EdgeFaaS A Function-based Framework for Edge Computing Runyu Jin Arizona state UniversityQirui Yang

2025-05-03 0 0 833.57KB 14 页 10玖币

侵权投诉

EdgeFaaS: A Function-based Framework for Edge Computing

Runyu Jin

Arizona state University

Qirui Yang

Arizona state University

Abstract—The rapid growth of data generated from Internet of

Things (IoTs) such as smart phones and smart home devices

presents new challenges to cloud computing in transferring,

storing, and processing the data. With increasingly more

powerful edge devices, edge computing, on the other hand,

has the potential to better responsiveness, privacy, and cost

efﬁciency. However, resources across the cloud and edge are

highly distributed and highly diverse. To address these chal-

lenges, this paper proposes EdgeFaaS, a Function-as-a-Service

(FaaS) based computing framework that supports the ﬂexible,

convenient, and optimized use of distributed and heterogeneous

resources across IoT, edge, and cloud systems. EdgeFaaS

allows cluster resources and individual devices to be managed

under the same framework and provide computational and

storage resources for functions. It provides virtual function and

virtual storage interfaces for consistent function management

and storage management across heterogeneous compute and

storage resources. It automatically optimizes the scheduling

of functions and placement of data according to their per-

formance and privacy requirements. EdgeFaaS is evaluated

based on two edge workﬂows: video analytics workﬂow and

federated learning workﬂow, both of which are representative

edge applications and involve large amounts of input data

generated from edge devices.

1. INTRODUCTION

We are living in the big data era where explosive data

comes largely from edge devices such as smart phones,

and smart home devices. The ubiquitous edge devices can

generate 400 Zettabytes of data per year [16]. This big

amount of data has presented important challenges to our ex-

isting computational infrastructure. The previous cloud-only

solution fails to transfer, store or process such vast amounts

of data efﬁciently and cost-effectively. What’s more, many

edge applications come with latency and privacy require-

ments where cloud-only solution falls short. At the same

time, edge devices have become unprecedented powerful in

both computation and storage, providing new opportunities

for the new computing paradigm, “edge computing”.

Contrary to the rapid growth of edge computing, there is

a lack of suitable computational infrastructure that supports

the unique characteristics of edge applications and edge

systems. Edge applications are highly data driven. Most edge

applications perform computations based on data collected

by IoT devices. For example, the smart home IoT devices

collect home metrics through different sensors which is

then analysed in the home hub or cloud. Another example

would be surveillance systems, where streams of videos are

generated from cameras and serves as input data for image

processing edge applications. The edge applications uses the

IoT devices generated data as input. The edge computational

infrastructure should transfer or store the large amount of

data generated from IoT devices with care to improve the

performance.

Edge applications are also latency sensitive. Due to edge

devices’ close interaction with the physical world by sensing

and actuating, many edge applications have low latency or

even real-time requirement. For example, in autonomous

driving, the system’s response time is capped at 100ms [29]

for some scenarios such as sensing and processing the

surrounding environment which is nearly real time. Another

example is for security alarms where the devices are ex-

pected to react immediately to detected security issues. The

edge computational infrastructure should process the vast

amount of data generated from the edge devices and avoid

the expensive data transfer latency from edge devices to

cloud [25].

Furthermore, edge applications have privacy require-

ment. Data collected from sensors and edge devices may

contain physical world or personal information that is crit-

ical to privacy. For example, medical edge devices capture

patient-private data and should not be exposed to any unre-

lated people. Computation that involves private data calls

for the data storage and processing to happen on local

edge devices to avoid possible data leakage during network

transportation and on the cloud.

At the same time, edge systems are highly heteroge-

neous and highly distributed. From edge to cloud, there

are three tiers of resources distributed over IoT devices,

edge servers, and cloud datacenters. Within each tier, the

resources are also distributed across different geographical

areas and administrative domains. An edge system is built

upon highly heterogeneous resources. IoT devices, edge

servers, and cloud servers are drastically different tiers in the

system, whereas within each tier the resources, especially

the devices in the IoT tier, are also highly diverse in terms

of processors/accelerators, memory/storage, OS/hypervisor,

and libraries/runtimes.

To better cater to the unique characteristics of edge

applications and edge systems, we introduce EdgeFaaS,

a function-based distributed computing platform for edge

applications.

First, EdgeFaaS allows cluster resources and individual

devices to be managed under the same framework and

provides computational and storage resources for functions.

The current FaaS platforms such as OpenFaas [14], Open-

Whisk [4], and AWS Lambda [12] can only run applications

on one device or one cluster. Applications cannot exploit

both the edge devices low data transfer cost and the cloud

strong computation power. EdgeFaaS can register different

clusters and schedule the computation based on the data

proximity and latency requirement to different clusters.

Second, it provides virtual function and virtual storage

interfaces for consistent function management and storage

management across heterogeneous compute and storage re-

sources. For the functions, the locations of input and output

data, cluster gateways, all these are virtualized in Edge-

FaaS and users do not have direct access. Each function

deployment, invocation are conducted through the exposed

EdgeFaaS gateway and all the resources are transparent to

the user. These help protect the data and function privacy.

Besides, it frees users from the hassle of function manage-

ment and storage management as users do not need to man-

ually participate in distributing functions across heteroge-

neous computing and storage resources. The manageability

of edge computing is improved as function and storage are

virtualized and managed by cautiously designed RESTful

APIs. Likewise, the storage resources are also virtualized

to be transparent to users. Users only get access to their

own data through the RESTful APIs and is free from the

complexity of storage resource management.

Third, it automatically optimizes the scheduling of func-

tions and placement of data according to their performance

and privacy requirements. To fullﬁll the latency-sensitive

feature of edge applications, EdgeFaaS place the functions

and data based on data locality to save the data transfer

latency from IoT devices to cloud. This feature also satisﬁes

the edge data privacy requirements by storing the device-

collected data locally without transferring it to the remote

cloud.

We also introduce two representative edge workﬂows to

illustrate the usage of EdgeFaaS. The two workﬂows are

video analytic workﬂows and federated learning workﬂows,

which are representative of a large portion of today’s edge

applications. The video analytic workﬂow forms a pipeline

of continuous video streams generation, motion detection,

face detection, face extraction, and face recognition. The

whole pipeline involves ﬁve stages and uses object storage

and message queue for storing the input and output of

each stage. The federated learning workﬂow involves the

distributed training on IoT devices and model aggregation on

edge and cloud clusters. Each IoT device trains the model lo-

cally using the local generated data. It then passes the trained

model to edge cluster for aggregation. The edge aggregated

model is ﬁnally contributed to cloud for ﬁnal aggregation.

Both workﬂows involve large amounts of data and exhibit

the three edge applications characteristics described above.

We implement and evaluate EdgeFaaS using the open

source FaaS platform, OpenFaaS [14]. We use the two edge

workﬂows to study the usability of EdgeFaaS. First, we

show EdgeFaaS can manage different computation and stor-

age resources. Second, we show how EdgeFaaS schedules

the function execution and data storage among different

resources to achieve the application requirements. Third, we

show how EdgeFaaS can be further extended to run other

edge applications.

In summary, the contributions of this work are as fol-

lows: 1) we characterize the unique challenges brought by

edge applications and the underlying systems; 2) we design

and implement EdgeFaaS to address these challenges by

providing a uniﬁed computing platform; 3) we provide two

workﬂows using EdgeFaaS which can represent a large

portion of edge applications; 4) we demonstrate that Edge-

FaaS can effectively execute edge workloads to meet edge

applications’ requirements.

In the following paper, Section 2 provides a background

of edge computing, function-as-a-service, and federated

learning. Section 3 provides an overview of the design and

implementation of EdgeFaaS. Section 4 introduces the two

representative edge workﬂows. Section 5 evaluates Edge-

FaaS based on the two workﬂows. Section 6 introduces

related works and Section 7 concludes the paper.

2. BACKGROUND

2.1. Edge Computing

Edge computing is a computing paradigm that conducts

computation close to the data source. With the proliferation

of edge devices, even the fast-growing cloud is believed to

be inadequate in handling the explosive demand of data col-

lection, transportation and processing. The emerging edge

computing, which expands computing infrastructure from

the cloud to the computing resources on edge including edge

devices, edge servers, and edge clusters [34], is a promising

paradigm to resolve the challenges on the cloud side.

A primary advantage of edge computing is the fast re-

sponse of applications. The ubiquitous edge devices can gen-

erate 400 Zettabytes of data per year [16]. The vast amount

of data transportation works have pushed the network to be

the bottleneck. Processing such large and fast data in time

places a huge burden on the computation resources of cloud

data centers. Both the computation and network latencies in

the cloud largely slow down applications’ response time,

which have pushed computation to move closer to the data

source. Doing computation on the ubiquitous edge devices

in a scale-out manner can reduce this unnecessary network

overhead by processing the data locally to improve the

response time.

Edge computing can ofﬂoad computation to peer edge

devices, nearby edge data centers, and farther cloud. All

these heterogeneous resources are utilized by edge comput-

ing to provide high scalability for different amount of work.

Another key advantage of edge computing is data pri-

vacy. Private data like personal video footage and data

collected by virtual assistant for model training can be better

protected on the edge then in the cloud. One one hand, edge

computing allows data to be processed on the user’s own

device which can be well protected by different security

features [2], [1]. On the other hand, it is relatively more

difﬁcult to ensure data privacy on shared cloud resources.

Finally, by utilizing the computing and storage resources

available on edge, the cost of cloud services can be reduced.

More computation are conducted using the edge resources

which is weaker and cheaper than cloud services. Users pay

less to use powerful cloud resources but in turn can get

similar services with acceptable performance.

2.2. Function-as-a-Service

Function-as-a-Service (FaaS) is a service-oriented com-

puting model that allows users to run functions without the

complexity of building and maintaining the infrastructure.

users deﬁne functions and FaaS deploys applications in

the unit of sandboxed functions. Functions can be chained

to work in a pipelined manner. A function is triggered by

function invocation. FaaS quickly and dynamically scales

up and down the number of function sandboxes on demand.

As soon as a request ﬁnishes, its function sandboxes can be

shut down to release resources and no state is guaranteed to

be preserved.

Compared to the other commonly used service models

such as IaaS and conventional PaaS, FaaS has several impor-

tant beneﬁts. First, FaaS simpliﬁes the development effort.

Developers are able to decompose complex applications

into smaller and simpler functions which helps reduce the

complexity of development and makes debugging easier.

The auto-scaling of functions gets easier too. Developers

only need to specify a few conﬁguration parameters to tune

the auto-scaling without the complexity of changing the

infrastructure and deploying more applications manually.

Another advantage of FaaS is the faster initialization.

FaaS deploys functions which usually have smaller image

sizes compared to entire applications. The functions belong-

ing to the same application can be deployed in parallel

to save the initialization time compared to a full-blown

application. FaaS also provides better cost effectiveness.

Edge devices are increasingly powerful in terms of speed,

capability, and capacity. For example, smartphones have

equipped with up to 8 core processors, tens of gigabytes of

memory and hundreds of gigabytes of storage to cater to the

various needs of applications. Finally, FaaS has the beneﬁts

of improving resource utilization. Resources are provided to

users in a ﬁner-grained unit which can avoid resource being

wasted and unutilized for IaaS and PaaS service models.

Also, each function requires limited resources and resources

can be better divided among different functions.

FaaS platforms typically involves a gateway and a FaaS

provider. The gateway provides RESTful APIs such as de-

ploy, invoke, delete operations on functions for users to send

requests to the FaaS platform. FaaS provider is a connector

to connect the FaaS gateway to the underlying computation

platform. Take OpenFaaS for example, it uses Kubernetes

as one of the underlying computation platforms and has a

faas-netes FaaS provider to connect the FaaS gateway to

the Kubernetes. Faas-netes parses the user requests from

the gateway and sends them to the Kubernetes for CRUD

for Functions, function auto-scaling, invocation of functions,

etc.

FaaS typically uses MinIO and S3 services as the ex-

ternal storage to store input and output data of a function.

These object stores provide interfaces for easy access to the

storage. Take MinIO as an example, it provides putObject,

getObject, makeBucket these interfaces for users to create

buckets and store and retrieve objects.

3. EdgeFaaS

EdgeFaaS manages different resources from IoT tier,

edge tier, to cloud tier. Each resource needs to be organized

as an OpenFaaS [14] resource which exposes a gateway

to EdgeFaaS in order to schedule functions to run on the

resource. EdgeFaaS provides a uniﬁed gateway which could

target different platforms using a scheduling mechanism of

user’s choice. It implements the same interfaces as Open-

FaaS but allows users to run applications using different

resources. So whenever an invocation is made or a deploy-

ment requested, EdgeFaaS is in the critical-path and acts like

a router, picking some most suitable resources for function

execution. In order to address the scalability and single point

of failure problem, EdgeFaaS can run as a deployment of

Kubernetes and utilizes the ReplicaSet provided by Kuber-

netes to maintain a stable set of replica pods running at any

given time. Figure 1 shows the architecture of EdgeFaaS. We

will explain each module in details in the following sections.

Besides OpenFaaS functions, the resource also needs to have

multiple storage backends and monitoring systems deployed.

By default, EdgeFaaS uses Minio [7] and Prometheus [10].

Minio is used to store input, output, and intermediate data

and Prometheus is to provide real-time resource usage stats

for EdgeFaaS function management.

Figure 1: EdgeFaaS Architecture

3.1. Resource Management

EdgeFaaS manages heterogeneous resources through re-

source management. Each heterogeneous resource from IoT

tier, edge tier, to cloud tier is organized as an OpenFaaS

resource. The OpenFaaS resource can be a group of data

center nodes forming a kubernetes resource which host

OpenFaaS services atop of it or can be a Faasd service,

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：

EdgeFaaS:AFunction-basedFrameworkforEdgeComputingRunyuJinArizonastateUniversityQiruiYangArizonastateUniversityAbstract—TherapidgrowthofdatageneratedfromInternetofThings(IoTs)suchassmartphonesandsmarthomedevicespresentsnewchallengestocloudcomputingintransferring,storing,andprocessingthedata.Withincre...

展开>> 收起<<

EdgeFaaS A Function-based Framework for Edge Computing Runyu Jin Arizona state UniversityQirui Yang.pdf

共14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！

EdgeFaaS A Function-based Framework for Edge Computing Runyu Jin Arizona state UniversityQirui Yang

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: