Section: New Results

Design methods for reconfiguration controller design in computing systems

We apply the results of the previous axes of the team's activity, as well as other control techniques, to a range of infrastructures of different natures, but sharing a transversal problem of reconfiguration control design. From this very diversity of validations and experiences, we draw a synthesis of the whole approach, towards a general view of Feedback Control as MAPE-K loop in Autonomic Computing [7], [9].

Self-adaptative distributed systems

Participants : Quang Pham Tran Anh, Eric Rutten, Hamza Sahli.

Complex Autonomic Computing Systems, as found typically in distributed systems, must involve multiple management loops, addressing different subproblems of the general management, and using different modeling, decision and control approaches (discrete [3], continuous, stochastic, machine-learning based, ...) They are generally addressing deployment and allocation of computations on resources w.r.t. QoS, load, faults, ... but following different, complementary approaches. The similarities and recurring patterns are considered as in Section 7.1.2 . Their execution needs to be distributed w.r.t. different characteristics such as latency (as in Fog and Edge Computing) or load. We are studying Software Architectures to address the design of such complex systems.

Self-adaptation of micro-services in Fog/Edge and Cloud computing

Fog systems are a recent trend of distributed computing having vastly ubiquitous architectures and distinct requirements making their design difficult and complex. Fog computing is based on leveraging both resource-scarce computing nodes around the Edge to perform latency and delay sensitive tasks and Cloud servers for the more intensive computation.

In this work, we present a formal model defining spatial and structural aspects of Fog-based systems using Bigraphical Reactive Systems, a fully graphical process algebraic formalism. The model is extended with reaction rules to represent the dynamic behavior of Fog systems in terms of self-adaptation. The notion of bigraph patterns is used in conjunction with boolean and temporal operators to encode spatio-temporal properties inherent to Fog systems and applications. The feasibility of the modelling approach is demonstrated via a motivating case study and various self-adaptation scenarios.

This work is done in cooperation with the Inria team Stack in Nantes, and published in the FOCLASA workshop, co-located with the SFEM conference [13].

Autonomic management in Software Defined Networks

In the framework of our cooperation with Nokia Bell-labs (See Section 8.1.2), and the Dyonisos team at Inria Rennes, we are considering the management of Software Defined Networks (SDN), involving Data-Centers and accelerators.

The main approach AI / Machine Learning approaches, developed in Rennes. An ongoing topic is to consider that these reinforcement learning based approaches involve questions of trust, and we are beginning to consider their composition with controllers based e.g. on Control Theory, in order to maintain guarantees on the behaviors of the managed system.

High-Performance Grid Computing

Cloud and HPC (High-Performance Computing) systems have increasingly become more varying in their behavior, in particular in aspects such as performance and power consumption, and the fact that they are becoming less predictable demands more runtime management [10].

A Control-Theory based approach to minimize cluster underuse

Participants : Abdul Hafeez Ali, Raphaël Bleuse, Bogdan Robu, Eric Rutten.

One such problem is found in the context of CiGri, a simple, lightweight, scalable and fault tolerant grid system which exploits the unused resources of a set of computing clusters. In this work, we consider autonomic administration in HPC systems for scientific workflows management through a control theoretical approach. We propose a model described by parameters related to the key aspects of the infrastructure thus achieving a deterministic dynamical representation that covers the diverse and time-varying behaviors of the real computing system. We propose a model-predictive control loop to achieve two different objectives: maximize cluster utilization by best-effort jobs and control the file server’s load in the presence of external disturbances. The accuracy of the prediction relies on a parameter estimation scheme based on the EKF (Extended Kalman Filter) to adjust the predictive-model to the real system, making the approach adaptive to parametric variations in the infrastructure. The closed loop strategy shows performance improvement and consequently a reduction in the total computation time. The problem is addressed in a general way, to allow the implementation on similar HPC platforms, as well as scalability to different infrastructures.

This work is done in cooperation with the Datamove team of Inria/LIG, and Gipsa-lab. Some results were published in the CCTA conference [14]. It was the topic of the Master's thesis of Abdul Hafeez Ali [16].

Combining Scheduling and Autonomic Computing for Parallel Computing Resource Management

Participants : Raphaël Bleuse, Eric Rutten.

This research topic aims at studying the relationships between scheduling and autonomic computing techniques to manage resources for parallel computing platforms. The performance of such platforms has greatly improved (149 petaflops as of November 2019 [20]) at the cost of a greater complexity: the platforms now contain several millions of computing units. While these computation units are diverse, one has to consider other constraints such as the amount of free memory, the available bandwidth, or the energetic envelope. The variety of resources to manage builds complexity up on its own. For example, the performance of the platforms depends on the sequencing of the operations, the structure (or lack thereof) of the processed data, or the combination of application running simultaneously.

Scheduling techniques offer great tools to study/guaranty performances of the platforms, but they often rely on complex modeling of the platforms. They furthermore face scaling difficulties to match the complexity of new platforms. Autonomic computing manages the platform during runtime (on-line) in order to respond to the variability. This approach is structured around the concept of feedback loops.

The scheduling community has studied techniques relying on autonomic notions, but it has failed to link the notions up. We are starting to address this topic.

High-Performance Embedded Computing

Participants : Soguy Mak Kare Gueye, Stéphane Mocanu, Eric Rutten.

This topics build upon our experience in reconfiguration control in DPR FPGA [2].

Implementing self-adaptive embedded systems, such as UAV drones, involves an offline provisioning of the several implementations of the embedded functionalities with different characteristics in resource usage and performance in order for the system to dynamically adapt itself under uncertainties. We propose an autonomic control architecture for self-adaptive and self-reconfigurable FPGA-based embedded systems. The control architecture is structured in three layers: a mission manager, a reconfiguration manager and a scheduling manager. This work is in the framework of the ANR project HPeC (see Section 9.2.1).

DPR FPGA and discrete control for reconfiguration

In this work we focus on the design of the reconfiguration manager. We propose a design approach using automata-based discrete control. It involves reactive programming that provides formal semantics, and discrete controller synthesis from declarative objectives.

Ongoing work concerns experimental validation, where upon the availability of hardware implementations of vision, detection and tracking tasks, a demonstrator is being built integrating our controller.

Mission management and stochastic control

In the Mission Management workpackage of the ANR project HPeC, a concurrent control methodology is constructed for the optimal mission planning of a U.A.V. in stochastic environnement. The control approach is based on parallel ressource sharing Partially Observable Markov Decision Processes modeling of the mission. The parallel POMDP are reduced to discrete Markov Decision Models using Bayesian Networks evidence for state identification. The control synthesis is an iterative two step procedure : first MDP are solved for the optimisation of a finite horizon cost problem ; then the possible ressource conflicts between parallel actions are solved either by a priority policy or by a QoS degradation of actions, e.g., like using a lower resolution version of the image processing task if the ressource availability is critical.

This work was performed in the framework of the PhD of Chabha Hireche, defended in nov. 2019 [17].

IoT and Cyberphysical Systems

Participants : Neil Ayeb, Ayan Hore, Fabien Lefevre, Stéphane Mocanu, Jan Pristas, Eric Rutten, Gaetan Sorin, Mohsen Zargarani.

Device management

The research topic is targeting an adaptative and decentralized management for the IoT. It will contribute design methods for processes in virtualized gateways in order to enhance IoT infrastructures. More precisely, it concerns Device Management (DM) in the case of large numbers of connected sensors and actuators, as can be found in Smart Home and Building, Smart Electricity grids, and industrial frameworks as in Industry 4.0.

Device Management is currently industrially deployed for LAN devices, phones and workstation management. Internet of Things (IoT) devices are massive, dynamic, heterogeneous, and inter-operable. Existing solutions are not suitable for IoT management. This work in an industrial environment addresses these limitations with a novel autonomic and distributed approach for the DM.

This work is in the framework of the Inria/Orange labs joint laboratory (see Section 8.1.1), and supported by the CIFRE PhD thesis grant of Neïl Ayeb, starting dec. 2017. It was awarded a best paper distinction at the Doctoral Symposium of ICAC 2019 [12].

Security in SCADA industrial systems

We focus mainly on vulnerability search, automatic attack vectors synthesis and intrusion detection [11]. Model checking techniques are used for vulnerability search and automatic attack vectors construction. Intrusion detection is mainly based on process-oriented detection with a technical approach from run-time monitoring. The LTL formalism is used to express safety properties which are mined on an attack-free dataset. The resulting monitors are used for fast intrusion detections. A demonstrator of attack/defense scenario in SCADA systems has been built on the existing G-ICS lab (hosted by ENSE3/Grenoble-INP). This work is in the framework of the ANR project Sacade on cybersecurity of industrial systems (see Section 9.2.2).

One of important results is the realization of a Hardware-in-the-loop SCADA Cyberange based on a electronic interface card that allows to interface real-world PLC with a software simulation [21]. The entire system is available in open-source including the electronic card fabrication files (http://gics-hil.gforge.inria.fr/). Interfacing system allow connection with various commercial simulation software but also with “home made” simulators [15]. The work is also supported by Grenoble Alpes Cybersecurity Institute (see Section 9.1.1) and Pulse program of IRT NANOELEC.

Ongoing work concerns the complementary topic of analysis and identification of reaction mechanisms for self-protection in cybersecurity, where, beyond classical defense mechanisms that detect intrusions and attacks or assess the kind of danger that is caused by them, we explore models and control techniques for the automated reaction to attacks, in order to use detection information to take the appropriate defense and repair actions. A first approach was developed in the M2R internship by Ayan Hore [18]