Section: New Results

Infrastructure-level support

Autonomic Cloud and Big-Data systems

This activity continues work started several years ago in the Sardes Inria-team, before it split into Erods (at LIG) and Ctrl-A (at Inria).

Coordination in multiple-loop autonomic Cloud systems

Participants : Soguy Gueye, Gwenaël Delaval, Stéphane Mocanu, Bogdan Robu, Eric Rutten.

Complex computing systems are increasingly self-adaptive, with an autonomic computing approach for their administration. Real systems require the co-existence of multiple autonomic management loops, each complex to design. However their uncoordinated co-existence leads to performance degradation and possibly to inconsistency. There is a need for methodological supports facilitating the coordination of multiple autonomic managers. We address this problem in the context of the ANR project Ctrl-Green (see 8.2.1 ), in cooperation with LIG (N. de Palma) in the framework of the PhD of S. Gueye. We propose a method focusing on the discrete control of the interactions of managers [7] [9] . We follow a component-based approach and explore modular discrete control, allowing to break down the combinatorial complexity inherent to the state-space exploration technique [13] . This improves scalability of the approach and allows constructing a hierarchical control. It also allows re-using complex managers in different contexts without modifying their control specifications. We build a component-based coordination of managers, with introspection, adaptivity and reconfiguration. We validate our method on a multiple-loop multi-tier system.

We are currently working on the distributed execution of modular controllers and on considering more control objectives, beyond purely discrete or logical ones, evaluating the new tool ReaX developed at Inria Rennes (Sumo) (see 6.2 ) and exploring continuous or stochastic control of servers provisioning.

Control for Big data

Participants : Bogdan Robu, Mihaly Berekmeri, Nicolas Marchand.

To deal with the issue of ensuring performance constraints while also minimizing costs in systems for Big Data analytics based on the parallel programming paradigm MapReduce, we propose a control theoretical approach, based on techniques that have already proved their usefulness for the control community. We develop an algorithm to create the first linear dynamic model for a Big Data MapReduce system, running a concurrent workload. Furthermore we identify two major performance constraint use cases: relaxed-minimal resource and strict performance constraints. For the first case we developed a feedback control mechanism and, to minimize the number of control actuations, an event-based feedback controller. For the second case we add a feedforward controller that efficiently suppresses the effects of large workload size variations. The work is validated in a simulated Matlab environment build at GIPSA-lab and online on a real 60 node MapReduce cluster (part of GRID 500), running a data intensive Business Intelligence workload. Our experiments demonstrate the success of the control strategies employed in assuring service time constraints [17] , [18] .

This work is performed in cooperation with LIG (S. Bouchenak) in the framework of the PhD of M. Berekmeri.

Reconfiguration control in DPR FPGA

Participant : Eric Rutten.

Dynamically reconfigurable hardware has been identified as a promising solution for the design of energy efficient embedded systems. However, its adoption is limited by the costly design effort including verification and validation, which is even more complex than for non dynamically reconfigurable systems. We work on this topic in the context of a ensign environment, developed in the framework of the ANR project Famous, in cooperation with LabSticc in Lorient and Inria Lille (DaRT team) [10] . We propose a tool-supported formal method to automatically design a correct-by-construction control of the reconfiguration. By representing system behaviors with automata, we exploit automated algorithms to synthesize controllers that safely enforce reconfiguration strategies formulated as properties to be satisfied by control. We design generic modeling patterns for a class of reconfigurable architectures, taking into account both hardware architecture and applications, as well as relevant control objectives. We validate our approach on two case studies implemented on FPGAs [1] .

We are currently valorizing results in more publications, and extending the use of control techniques by evaluating the new tool ReaX developed at Inria Rennes (Sumo) in the framework of the ANR Ctrl-Green project (see 6.2 and 8.2.1 ).

Autonomic memory management in HPC

Participants : Naweiluo Zhou, Gwenaël Delaval, Bogdan Robu, Eric Rutten.

Concurrent programs need to manage the time trade-off between synchronization and computing. A high concurrency level may decrease computing time but at the same time increase synchronization cost among threads. The traditional way to handle synchronization problems is through implementing locks. However locks suffer from the likelihood of deadlocks, vulnerability to failures, faults etc.. Software Transactional Memory (STM) has emerged as a promising technique to address synchronization issues through transactions. In STM, blocks of instructions accessing the shared data are wrapped into transactions. In STM each transaction executes speculatively, and conflicts may be aroused when two transactions are trying to modify the same area simultaneously. A way to reduce conflicts is by adjusting concurrency levels. A suitable concurrency level can maximize program performance. However, there is no universal rule to decide the best concurrency level for a program from an offline view. Hence, it becomes necessary to adopt a dynamical tuning strategy to better manage a STM system, so that a program can achieve a better performance. In the context of the action-team HPES of the Labex Persyval-lab(https://persyval-lab.org/en/sites/hpes ) (see 8.1 ), we explore the autonomic computing approach and control techniques to address these runtime tuning problems as a feedback control loop to automate the choices of concurrency levels, conflict management policies, and other parameters, with the objective of optimizing program execution time. This work is performed in cooperation with LIG (J.F. Méhaut) in the framework of the PhD of N. Zhou.

Control of smart environments

Participants : Julio Angel Cano Romero, Mengxuan Zhao, Eric Rutten, Hassane Alla [Gipsa-lab] .

Generic supervision architecture

New application domains of control, such as in the Internet of Things (IoT) and Smart Environments, require generic control rules enabling the systematization and the automation of the controller synthesis. We are working on an approach for the generation of Discrete Supervisory Controllers for these applications. A general modeling framework is proposed for the application domain of smart home. We formalize the design of the environment manager as a Discrete Controller Synthesis (DCS) problem, w.r.t. multiple constraints and objectives, for example logical issues of mutual exclusion, bounding of power peaks. We validate our models and manager computations with the BZR language and an experimental simulator [15] . This work is performed in cooperation with Orange labs (G. Privat) in the framework of the Cifre PhD of M. Zhao.

Rule-based specification

In the Internet of things, Event - Condition - Action (ECA) are used as a flexible tool to govern the relations between sensors and actuators. Runtime coordination and formal analysis becomes a necessity to avoid side effects mainly when applications are critical. In cooperation with CEA LETI/DACLE, we have worked on a case study for safe applications development in IoT and smart home environments [11] .