Research Challenges for Heterogeneous CPS Design

by   Shuvra S. Bhattacharyya, et al.

Heterogeneous computing is widely used at all levels of computing from data center to edge due to its power/performance characteristics. However, heterogeneity presents challenges. Interoperability—the management of workloads across heterogeneous resources—requires more careful design than is the case for homogeneous platforms. Cyber-physical systems present additional challenges. This article considers research challenges in heterogeneous CPS design, including interoperability, physical modeling, models of computation, self-awareness and adaptation, architecture, and scheduling.


page 1

page 2

page 3

page 4


Learning-Assisted Secure End-to-End Network Slicing for Cyber-Physical Systems

There is a pressing need to interconnect physical systems such as power ...

A new step for computing

The data center of tomorrow is a data center made up of heterogeneous sy...

Cyber Physical Systems: Prospects and Challenges

Cyber physical systems CPSs embodies the conception as well as the imple...

Merge: An Architecture for Interconnected Testbed Ecosystems

In the cybersecurity research community, there is no one-size-fits-all s...

Toward Edge-enabled Cyber-Physical Systems Testbeds

The use of edge computing can be extremely valuable in support of CPS ef...

TEACHING – Trustworthy autonomous cyber-physical applications through human-centred intelligence

This paper discusses the perspective of the H2020 TEACHING project on th...

Robustness Analysis for Battery Supported Cyber-Physical Systems

This paper establishes a novel analytical approach to quantify robustnes...


Heterogeneous computing offers the potential to streamline execution of key tasks for processing, sensing, actuation, and communication using devices that are better suited to those tasks than architectures composed from collections of identical devices. This potential is of great utility for cyber-physical systems (CPSs), where constraints on energy consumption, cost and real-time performance often motivate the investigation of highly streamlined solutions. However, increased use of heterogeneity leads to complex challenges and important needs associated with interoperability and model-based design in CPSs. This paper outlines challenges in heterogeneous CPS design, and motivates the need for approaches to system-level design that are based on complementary collections of compact system-level models.

Interoperability in Cps

Interoperability has been studied in many different forms in the context of heterogeneous computing and CPS. In this section, we review a small sampling of representative directions of investigation. A major direction of the recent emphasis in heterogeneous computing has focused on interoperability in the context of cloud computing (e.g., see [14]). Interoperability in this context involves both the management of application workloads across heterogeneous collections of resources associated with a given cloud computing service provider as well as the deployment of workloads across resources of different providers.

Givehchi et al. investigate interoperability challenges in industrial cyber-physical systems involving the networked management of data from heterogeneous field devices, such as I/O devices, sensors, and actuators [8]. They introduce an interoperability layer for connecting the physical and cyber layers in networked factory systems in such a way that legacy devices can be integrated without modification.

Gürdür et al. present a survey of methods for assessing interoperability in tool chains for CPS [9]. They identify numerous assessment models and focus on fourteen of the most popular models, which have been introduced over a period spanning 1980–2007. Their investigation found that most of the assessment models focus on isolated types of interoperability, and rely on complex metrics, which limits their usability in practical CPS design contexts.

In this paper, we discuss approaches for enhancing interoperability and heterogeneous CPS design based on the use of complementary modeling strategies, which abstract different concerns in the design process through well-defined, formal modeling concepts. We emphasize the diversity of different design concerns that may be modeled in this way, and the need for compactness in the models that are employed.

Compact System-Level Models

Raising the level of abstraction in design processes for CPS can facilitate interoperability by making it easier to reason about the behavior of subsystems in a design and interactions between them. However, due to the multi-faceted nature of CPS system design, no single abstraction or small set of abstractions is adequate for design of all systems. Instead, the abstractions to employ must be selected and applied in complementary ways that are well matched to the targeted class of applications, and the objectives and constraints that are involved in their design.

Given the complexity of modern CPS systems, the size of the models in the employed abstractions is an important consideration in their formulation or selection. The transition from assembly language to high-level languages such as FORTRAN or C, which began many decades ago, can be considered as an increase in the level of abstraction. However, modern CPS systems involve hundreds of thousands to tens of millions of lines of high-level language code or more. The compactness of the models that are involved in the abstractions becomes an important concern to facilitate human understanding and tractable analysis of the models.

Strategic application of compact models is important, for example, in the paradigm of dynamic, data driven applications systems (DDDAS), where an executing model of an application is integrated into a feedback loop with instrumentation processes that supply data to the model [4]. Accurate, compact models are useful for real-time adaptation of DDDAS models based on dynamic changes in the data acquired from instrumentation, and conversely, for control of the instrumentation processes by the executing models.

The motivations above for diverse and compact abstractions leads us to advocate the concept of compact system-level models as a central concept in the design and implementation of CPSs. Many different types of models are relevant to CPS design. Some prominent examples include the following.

  • Models of physical phenomena [15]. Computing is a physical act: it takes time and energy; the reliability of the result depends on the physics of the computing system. Taking all these physical phenomena into account in multi-billion-transistor systems is extremely challenging.

  • Models of computation. A model of computation defines how an interconnected set of components interact to perform computation. A few examples of important classes of models of computation include dataflow models, state machines, and discrete event models. Models of computation may impose restrictions on how components are defined or interact that make important analysis or optimization problems become tractable (e.g, see [7]). In contrast, fundamental analysis problems, such as whether a program halts or has bounded memory requirements, are undecidable in conventional programming languages for general-purpose computing. Models of computation contribute to modeling compactness by abstracting implementation details of individual functional components and their coordination.

  • Models of self-awareness and adaptation [6]

    . Stochastic models provide systems with compact, run-time-ready models that they can use to estimate their own state. Training allows us to capture complex models, so long as we have sufficient training data. Once trained, those models can be evaluated much more efficiently on the platform. Their results allow the system to reflect on its own power and thermal behavior. Managing power and thermal behavior is critical to maintaining system longevity.

  • Models of architecture. While models of computation focus on capturing the algorithmic behavior of application systems, models of architecture provide compact abstractions of the hardware on which the algorithms are mapped [13]. Models of architecture are formulated to enable efficient, reproducible estimation of nonfunctional costs associated with executing applications that are described in terms of a given model of computation. These costs include important metrics for efficiency evaluation, such as latency, throughput, memory requirements, and energy consumption. A key concept in the formulation of models of architecture is the decomposition of application execution into quantized units of communication and computation, and the estimation of costs in terms of these abstract units. Models of architecture are more constrained and operate at a higher level of abstraction compared to hardware description languages, such as Verilog and VHDL.

  • Scheduling models. Scheduling is an important aspect of implementation that is abstracted away by models of computation. Scheduling involves the assignment of computational tasks to processing resources and the ordering of tasks that share common resources. Scheduling often has major impact on metrics for efficiency evaluation, including the ones listed above. Model-based scheduling representations provide formal, platform-independent approaches for representing, reasoning about, and transforming schedules [3].

A design methodology based on compact system-level models for CPS involves the selection of such models, and the definition of how representations and design tools associated with these models are cooperatively applied in system design processes. While there are trade-offs between model complexity and accuracy that may be involved in the models that are employed, restricting attention to only the highest fidelity models may severely limit the extent of the design space that can be investigated.

Modeling Example

An example of a complex subsystem design using multiple forms of compact system-level models is the the MDP framework for Adaptive DPD (digital predistortion) Systems (MADS) [12]. DPD is a type of algorithm that is used to counteract nonlinearities in power amplifiers (PAs) to improve the quality of wireless communications signals (e.g., see  [1]). The design and configuration of DPD systems involves complex trade-offs among signal quality, energy efficiency and real-time performance. The MADS framework is demonstrated by mapping it into an optimized implementation on a CPU/GPU platform. The model-based design of the MADS framework is illustrated in Figure 1.

Fig. 1: An illustration of the MADS framework (adapted from [12]).

The MADS framework illustrates an approach to several of the challenges associated with heterogeneous CPS design discussed in this paper. MADS applies a model of the physics involved in a communications transmitter to define, simulate and fine-tune the core predistortion algorithm that is employed. A Markov decision process (MDP) is employed in MADS as a model that provides self-awareness and adaptation capabilities. MDPs are probabilistic models that are used to derive adaptation policies in uncertain environments. In particular, MDPs are used in the context of environments that are characterized using memoryless probability distributions — that is, the distribution of the next state is dependent only on the current state, and not on the trajectory of prior states that led to the current state. In MADS, MDP-based DPD architecture adaptation is performed with the objective of jointly optimizing signal quality, system throughput, and power consumption.

In general, MDP models can become large and unwieldy to employ in complex applications. To help ensure compactness of the MDP model that is employed, a hierarchical MDP [10] structure is designed, as illustrated in the lower left part of Figure 1.

Parameterized dataflow [2] is used in MADS as a model of computation to represent the algorithms employed for adaptation and DPD operation, and model their interactions. In parameterized dataflow, the design for a signal processing system is decomposed into three cooperating dataflow graphs, called the init graph, subinit graph, and body graph (see Figure 1). The body graph represents the core signal processing functionality, while the init and subinit graphs represent functionality for dynamic manipulation of parameters in the body graph. The init and subinit graphs differ in the frequency with which the associated parameter adaptation operations are carried out, with subinit graph operations being more frequent [2]. In MADS, the parameterized dataflow model is used as a starting point to map the MDP-equipped adaptive system into a CPU/GPU implementation.

For more details on the MADS framework, including the different design components illustrated in Figure 1, we refer the reader to the presentation by Li et. al [12].


Many of the state-of-the art methods for CPS design and implementation are not model-based or involve a focus on individual model types — for example, the development of software synthesis techniques for specific models of computation or reconfigurable architectures based on specific models for self-awareness and adaptivity. The study of design methodologies based on cooperating compact system-level modeling approaches is a broad area that is ripe for further study. For example, deeper understanding is needed for many modeling techniques on how these models may be adapted or parameterized to provide more flexible trade-offs between model compactness and accuracy. Some compact modeling adaptations, such as hierarchical and factored MDPs [10, 5] or the multirate versus homogeneous synchronous dataflow models of computation [11] (to name just a few), are established in the literature but are not applied in practice to their full potential. More diverse families of compact models, more sophisticated design tool support for applying and integrating them, and more concrete ways to assess the novel trade-offs introduced by such models are all representative directions for future research that can help to address the complexities and opportunities presented by heterogeneous CPS design.


We thank the following people who contributed to discussions about heterogeneous computing and interoperability that have influenced this article: Alvaro Cardenas, Roger Chamberlain, Tam Chantem, Changhee Jung, Miriam Leeser, Shivakant Mishra, Mahdi Nikdast, Massimiliano Pierobon, Aviral Shrivastava, Heechul Yun, and Ting Zhu. This work was supported in part by the U.S. National Science Foundation under Grants CNS1514425 and CNS151304, and by the U.S. Air Force Office of Scientific Research under Grant FA9550-18-1-0068.


  • [1] L. Anttila, P. Händel, and M. Valkama (2010) Joint mitigation of power amplifier and I/Q modulator impairments in broadband direct-conversion transmitters. IEEE Transactions on Microwave Theory and Techniques 58 (4), pp. 730–739. Cited by: MODELING EXAMPLE.
  • [2] B. Bhattacharya and S. S. Bhattacharyya (2001-10) PARAMETERIZED dataflow modeling for DSP systems. IEEE Transactions on Signal Processing 49 (10), pp. 2408–2421. Cited by: MODELING EXAMPLE.
  • [3] S. S. Bhattacharyya and J. Lilius (2018) Model-based representations for dataflow schedules. In Principles of Modeling: Essays Dedicated to Edward A. Lee on the Occasion of His 60th Birthday, pp. 88–105. Cited by: 5th item.
  • [4] E. P. Blasch, S. Ravela, and A. J. Aved (Eds.) (2018) Handbook of dynamic data driven applications systems. Springer. Cited by: COMPACT SYSTEM-LEVEL MODELS.
  • [5] C. Boutilier, R. Dearden, and M. Goldszmidt (1995) Exploiting structure in policy construction. In

    Proceedings of the International Joint Conference on Artificial Intelligence

    pp. 1104–1111. Cited by: OUTLOOK.
  • [6] N. Dutt, A. Jantsch, and S. Sarma (2016-02) Toward smart embedded systems: a self-aware system-on-chip (SoC) perspective. ACM Transactions on Embedded Computing Systems 15 (2). Note: Article 22, 27 pages Cited by: 3rd item.
  • [7] J. Eker, J. W. Janneck, E. A. Lee, J. Liu, X. Liu, J. Ludvig, S. Neuendorffer, S. Sachs, and Y. Xiong (2003-01) Taming heterogeneity — the Ptolemy approach. Proceedings of the IEEE 91 (1), pp. 127–144. Cited by: 2nd item.
  • [8] O. Givehchi, K. Landsdorf, P. Simoens, and A. W. Colombo (2017) Interoperability for industrial cyber-physical systems: an approach for legacy systems. IEEE Transactions on Industrial Informatics 13 (6), pp. 3370–3378. Cited by: INTEROPERABILITY IN CPS.
  • [9] D. Gürdür, F. Asplund, and J. El-khoury (2016) Measuring tool chain interoperability in cyber-physical systems. In Proceedings of the System of Systems Engineering Conference, Cited by: INTEROPERABILITY IN CPS.
  • [10] A. Jonsson and A. Barto (2006) Causal graph based decomposition of factored MDPs.

    Journal of Machine Learning Research

    7, pp. 2259–2301.
  • [11] E. A. Lee and D. G. Messerschmitt (1987-09) Synchronous dataflow. Proceedings of the IEEE 75 (9), pp. 1235–1245. Cited by: OUTLOOK.
  • [12] L. Li, P. Deaville, A. Sapio, L. Anttila, M. Valkama, M. Wolf, and S. S. Bhattacharyya (2019-12) MADS: a framework for design and implementation of adaptive digital predistortion systems. IEEE Journal on Emerging and Selected Topics in Circuits and Systems 9 (4), pp. 712–722. Cited by: Fig. 1, MODELING EXAMPLE, MODELING EXAMPLE.
  • [13] M. Pelcat et al. (2018-10) Reproducible evaluation of system efficiency with a model of architecture: from theory to practice. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 37 (10), pp. 2050–2063. Cited by: 4th item.
  • [14] R. Ranjan (2014) The cloud interoperability challenge. Computer 1 (2), pp. 20–24. Cited by: INTEROPERABILITY IN CPS.
  • [15] M. Wolf (2016) The physics of computing. Morgan Kaufmann Publishers Inc.. Cited by: 1st item.