<h3 class="showhide_heading" id="Participants">Participants</h3>
<p>Tamas Elteto, Cécile Germain-Renaud, Balazs Kegl, Julien Perez, Michèle Sebag, Xiangliang Zhang<br />
 </p>
<h3 class="showhide_heading" id="Research_Themes">Research Themes</h3>
<p>Autonomic computing (Kephart & Chess 2003) targets self-optimization, self-configuration, self-healing, and self-protection of computing systems, mostly distributed. As a specialization to the grid context, our research targets self-regulation and maintenance under the constraints of production grids.</p>
<p>The motivation for an approach of grid modelling based on statistical analysis, computational learning and, to some extent, data mining, is the proved complexity of the individual components of the grid, and the potentialization effect of their interaction. Due to the large range of dynamic resources (basic hardware and middleware, but also software usage rights, sensors, etc.), the collective behaviour of users, and the mutualisation paradigm that founds the grid concept, such a large distributed system cannot be modelled only through a-priori analysis.</p>
<p>The operational basis for this research is the creation of the <i><a href="http://www.grid-observatory.org/">Grid Observatory</a></i>activity, chaired by C. Germain, in the EGEE-III project. The goal of this activity is to integrate the collection and publication of data on the behaviour grid users, exploiting the rich monitoring infrastructure of EGEE, with the development of models and of an ontology for the domain knowledge.</p>
<p>Concise representations of the grid behaviour (filtering, dimensionality reduction, possibly clustering) are necessary in order to make both analysis and publication manageable. Collaborations with research projects have already started (CoreGrid, KDUbiq) in order to develop the relevant concepts and schemas for each domain.</p>
<p>The second step is to characterize the quantities that are internally used in on-line grid management, such as job duration, or the volume of data transfers, or which are critical to dimensioning, such as the cluster and queues loads. Both intrinsic descriptions (the requests) and middleware dependent metrics must be addressed. Preliminary studies \cite{le poster de Julien} have shown that, even for these simple components, the characterization is far from trivial, long-range dependence being the norm.</p>
<p>The next step is to go beyond this collection of profiles, by characterizing, and finally explaining, their interactions. Integrating the grid new concept of Virtual Organization (VO) might be a key point. Grid users are organized along VOs, which define the qualitative (granted or denied) as well as the quantitative (share) software and hardware access rights. Correlated activity (computation, file access, database requests) will be created by the common timelines of related institutions or individuals; these correlations are both temporal (deadlines, “interesting” experimental events) and spatial (the researchers, the data, and the available computing power of a VO are not uniformly distributed over the grid). The characteristics of complex system also appear in the VO structure, with small-world graphs (Foster).</p>
<p>The two main application areas are scheduling on one hand, and real-time fault diagnosis and prediction on the other hand.<br />
 </p>
<h4 class="showhide_heading" id="Scheduling">Scheduling</h4>
<p><br />
 </p>
<ul>
<li>Optimization is required to assess the effectiveness of grid scheduling policies. Post-mortem analysis, where all information is derived from the analysis of traces, is just an instance of a classical optimization problem, for which the grid scale will require efficient approximation methods.</li>
<li>In an on-line setting, and in relation with the reactive grid framework (reference sur la partie agir), the scheduler should learn on-line an optimal policy maximizing both the long-term expected productivity and utility (Jensen, vengerov) as a function of the system current state, which is the goal of the above-mentioned models.</li>
</ul>
<h4 class="showhide_heading" id="Fault_diagnosis_and_prediction">Fault diagnosis and prediction</h4>
<p>Efficient end-to-end probing, where commands or transactions are sent from highly reliable sources and their results analyzed on-line, require an adaptive and hierarchical probing scheme (Rish). An alternative approach is to exploit the actual production of the grid as probes (Schuster). The final stepis to integrate both approaches, typically by considering the optimal balance of probes and passive analysis as a multi-criteria optimization problem.The following challenges can be envisioned.<br />
 </p>
<ul>
<li>The size of the datasets, even after filtering, will require scaling the standard techniques usually used on small or moderately large sets. In an interesting interplay between the grid as an object of research and the grid as a tool, it might be possible to use the grid itself as a computational resource to speed up the analysis.</li>
<li>It will be necessary to adapt the algorithms so they can deal with the structured nature of the data. Third, since the ultimate goal of this research project is to contribute to the understanding the grid, it is very desirable to create models that can be interpreted by human experts.</li>
</ul>
<h3 class="showhide_heading" id="Related_projects">Related projects</h3>
<p>The grid Observatory project is supported by both the EU project EGEE<br />
 and by the DIGITEO fundation.<br />
<br />
<br />
 </p>
<p> </p>