Jean-Marc Vincent

Random graph generation for scheduling simulations

In parallel and distributed systems, validation of scheduling heuristics is usually done by simulation on randomly generated synthetic workloads, typically represented by task graphs. Since there is no single generation method that models all possible workloads for scheduling problems, researchers often re-implement the classical generation algorithms or even implement ad hoc ones. A bad choice of generation method can mislead the validation of the algorithm due to biases it can induce. Moreover, different implementations of the same randomized generation method may produce slightly different graphs. These problems can harm the experimental comparison of scheduling algorithms. In order to provide a comparison basis we propose GGen -- a unified and standard implementation of classical task graph generation methods used in the scheduling domain. We also provide an in-depth analysis of each generation method, emphasizing important graph properties that may influence scheduling algorithms.

Visualisation of Distributed Applications for Performance Debugging

This paper presents a method to perform visualisations of the behaviour of distributed applications, for performance analysis and debugging. This method is applied to a Java distributed application. Application level traces are recorded without any modification of the monitored applications nor of the JVMs. Trace recording includes records from the JVM, through the JVMPI, and records from the OS, through the data structure associated to each process. Recorded traces are visualised post mortem, using the interactive Paje visualisation tool, which can be conveniently specialised to visualise the dynamic behaviour of distributed Java applications. Applying this method to the execution of a book server, we were able to observe a situation where both the computation or the communications could be at the origin of a lack of performances. The observation helped finding the origin of the problem coming in this case from the computation.

A flexible checkpoint/restart model in distributed systems

Large scale applications running on new computing platforms with thousands of processors have to face with reliability problems. The failure of a single processor will cause the entire execution to fail. Most existing approaches to guarantee reliable executions are based on fault tolerance mechanisms. Coordinated checkpointing is one of the most popular technique to deal with failures in such platforms. This work presents a new model of coordinated Checkpoint/Restart mechanism for several types of computing platforms. The model is parametrized by the process failure distribution, the cost to save a global consistent state of processes and the number of computational resources. Through mathematical analysis of reliability, we apply this new model to compute the optimal interval between checkpoint times in order to minimize the average completion time. Model independency from the type of the failure law makes it completely flexible. We show that such a model may be used to reduce the checkpoint rate up to 20% in same cases and up to factor 4 the total overhead in same cases. Finally, we report some experiments based on simulations for random failure distributions corresponding to the two most popular laws, namely, the Poissons process and Weibulls law.

Perfect simulation and non-monotone Markovian systems

Perfect simulation, or coupling from the past, is an efficient technique for sampling the steady state of monotone discrete time Markov chains. Indeed, one only needs to consider two trajectories corresponding to minimal and maximal state in the system. We show here that even for non-monotone systems one only needs to compute two trajectories: an infimum and supremum envelope. Since the sequence of states obtained by taking infimum (resp. supremum) at each time step does not correspond to a feasible trajectory of the system, the envelopes might not couple or the coupling time might be larger. We show that the envelope approach is efficient for some classes of non-monotone queuing networks, such as networks of queues with batch arrivals, queues with fork and join nodes and/or with negative customers.

Physical cloth simulation on a PC cluster

Cloth simulation is of major interest in 3D animation, as it allows the realistic modeling of dressed humans. The goal of our work is to decrease computation time in order to obtain real time dynamics animation. This paper describes a cloth simulation and addresses the problem of parallelizing the implicit time integration and to couple a parallel execution with a standard visualization. We believe that this work could benefit to other applications based on a conjugate gradient solution and other applications of PC clusters.

Detection and analysis of resource usage anomalies in large distributed systems through multi-scale visualization

Understanding the behavior of large scale distributed systems is generally extremely difficult as it requires to observe a very large number of components over very large time.

Perfect simulation of index based routing queueing networks

Markovian queueing networks models are widely used for performance evaluation of computer systems, production lines, communication networks and so on. Routing strategies allocate clients to queues after the end of service. In many situations such as deterministic, probabilistic, or state dependent like Join the shortest queue routing, the routing function could be written in terms of index scheduling functions introduced in [3, 6].

How to Build the Best Macroscopic Description of Your Multi-Agent System?

The design and debugging of large-scale MAS require abstraction tools in order to work at a macroscopic level of description. Agent aggregation provides such abstractions by reducing the complexity of the microscopic description. Since it leads to an information loss, such a key process may be extremely harmful for the analysis if poorly executed. This paper presents measures inherited from information theory to evaluate abstractions and provide the experts with feedback regarding the quality of generated descriptions. Several evaluation techniques are applied to the spatial aggregation of an agent-based model of international relations. The information from on-line newspapers constitutes a complex microscopic description of agent states. Our approach is able to evaluate geographical abstractions used by the domain experts in order to provide efficient and meaningful macroscopic descriptions of the world global state.

Introducing Queuing Network-Based Performance Awareness in Autonomic Systems

This paper advocates for the introduction of performance awareness in autonomic systems. The motivation is to be able to predict the performance of a target configuration when a self-* feature is planning a system reconfiguration.We propose a global and partially automated process based on queues and queuing networks models. This process includes decomposing a distributed application into black boxes, identifying the queue model for each black box and assembling these models into a queuing network according to the candidate target configuration. Finally, performance prediction is performed either through simulation or analysis.This paper sketches the global process and focuses on the black box model identification step. This step is automated thanks to a load testing platform enhanced with a workload control loop. Model identification is then based on statistical tests. The model identification process is illustrated by experimental results.

The Best-Partitions Problem: How to Build Meaningful Aggregations

The design and the debugging of large distributed AI systems require abstraction tools to build tractable macroscopic descriptions. Data aggregation provides such tools by partitioning the system dimensions into aggregated pieces of information. Since this process leads to information losses, the partitions should be chosen with the greatest caution. While the number of possible partitions grows exponentially with the size of the system, this paper proposes an algorithm that exploits exogenous constraints regarding the system semantics in order to find the best partitions in a linear or polynomial time. Two constrained sets of partitions (hierarchical and ordered) are detailed and applied to spatial and temporal aggregation of an agent-based model of international relations. The algorithm succeeds in providing meaningful high-level abstractions for the system analysis.