Section: Scientific Foundations

Application and Resource Models

A second research direction consists in providing accurate, or at least realistic, models of applications and execution infrastructures. Such a goal has been the main concern of the SimGrid project for more than 10 years. Hence, this simulation toolkit provides most of the technological background to allow for the exploration of new scientific challenges. Moreover, simulation is a classical and efficient way to explore many “what-if” scenarios in order to better understand how an application behaves under various experimental conditions.

The Avalon team considers using simulation for application performance prediction. The scientific challenges lie in the diversity of applications and available execution environments. Moreover the behavior and performance of a given application may vary greatly if the execution context changes. Simulation allows us to explore many scenarios in a reasonable time, but this require to get a good understanding of both application structure and target environment.

A first focus is on HPC, regular, and parallel applications. For instance, we study those based on the message passing paradigm, as we have already developed some online and offline simulators. However, the different APIs provided by SimGrid allow us to also consider other kinds of applications, such as scientific workflows or CSPs.

A second focus is on data-intensive applications. It implies to also consider storage elements as a main modeling target. In the literature, the modeling of disk is either simplistic or done at a very-low level. This leads to unrealistic or intractable models that prevent the acquisition of sound information. Our goal is then to propose comprehensive models at the storage system level, e.g., one big disk bay accessed through the network. The main challenge associated to this objective is to analyze lots of logs of accesses to data to find patterns and derive sound models. The IN2P3 Computing Center gives us an easy access to such logs. Moreover a collaboration with CERN will allow us to validate the proposed model on an actual use case, the distributed data management system of the ATLAS experiment.

Modeling applications and infrastructures is in particular required to deal with energy concerns, as energy price is becoming a major limiting factor for large scale infrastructures. Physically monitoring the energy consumption of few resources is now becoming a reality; injecting such local measurements as a new parameter in multi-objective optimization models is also more and more common. However, dealing with energy consumption and energy efficiency at large scale is still a real challenge. This activity, initiated in the RESO team since 2008, is continued by the Avalon team by investigating energy consumption and efficiency on large scale (external, internal) monitoring of resources. Also, while physical resources start to be well mastered, another challenge is to deal with virtualized resources and environments.