Section: New Results

Language Based Fault-Tolerance

Participants : Pascal Fradet, Alain Girault, Gregor Goessler, Jean-Bernard Stefani, Martin Vassor.

Fault Ascription in Concurrent Systems

The failure of one component may entail a cascade of failures in other components; several components may also fail independently. In such cases, elucidating the exact scenario that led to the failure is a complex and tedious task that requires significant expertise.

The notion of causality (did an event $e$ cause an event $e^{\text{'}}$?) has been studied in many disciplines, including philosophy, logic, statistics, and law. The definitions of causality studied in these disciplines usually amount to variants of the counterfactual test “$e$ is a cause of $e^{\text{'}}$ if both $e$ and $e^{\text{'}}$ have occurred, and in a world that is as close as possible to the actual world but where $e$ does not occur, $e^{\text{'}}$ does not occur either”. In computer science, almost all definitions of logical causality — including the landmark definition of [54] and its derivatives — rely on a causal model that. However, this model may not be known, for instance in presence of black-box components. For such systems, we have been developing a framework for blaming that helps us establish the causal relationship between component failures and system failures, given an observed system execution trace. The analysis is based on a formalization of counterfactual reasoning [6].

In [16] we have discussed several shortcomings of existing approaches to counterfactual causality from the computer science perspective, and sketched lines of work to try and overcome these issues. In particular, research on counterfactual causality analysis has been marked, since its early days, by a succession of definitions of causality that are informally (in)validated against human intuition on mostly simple examples, see e.g.,   [54], [53]. We call this approach TEGAR, textbook example guided analysis refinement. As pointed out in  [48], it suffers from its dependence on the tiny number and incompleteness of examples in the literature, and from the lack of stability of the intuitive judgments against which the definitions are validated. We have argued that we need a formalization of counterfactual causality based on first principles, in the sense that causality definitions should not be driven by individual examples but constructed from a set of precisely specified requirements. Example of such requirements are robustness of causation under equivalence of models, and well-defined behavior under abstraction and refinement. To the best of our knowledge, none of the existing causality analysis techniques provides sufficient guarantees in this regard.

We are currently working on a revised version of our general semantic framework for fault ascription in  [50] that satisfies a set of formally stated requirements, and on its instantiation to acyclic models of computation, in order to compare our approach with the standard definition of actual causality proposed by Halpern and Pearl.

Tradeoff exploration between energy consumption and execution time

We have continued our work on multi-criteria scheduling, in two directions. First, in the context of dynamic applications that are launched and terminated on an embedded homogeneous multi-core chip, under execution time and energy consumption constraints, we have proposed a two layer adaptive scheduling method [14]. In the first layer, each application (represented as a DAG of tasks) is scheduled statically on subsets of cores: 2 cores, 3 cores, 4 cores, and so on. For each size of these sets (2, 3, 4, ...), there may be only one topology or several topologies. For instance, for 2 or 3 cores there is only one topology (a “line”), while for 4 cores there are three distinct topologies (“line”, “square”, and “T shape”). Moreover, for each topology, we generate statically several schedules, each one subject to a different total energy consumption constraint, and consequently with a different Worst-Case Reaction Time (WCRT). Coping with the energy consumption constraints is achieved thanks to Dynamic Frequency and Voltage Scaling (DVFS). In the second layer, we use these pre-generated static schedules to reconfigure dynamically the applications running on the multi-core each time a new application is launched or an existing one is stopped. The goal of the second layer is to perform a dynamic global optimization of the configuration, such that each running application meets a pre-defined quality-of-service constraint (translated into an upper bound on its WCRT) and such that the total energy consumption be minimized. For this, we (i) allocate a sufficient number of cores to each active application, (ii) allocate the unassigned cores to the applications yielding the largest gain in energy, and (iii) choose for each application the best topology for its subset of cores (i.e., better than the by default “line” topology). This is a joint work with Ismail Assayad (U. Casablanca, Morocco) who visits the team regularly.

Second, we have proposed the first of its kind multi-criteria scheduling heuristics for a DAG of tasks onto an homogeneous multi-core chip, optimizing the execution time, the reliability, the power consumption, and the temperature. Specifically, we have worked on the static scheduling minimizing the execution time of the application under the multiple constraints that the reliability, the power consumption, and the temperature remain below some given thresholds. There are multiple difficulties: (i) the reliability is not an invariant measure w.r.t. time, which makes it impossible to use backtrack-free scheduling algorithms such as list scheduling [28]; to overcome this, we adopt instead the Global System Failure Rate (GSFR) as a measure of the system's reliability, which is invariant with time [46]; (ii) keeping the power consumption under a given threshold requires to lower the voltage and frequency, but this has a negative impact both on the execution time and on the GSFR; keeping the GSFR below a given threshold requires to replicate the tasks on multiple cores, but this has a negative impact both on the execution time, on the power consumption, and on the temperature; (iii) keeping the temperature below a given threshold is even more difficult because the temperature continues to increase even after the activity stops, so each scheduling decision must be assessed not based on the current state of the chip (i.e., the temperature of each core) but on the state of the chip at the end of the candidate task, and cooling slacks must be inserted. We have proposed a multi-criteria scheduling heuristics to address these challenges. It produces a static schedule of the given application graph and the given architecture description, such that the GSFR, power, and temperature thresholds are satisfied, and such that the execution time is minimized. We then combine our heuristic with a variant of the $\epsilon$-constraint method [52] in order to produce, for a given application graph and a given architecture description, its entire Pareto front in the 4D space (exec. time, GSFR, power, temp.). This is a joint work with Athena Abdi and Hamid Zarandi from Amirkabir U., Iran, who have visited the team in 2016.

Concurrent flexible reversibility

Reversible concurrent models of computation provide natively what appears to be very fine-grained checkpoint and recovery capabilities. We have made this intuition clear by formally comparing a distributed algorithm for checkpointing and recovery based on causal information, and the distributed backtracking algorithm that lies at the heart of our reversible higher-order pi-calculus. We have shown that (a variant of) the reversible higher-order calculus with explicit rollback can faithfully encode a distributed causal checkpoint and recovery algorithm. The reverse is also true but under precise conditions, which restrict the ability to rollback a computation to an identified checkpoint. This work has currently not been published.