On Provisioning Procedural Geometry Workloads on Edge Architectures

Ilir Murturi

1 a

, Chao Jia

2 b

, Bernhard Kerbl

2 c

, Michael Wimmer

2 d

, Schahram Dustdar

1 e

and Christos Tsigkanos

1 f

Distributed Systems Group, TU Wien, Vienna, Austria

Research Unit of Computer Graphics, TU Wien, Vienna, Austria

Keywords:

Edge Architectures, Computational Workloads, Edge-cloud Continuum.

Abstract:

Contemporary applications such as those within Augmented or Virtual Reality (AR/VR) pose challenges for

software architectures supporting them, which have to adhere to stringent latency, data transmission, and

performance requirements. This manifests in processing 3D models, whose 3D contents are increasingly gen-

erated procedurally rather than explicitly, resulting in computational workloads (i.e., perceived as Procedural

Geometry Workloads) with particular characteristics and resource requirements. Traditionally, executing such

workloads takes place in resource-rich environments such as the cloud. However, the massive amount of data

transfer, heterogeneous devices, and networks involved affect latency, which in turn causes low-quality visual-

ization in user-facing applications (e.g., AR/VR). To overcome such challenges, processing elements available

close to end users can be leveraged to generate 3D models instead, and as such the edge emerges as a central

architectural entity. This paper describes such procedural geometry workloads, their particular characteristics,

and challenges to execute them on heterogeneous devices. Furthermore, we propose an architecture capable

of provisioning procedural geometry workloads in edge scenarios.

1 INTRODUCTION

Contemporary applications such as those within Aug-

mented or Virtual Reality (AR/VR) demand dedicated

software architectures, able to cope with stringent la-

tency, data transmission, and performance require-

ments. Those challenge the traditional cloud-centric

view, where computation and data reside in power-

ful cloud servers, but away from user-facing appli-

cations which may have to further overlay sensory

information obtained near user’s locations. The key

computational functionality within AR/VR consists

of processing high-quality 3D models, whose 3D con-

tents are increasingly generated procedurally rather

than explicitly, resulting in computational workloads

with particular characteristics and resource require-

ments. Visualizing and representing 3D models is

a resource-intensive process that involves both mas-

sive data transfer to user clients (in terms of both

volume and velocity), as well as demanding compu-

tation, with latency perceived by end users being a

https://orcid.org/0000-0003-0240-3834

https://orcid.org/0000-0003-2304-5976

https://orcid.org/0000-0002-5168-8648

https://orcid.org/0000-0002-9370-2663

https://orcid.org/0000-0001-6872-8821

https://orcid.org/0000-0002-9493-3404

major concern. In recent years, one prominent ap-

proach that has emerged to overcome latency delays

especially in pervasive applications suggests utilizing

computation entities in proximity to end users. Edge

Computing is a distributed computing paradigm that

places resources (e.g., compute and storage) at the

edge of the network (Shi and Dustdar, 2016). Edge

resources (i.e., perceived as edge devices) are typi-

cally resource-constrained and heterogeneous devices

that can process, store, and analyze data and deliver

efﬁcient and low-latency user-facing services. Such

edge devices can support applications by i) running

decision functions close to data-producing end users,

ii) processing computational workloads, and iii) min-

imizing the need to transmit data to the cloud.

Accurate digital architectural models are essential

to various practical applications such as urban plan-

ning (Vanegas et al., 2009), 3D navigation (Hilde-

brandt and Timm, 2014), natural and social phe-

nomena simulation (Heuveline et al., 2011; Jund

et al., 2012) in urban environments. Employing such

3D representations in contemporary near real-time

applications requires their abstraction or simpliﬁca-

tion (Visconti et al., 2021) to avoid storage and trans-

mission of the sheer volume of geometric informa-

tion that they include. One way to improve the user-

perceived quality of these visual models without mas-

354

Murturi, I., Jia, C., Kerbl, B., Wimmer, M., Dustdar, S. and Tsigkanos, C.

On Provisioning Procedural Geometry Workloads on Edge Architectures.

DOI: 10.5220/0010687800003058

In Proceedings of the 17th International Conference on Web Information Systems and Technologies (WEBIST 2021), pages 354-359

ISBN: 978-989-758-536-4; ISSN: 2184-3252

sive storage overhead is to procedurally generate the

geometric data on the ﬂy. Procedural modeling has

been well studied in the research area of computer

graphics and successfully applied in video games, the

movie industry, and AR/VR applications where com-

pelling and immersive virtual urban environments are

vital to the visual experience (Kim et al., 2018; M

uller

et al., 2006). By recursively applying a set of rules

that represent spatial transformations on geometric

shapes, procedural modeling can generate highly de-

tailed geometry starting from a very basic primitive

shape called an axiom (e.g., a box or a quadrilateral).

For example, windows or doors on a wall can be cre-

ated by subdividing a rectangle and applying extru-

sion operations and different textures to the resulting

small rectangles. Analogously, details on windows or

doors can be generated with further subdivision, ex-

trusion, and texture mapping rules.

The expressiveness of procedural geometry allows

for a compact representation of complex 3D phys-

ical environments with minimal resources. More-

over, procedural geometry is highly ﬂexible in that

the generation of geometric data can be tailored to

different requirements by controlling the set of rules

to be evaluated. Recently, efforts have also been

made to harness advances in parallel computing to

signiﬁcantly speed up procedural geometry genera-

tion (Steinberger et al., 2014). These particular char-

acteristics enable the distribution of procedural geom-

etry workloads on a wide range of devices to raise

performance. Aside from data center-grade servers,

edge devices such as single-board computers (SBC)

equipped with low-cost processors and embedded

systems specialized in parallel processing can also be

leveraged for load balancing and latency reduction.

Edge-based infrastructures are heterogeneous and

dynamic environments consisting of various devices

featuring different capabilities; resources available

may differ in terms of computational capabilities (i.e.,

from low-powered devices to server-grade hosts).

Such heterogeneity poses several challenges from a

software architecture perspective when executing pro-

cedural geometry workloads since they have par-

ticular processing requirements and low-latency de-

mands. The key theme is that processing elements

close to end users can be leveraged to generate 3D

models, to avoid user-perceived latency. In this paper,

we identify the current trend towards distributed com-

puting for visualization tasks, describe such procedu-

ral geometry workloads, their particular characteris-

tics, and challenges to execute them. Furthermore, we

propose an architecture capable of provisioning pro-

cedural geometry workloads in edge settings. Finally,

we outline a research agenda.

2 PERVASIVE VISUALIZATION

With the arrival of ubiquitous network connectiv-

ity and affordable, consumer-grade smartphones and

tablets, 2D and 3D visualizations that depend on live

remote data (e.g., video streaming, games and sim-

ulations) have become available on mobile devices.

In many cases, the raw data that must be transferred

in real-time is either limited or can be sufﬁciently

compressed, and end-user devices are capable of de-

compressing and performing the necessary visualiza-

tion tasks themselves. However, the concept of the

thin client is progressively becoming more important

for high-end pervasive visualization: complex visual

applications (e.g., Triple-A games) increasingly rely

on cutting-edge or specialized hardware capabilities,

opening the door for cloud gaming services that re-

lieve end-user devices of these requirements. The es-

tablished capabilities for streaming video at high res-

olutions can be exploited to deliver visual content that

is entirely generated by dedicated cloud services, such

as NVIDIA’s GeForce Now, Microsoft’s XCloud or

Google’s Stadia. However, as has become evident

during the COVID-19 pandemic, constant streaming

of high-resolution image data takes a heavy toll on

the available internet infrastructure. With the immi-

nent availability of yet faster 5G connectivity (high

bandwidth, low latency, low range) in many areas,

the exploitation of edge nodes is a logical next step

to both improve user experience and relieve some of

the infrastructure stress. However, edge computing

is still a relatively young concept and in the process

of being developed. In contrast to cloud services,

performing visualization tasks with the help of edge

nodes enables a range of scheduling and distribution

strategies and encourages the development of special-

ized solutions for different use cases. Reasonable ap-

proaches that economize on the available resources

require careful analysis of the implied workloads and

the design of multi-tiered, dynamic architectures.

3 PROCEDURAL GEOMETRY AS

A COMPUTATIONAL

WORKLOAD

Online repositories such as Google Earth

or 3DC-

ityDB

provide 3D models for several metropolitan

cities. Those are particularly useful e.g., in navigation

applications that aim to assist end-users while roam-

ing inside a city. Unfortunately, the 3D representa-

Google Earth, https://www.google.com/earth

The CityGML Database, https://www.3dcitydb.org/

On Provisioning Procedural Geometry Workloads on Edge Architectures

355

tions of those architectural models used in everyday

applications are often oversimpliﬁed and lack in terms

of quality when requested in high-quality 3D due to

the high latency caused by network congestion. As

a result, applying the aforementioned methodologies

to applications like AR/VR may result in unexpected

user-perceived latencies. Thus, novel applications in

AR/VR require novel paradigms and software archi-

tectures to cope with demands on latency, resource

management, and computation requirements.

Cloud

275

Edge devices

Fog Node

Base station

275

Request

rendering

Result/Data

Users

Data

Exchange

Sensory data

Cloud Layer

Edge and Fog Layers

Subject Layer

Workload

Processing

Figure 1: An example of an AR application reliant on high-

quality 3D data for use in civil engineering.

AR/VR applications are increasingly being pur-

sued in a plethora of ﬁelds ranging from gaming and

immersive tourism to industrial applications. Con-

sider for instance an AR application within civil en-

gineering (Figure 1), which, using CityGML models,

enables visualization of buildings by rendering them

into users’ smart devices such as smartphones, tablets,

or 3D glasses. Use cases may involve users such

as city inspectors roaming in the analyzing energy

infrastructure (i.e., heating or cooling) (Kaden and

Kolbe, 2014). In such a case, 3D visualization com-

bined with real-time sensor data (e.g., energy con-

sumption) enables users to perceive energy demands

for each building on their smart devices.

However, to enable the interaction with high-

quality 3D visualization on the users’ smart devices,

one must update the geometry of the observed build-

ing(s) as the user moves. Procedural geometry gener-

ation is an intensive process, while representing high-

quality 3D visualizations on the user’s smart device

at interactive rates demands fast processing and low

latency. To meet these demands, a user’s smart de-

vice should take advantage of available computing re-

sources in proximity (i.e., server-grade fog devices)

to ﬁnd the most suitable place to execute geometry

workloads (1); this may take place in single-board

computers, in server-grade fog nodes in mobile base

stations, or in the cloud. After a computation request

is made, various data is exchanged between compu-

tation entities and system components (2-3) to decide

where to compute the workload as well as to provide

the resulting geometry to the user’s device (at the sub-

ject layer).

The ever-growing demand for high ﬁdelity graph-

ics and visualization applications entails processing

enormous geometry data consisting of numerous tri-

angles; for instance, around 1 million triangles are re-

quired to model an area of 0.05 km

with a moderate

level of detail (LOD) near the center of the city of Vi-

enna, resulting in at least 30 MiB of raw mesh data. In

contrast, an analogous procedural generation of raw

mesh data requires less than 0.1 MiB in input param-

eters that deﬁne the position, dimension and orien-

tation of each axiom shape, as well as style param-

eters upon which detailed geometries can be gener-

ated. These style input parameters can for instance be

extracted from high-quality imagery (aerial or drive-

through) and applied on top of CityGML models in

which buildings are represented by simple boxes that

serve as a geometric baseline for the procedural gen-

eration (Figure 2).

Figure 2: Illustration of procedural generation workload.

Starting from a CityGML baseline model (left), detailed ge-

ometry is generated and delivered to the end user (right).

Given the amount of required raw mesh data, pro-

cedural generation workloads typically involve very

intensive computation. This manifests as a trade-off

for the full 3D data that would be equivalently trans-

ferred. The automatic generation of 3D data is con-

ﬁgurable; one may compute a certain level of detail

only, do so dynamically and depending on computa-

tion or time budgets. The (conﬁgurable) level of de-

tail emerges naturally as another factor subject to opti-

mization. Moreover, since geometric details for areas

far away from the current point of view of the user are

insigniﬁcant, their computation can be omitted, and

rule evaluation can be terminated at a lower LOD for

those buildings, thereby greatly eliminating unneces-

sary computation. Finally, the generated mesh data

for areas that are frequently requested by end users

can be cached and reused.

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

356

Runtime

Orchestration

PG Configuration

Management

Resource

Management

Workload

Processing

Smart City

3D Models

Procedural Geometry

Workload

GPU

CPU

Edge Layer

Latency (milliseconds)

< 30ms

< 5ms

< 10 ms

Legend:

4G / 5G network

connection

WAN

connection

Optional network

connection

Million of devices

Hundreds of devices

Adaptation

Management

Workload

Processing

PG Configuration

Manager

Monitoring

Workload

Processing

Edge Node

CPU Battery

Node

GPUCPU

Node

Request

Edge Layer

Edge

Application

Cloud Layer

Fog Layer

Figure 3: An overview of the proposed architecture and the interactions between layers.

4 SOFTWARE ARCHITECTURE

CONSIDERATIONS

Edge Computing is positioned as an important archi-

tectural layer between cloud and end users (Figure 1).

A platform for processing 3D models may leverage

the decentralized nature of such infrastructures, con-

sisting largely of three entities: ﬁrst, the user that re-

quests 3D processing; second, the software compo-

nents that support ﬁnding the most suitable entity to

execute the procedural geometry workload and cope

with the dynamicity and uncertainty of the edge en-

vironment; and third, an orchestration layer, typically

located in the cloud and responsible for monitoring

the overall system and deployment of software com-

ponents in the Edge-Cloud infrastructure.

The cloud is a resource-rich environment and

provides advanced features for both service providers

and service consumers; thus, we advocate its role for

orchestration, performing 3D model generation, and

providing the required resource management. The

edge layer comprises devices placed in proximity

to end users with different characteristics and hard-

ware conﬁgurations, divided into sub-layers of i) fog

(i.e., powerful devices) and ii) edge (low-powered

devices).

The fog layer includes a set of stationary powerful

devices (i.e., physical or virtual) with different hard-

ware conﬁgurations (i.e., CPU, GPU, storage, etc.).

The role of the fog layer is interchangeable from exe-

cuting geometry workloads with low latency to an in-

termediary layer for managing, communicating, and

exchanging resources between different edge devices

and the cloud. In addition, fog devices may provide

storage for 3D models with different sizes (i.e., from

MB to GB). The edge layer consists of numerous de-

vices that can simultaneously request visualization;

those are usually resource-constrained devices (i.e.,

CPU-based and battery-powered) — however, often

with enough hardware capabilities to execute speciﬁc

geometry workloads (such as modern smartphones).

Figure 3 illustrates different architectural mani-

festations to execute procedural geometry workloads.

Edge applications operated by end users and hosted

on smart devices interact with other system compo-

nents deployed in the Edge-Cloud continuum. An

edge application internals view consists of three com-

ponents: i) procedural geometry conﬁguration man-

ager, ii) monitoring, and iii) workload processing.

On Provisioning Procedural Geometry Workloads on Edge Architectures

357

The conﬁguration manager enables users to express

their goals (e.g., visualizing based on the user’s lo-

cation and within a speciﬁed radius). Essentially, a

user conﬁgures spatial boundaries (i.e., a building or

a neighborhood), and the amount of detail desired to

represent the virtual world. The monitoring compo-

nent is responsible for observing internal hardware

resource states: i) internal hardware resources and

their utilization and ii) quality of the communication

link (i.e., status, latency, and bandwidth). The work-

load processing component is responsible for execut-

ing procedural geometry workloads (i.e., workloads

can be packaged into software containers within the

overall service-based architecture (Dustdar and Mur-

turi, 2021)). As illustrated in Figure 3, the process

starts (1) when a user expresses her goal via the edge

application. Then, the request with hardware infor-

mation is forwarded (2-3) to the adaptation manager

which interprets the goal and decides where to exe-

cute the workload.

The adaptation management component is respon-

sible for identifying devices needed to achieve a user

goal with the lowest possible latency. If the goal is

achievable on the user’s device, it forwards the re-

quired data to the host. If the goal is not achievable

locally, the adaptation component attempts to gener-

ate a deployment plan that maps the workload to other

available devices. As illustrated in Figure 3, procedu-

ral generation may occur in the cloud, fog, and edge

devices. To generate valid deployment plans, adapta-

tion must consider several factors such as device hard-

ware requirements, network metrics, and the time re-

quired to transfer (un)processed geometry. To gener-

ate optimal plans, the adaptation component requires

ﬁne-grained information of the infrastructure (4).

The cloud part has a supportive role, which in-

cludes procedural geometry conﬁguration manage-

ment, data storage (e.g., 3D models), resource man-

agement, geometry workload generation, and overall

orchestration. As illustrated in Figure 3, the user’s

request can be forwarded (5) directly to the cloud

as well if no other solution is feasible. The re-

source management component comprises a set of

functionalities from resource discovery (i.e., discov-

ering available edge devices) to context monitoring

(i.e., monitoring hardware infrastructure and updating

its status when changes occur). Orchestration entails

where the software components must be placed, aim-

ing for reliable and low-latency service to end users.

Recent developments in IoT-based systems have

shown that systems can be engineered, deployed, and

executed in Edge-Cloud infrastructures (Alkhabbas

et al., 2020). At the same time, software components

can easily self-adapt to dynamic changes in their de-

ployment topologies when the quality of their services

is degraded (Brogi et al., 2020). Finally, as shown

in Figure 3, software components can be placed on

different devices yielding different deployment con-

ﬁgurations. More speciﬁcally, software components

that face high requests from a particular region can

be placed in proximity to the end-users. For in-

stance, if the procedural geometry generation for a

particular city area occurs mostly on the user devices,

then the orchestration mechanism must instantiate the

data storage component with associated data (i.e., 3D

models) on the nearest fog devices to the users. As a

result, data can be forwarded faster to the end-users

from the edge layer rather than from the cloud via

WAN connection.

5 AN EMERGING RESEARCH

AGENDA

Satisfying the dynamic and stringent requirements of

contemporary applications such as those in AR/VR is

challenging for centralized cloud-based systems. Pro-

cessing 3D models and transferring vast amounts of

data to user-facing devices over the internet incur la-

tencies and result in user experience degradation. We

discussed aspects emerging from latency and compu-

tation requirements and how edge architectures can

address the requirements and support procedural ge-

ometry workloads. Thus, we sketched an architecture

capable of provisioning such workloads in edge com-

puting scenarios.

As future work, we aim at providing a complete

technical framework for the processing of geome-

try workloads on edge-based architectures; this in-

cludes both technical and architectural aspects. En-

capsulating procedural generation appropriately such

as it being able to execute on heterogeneous hardware

platforms is a challenging task, as such workloads

are required to take advantage of specialized hard-

ware (such as GPUs) when available, yielding dif-

ferent conﬁgurations. Subsequently, our vision en-

tails them to be containerized, such that a service-

based architecture emerges across the device-to-cloud

continuum. Performance aspects of different geome-

try workloads executed on state-of-the-art resource-

constrained and powerful devices need to be care-

fully considered. Besides that, assessing deployment

tradeoffs in terms of quality, performance, and cost

is highly desired. Regarding deployment, the edge

topology may not be static, and components may need

to be scaled or migrated to comply with other con-

straints like energy, latency, or device movement, in-

troducing dynamicity. Finally, we identify three main

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

358

research challenges that must be further investigated

in the future:

• Procedural Geometry Workload Conﬁgura-

tion. A platform for the described scenarios needs

to hide operational complexity from application

end users and developers. In particular, devel-

opers should be able to express in a high-level

way the context in which particular procedural ge-

ometry executions are allowed to run. Thus, a

novel domain-speciﬁc language (DSL) for spec-

ifying the high-level constraints such as Quality

of Service (QoS) and hardware requirements re-

mains a critical task.

• Resource Discovery across the Edge-cloud

Continuum. A fundamental aspect in described

scenarios is discovering resources such as edge

devices or IoT resources (e.g., sensors, actuators,

etc.) as they become available in the city’s envi-

ronment (Murturi and Dustdar, 2021). Thus, re-

source discovery in heterogeneous and dynamic

edge-based settings remain among the main re-

search challenges.

• Resilient Geometry Workload Runtime. The

end users and edge devices providing computa-

tional resources can be mostly mobile. As a result,

preserving optimal QoS in the face of client or

resource mobility is another prominent research

challenge that needs to be addressed in the future.

ACKNOWLEDGMENT

Research supported in part by the Research Cluster

“Smart Communities and Technologies (Smart CT)”

at TU Wien, the EU’s Horizon 2020 Research and

Innovation Programme under grant agreement No.

871525 and by Austrian Science Foundation’s (FWF)

project M 2778-N “EDENSPACE”.

REFERENCES

Alkhabbas, F., Murturi, I., Spalazzese, R., Davidsson, P.,

and Dustdar, S. (2020). A goal-driven approach for

deploying self-adaptive iot systems. In 2020 IEEE

International Conference on Software Architecture

(ICSA), pages 146–156. IEEE.

Brogi, A., Forti, S., Guerrero, C., and Lera, I. (2020). How

to place your apps in the fog: State of the art and

open challenges. Software: Practice and Experience,

50(5):719–740.

Dustdar, S. and Murturi, I. (2021). Towards IoT Processes

on the Edge, pages 167–178. Springer International

Publishing, Cham.

Heuveline, V., Ritterbusch, S., and Ronnas, S. (2011).

Augmented reality for urban simulation visualization.

Preprint Series of the Engineering Mathematics and

Computing Lab, (16).

Hildebrandt, D. and Timm, R. (2014). An assisting, con-

strained 3d navigation technique for multiscale virtual

3d city models. GeoInformatica, 18(3):537–567.

Jund, T., Kraemer, P., and Cazier, D. (2012). A uniﬁed

structure for crowd simulation. Comput. Animat. Vir-

tual Worlds, 23(3-4):311–320.

Kaden, R. and Kolbe, T. H. (2014). Simulation-based total

energy demand estimation of buildings using semantic

3d city models. International Journal of 3-D Informa-

tion Modeling (IJ3DIM), 3(2):35–53.

Kim, J., Kavak, H., and Crooks, A. (2018). Procedural city

generation beyond game development. ACM SIGSPA-

TIAL Special, 10(2):34–41.

uller, P., Wonka, P., Haegler, S., Ulmer, A., and Gool,

L. V. (2006). Procedural modeling of buildings. ACM

Trans. Graph., 25(3):614–623.

Murturi, I. and Dustdar, S. (2021). A decentralized ap-

proach for resource discovery using metadata replica-

tion in edge networks. IEEE Transactions on Services

Computing.

Shi, W. and Dustdar, S. (2016). The promise of edge com-

puting. Computer, 49(5):78–81.

Steinberger, M., Kenzel, M., Kainz, B., M

uller, J., Wonka,

P., and Schmalstieg, D. (2014). Parallel generation

of architecture on the GPU. Comput. Graph. Forum,

33(2):73–82.

Vanegas, C. A., Aliaga, D. G., Benes, B., and Waddell,

P. (2009). Interactive design of urban spaces using

geometrical and behavioral modeling. ACM Trans.

Graph., 28(5):111.

Visconti, E., Tsigkanos, C., Hu, Z., and Ghezzi, C. (2021).

Model-driven engineering city spaces via bidirec-

tional model transformations. Software and Systems

Modeling, pages 1–20.

On Provisioning Procedural Geometry Workloads on Edge Architectures

359