Section: New Results

Tools for Performance Evaluation

Participants : Jean-Michel Fourneau, Brigitte Plateau, Jean-Marc Vincent.

Model Checking

This is collaborative work with Stavros Tripakis (Cadence Research Laboratories)

Exhaustive verification often suffers from the state-explosion problem, where the reachable state space is too large to fit in main memory. For this reason, and because of disk swapping, once the main memory is full very little progress is made, and the process is not scalable. To alleviate this, partial verification methods have been proposed, some based on randomized exploration, mostly in the form of random walks. In , we enhance partial, randomized state-space exploration methods with the concept of resource-awareness: the exploration algorithm is made aware of the limits on resources, in particular memory and time. We present a memory-aware algorithm that by design never stores more states than those that fit in main memory. We also propose criteria to compare this algorithm with similar other algorithms. We study properties of such algorithms both theoretically on simple classes of state spaces and experimentally on some preliminary case studies.

Achieving automatic performance modelling of black boxes for self-sizing

This is a collaborative work with Nabila Salmi (France Télécom), Bruno Dillenseger (France Télécom)

Modern distributed systems are characterized by a growing complexity of their architecture, functionalities and workload. This complexity, and in particular significant workloads, often lead to quality of service loss, saturation and sometimes unavailability of on-line services. To avoid troubles caused by important workloads and fulfill a given level of quality of service (such as response time), systems need to self-manage , for instance by tuning or strengthening one tier through replication. This autonomic feature requires performance modelling of systems. In [29] , we developed an automatic identification process providing a queuing model for a part of distributed system considered as black box. This process is a part of a general approach targetting self-sizing for distributed systems and is based on a theoretical and experimental approach. We show how to derive automatically the performance model of one black box considered as a constituent of a distributed system, starting from load injection experiments. This model is determined progressively, using self-regulated test injections, from statistical analysis of measured metrics, namely response time. This process is illustrated through experimental results.

Stochastic Automata Networks

With excellent cost/performance trade-offs and good scalability, multiprocessor systems are becoming attractive alternatives when high performance, reliability and availability are needed. They are now more popular in universities, research labs and industries. In these communities, life-critical applications requiring high degrees of precision and performance are executed and controlled. Thus, it is important for the developers of such applications to analyze during the design phase how hardware, software and performance related failures affect the quality of service delivered to the users. This analysis can be conducted using modeling techniques such as transition systems. However, the high complexity of such systems (large state space) makes them difficult to analyze.

In [37] , we present a new approach to obtain the Reachable State Space (RSS) of a structured model which uses functional transitions. We use Multi-valued Decision Diagrams (MDD) to store sets of reachable spaces and Stochastic Automata Networks (SAN) formalism to describe structured models. We propose a method to generate a compact MDD description taking advantage of the modular structure of SAN formalism, which also allows one to represent the transition rate matrix of a continuous-time Markov chain by means of a sum of generalized Kronecker products.