven Hand, Mor Harchol-Balter, and John Wilkes. 2020. Borg: The next Gener-
ation. In EuroSys.
[56]
Cong Wang, Bin Hu, and Hongyi Wu. 2022. Energy Minimization for Federated
Asynchronous Learning on Battery-Powered Mobile Devices via Application
Co-running. In ICDCS.
[57]
Qizhen Weng, Wencong Xiao, Yinghao Yu, Wei Wang, Cheng Wang, Jian He,
Yong Li, Liping Zhang, Wei Lin, and Yu Ding. 2022. MLaaS in the Wild: Workload
Analysis and Scheduling in Large-Scale Heterogeneous GPU Clusters. In USENIX
NSDI.
[58]
Philipp Wiesner, Ilja Behnke, and Odej Kao. 2023. A Testbed for Carbon-Aware
Applications and Systems. arXiv:2306.09774 [cs.DC]
[59]
Philipp Wiesner, Ilja Behnke, Dominik Scheinert, Kordian Gontarska, and Lauritz
Thamsen. 2021. Let’s Wait Awhile: How Temporal Workload Shifting Can Reduce
Carbon Emissions in the Cloud. In ACM Middleware.
[60]
Philipp Wiesner, Dominik Scheinert, Thorsten Wittkopp, Lauritz Thamsen, and
Odej Kao. 2022. Cucumber: Renewable-Aware Admission Control for Delay-
Tolerant Cloud and Edge Workloads. In International European Conference on
Parallel and Distributed Computing (Euro-Par).
[61]
Carole-Jean Wu, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani,
Kiwan Maeng, Gloria Chang, Fiona Aga Behram, Jinshi Huang, Charles Bai,
Michael Gschwind, Anurag Gupta, Myle Ott, Anastasia Melnikov, Salvatore
Candido, David Brooks, Geeta Chauhan, Benjamin Lee, Hsien-Hsin S. Lee, Bugra
Akyildiz, Maximilian Balandat, Joe Spisak, Ravi Jain, Mike Rabbat, and Kim M.
Hazelwood. 2022. Sustainable AI: Environmental Implications, Challenges and
Opportunities. In MLSys.
[62]
Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. 2019.
Federated learning. Morgan & Claypool Publishers.
[63]
Zhaohui Yang, Mingzhe Chen, Walid Saad, Choong Seon Hong, and Moham-
mad Shikh-Bahaei. 2021. Energy Efficient Federated Learning Over Wireless
Communication Networks. IEEE Transactions on Wireless Communications 20, 3
(2021).
[64]
Ashkan Yousefpour, Shen Guo, Ashish Shenoy, Sayan Ghosh, Pierre Stock, Kiwan
Maeng, Schalk-Willem Krüger, Michael Rabbat, Carole-Jean Wu, and Ilya Mironov.
2023. Green Federated Learning. arXiv:2303.14604 [cs.LG]
[65]
Sai Qian Zhang, Jieyu Lin, and Qi Zhang. 2022. A Multi-Agent Reinforcement
Learning Approach for Efficient Client Selection in Federated Learning. In AAAI.
https://doi.org/10.1609/aaai.v36i8.20894
[66]
Jiajia Zheng, Andrew A. Chien, and Sangwon Suh. 2020. Mitigating Curtailment
and Carbon Emissions through Load Migration between Data Centers. Joule 4,
384
FedZero: Leveraging Renewable Excess Energy in Federated Learning E-Energy ’24, June 04–07, 2024, Singapore, Singapore
10 (2020).
[67]
Zhi Zhou, Fangming Liu, Yong Xu, Ruolan Zou, Hong Xu, John C.S. Lui, and
Hai Jin. 2013. Carbon-Aware Load Balancing for Geo-distributed Cloud Services.
In 21st Int. Symposium on Modelling, Analysis and Simulation of Computer and
Telecommunication Systems (MASCOTS).
[68]
Chen Zhu, Zheng Xu, Mingqing Chen, Jakub Konečný, Andrew Hard, and Tom
Goldstein. 2022. Diurnal or Nocturnal? Federated Learning of Multi-branch
Networks from Periodically Shifting Distributions. In ICLR.
A TABLE WITH ALL RESULTS
Note, that the Upper Bound baseline is not constrained by capacity
or energy availability and therefore also uses grid energy.
Dataset & model Approach
Global scenario Co-located scenario
Target Best Time-to- Energy-to- Target Best Time-to- Energy-to-
accuracy accuracy accuracy accuracy accuracy accuracy accuracy accuracy
Upper bound
64.7 %
68.3 % 1.6 d 91.1 kWh
65.5 %
68.3 % 2.0 d 117.5 kWh
Random 64.7 % 6.7 d 80.6 kWh 65.5 % 6.7 d 101.0 kWh
Random 1.3n 66.0 % 4.7 d 79.2 kWh 66.6 % 5.3 d 113.4 kWh
CIFAR-100 Random fc 65.4 % 6.4 d 89.8 kWh 65.7 % 6.5 d 97.8 kWh
DenseNet-121 Oort 65.9 % 4.7 d 96.1 kWh 65.9 % 6.5 d 130.9 kWh
Oort 1.3n 66.4 % 4.5 d 103.8 kWh 66.4 % 5.4 d 138.7 kWh
Oort fc 65.8 % 5.3 d 102.4 kWh 66.1 % 6.4 d 126.7 kWh
FedZero 66.8 % 3.6 d 70.4 kWh 66.5 % 4.5 d 96.4 kWh
Upper bound
62.4 %
64.1 % 1.4 d 81.3 kWh
62.8 %
64.1 % 1.8 d 105.7 kWh
Random 62.4 % 6.7 d 99.8 kWh 62.8 % 5.7 d 92.0 kWh
Random 1.3n 63.1 % 5.6 d 109.6 kWh 63.3 % 3.7 d 86.0 kWh
Tiny ImageNet Random fc 63.0 % 5.6 d 96.4 kWh 63.1 % 6.5 d 102.1 kWh
EfficientNet-B1 Oort 63.3 % 3.8 d 88.1 kWh 62.7 % - -
Oort 1.3n 63.2 % 3.3 d 90.2 kWh 63.5 % 3.7 d 112.4 kWh
Oort fc 63.1 % 3.9 d 89.0 kWh 59.7 % - -
FedZero 63.7 % 2.8 d 64.8 kWh 63.8 % 3.4 d 76.6 kWh
Upper bound
50.4 %
53.3 % 1.4 d 82.5 kWh
50.9 %
53.3 % 1.8 d 104.2 kWh
Random 50.4 % 6.7 d 131.5 kWh 50.9 % 5.7 d 93.2 kWh
Random 1.3n 50.7 % 4.6 d 97.9 kWh 51.5 % 4.5 d 90.0 kWh
Shakespeare Random fc 52.0 % 2.8 d 57.3 kWh 52.1 % 3.7 d 59.2 kWh
LSTM Oort 50.2 % - - 50.7 % - -
Oort 1.3n 50.4 % - - 51.7 % 4.5 d 95.4 kWh
Oort fc 50.5 % 6.7 d 157.4 kWh 50.6 % - -
FedZero 53.1 % 1.8 d 40.0 kWh 53.1 % 2.3 d 42.8 kWh
Upper bound
83.6 %
87.9 % 2.7 d 105.6 kWh
82.8 %
87.9 % 2.3 d 91.0 kWh
Random 83.6 % 7.0 d 102.4 kWh 82.8 % 6.7 d 84.2 kWh
Random 1.3n 85.2 % 5.5 d 103.5 kWh 85.1 % 4.3 d 80.8 kWh
Google Speech Random fc 85.0 % 5.7 d 95.6 kWh 83.0 % 6.7 d 83.8 kWh
KWT-1 Oort 86.1 % 4.5 d 99.0 kWh 86.2 % 3.7 d 78.2 kWh
Oort 1.3n 86.8 % 3.9 d 103.5 kWh 86.4 % 3.4 d 85.0 kWh
Oort fc 87.0 % 3.8 d 86.2 kWh 85.6 % 3.7 d 78.1 kWh
FedZero 87.3 % 3.6 d 77.4 kWh 87.5 % 2.6 d 65.6 kWh
385
