Aggregation and multiresolution representation of applications trace informations in the context of embedded systems
- Niveau : Master recherche
- Responsable : Guillaume Huard (Guillaume.Huard@imag.fr) and Jean-Marc Vincent (Jean-Marc.Vincent@imag.fr)
- Durée : 4 à 6 mois
Guillaume Huard, Maître de Conférences UJF, LIG-MOAIS
Jean-Marc Vincent, Maître de Conférences UJF, LIG-MESCAL
Nowadays, the complexity of embedded systems is growing at a fast pace. From an architectural point of view, they are evolving to include an increasing number of computation cores, potentially heterogeneous, that communicate using dedicated interconnection networks rather than shared memory. From a software point of view, they embed a standalone operating system which tends to become more elaborate with each new generation. Today, embedded systems are already able to include a simple linux kernel derivative along with libraries to support parts of the POSIX API.
Programming applications for these new systems is a tremendously difficult task, partly because of the inherent complexity of parallel and distributed development and partly because of inadequate environments for analysing, debugging and tuning application code. The main issue with these inadequate environments is related to the size of applications traces, even simple video decoders produce several gigabytes of traces when executing for a handful of seconds. Dealing with such large amount of information raises several issues including storage management, noise elimination, patterns recognition and visual representation.
This proposal focuses on visual representation of large application traces. The problem is to represent on a limited space, typically a screen with a fixed resolution, a meaningful information about an application trace. Usual visualization tools perform this task either by overdrawing data onto the screen, thus eliminating parts of the information, or by averaging values, thus smoothing everything. In both cases interesting informations might dispear from the visualization. With this proposal, our intention is to propose new aggregation mechanics to enable users to guide the visualization tool toward interesting and meaningful representations.
The Moais and Mescal teams focus their research on large scale systems and parallel applications. They have a strong extertise regarding parallel algorithms, environment for parallel programming, performance evalua- tion of distributed systems, middleware for clusters and grids and scheduling. Recently some of their research efforts have been related to the analysis of trace data produced by distributed applications. We can mention Triva [Sch09, SLV11], a tool for the visualization of large scale applications execution. Triva is able to perform the aforementionned aggregation techniques in order to build a compact and dynamic visual representation of the execution. Furthermore, Triva support alternative points of view (platform topology, communication structures, treemap) that help understanding the end result. We can also mentionned some recent result about the use of statistical analysis to outline peculiar behaviors in large scale distributed systems [JKVA11].
These teams are involved in the SoC-Trace project which aims at developping new methods and tools to collect, analyse and visualize traces of applications for embedded systems. SoC-Trace is a FUI funded project that gather academic (Joseph Fourier University and INRIA) and industrial (ST Microelectronics, Magillem and ProBayes) partners and is scheduled for next four years.
The first objective of this internship is to conduct a comprehensive bibliographical study about scientific vi- sualization and application traces analysis (statistical, structural) and visualization. Then, the second objective is to propose new aggregation methods suited to applications analysis. More precisely, the candidate will have to propose aggregation operators and to demonstrate on concrete traces that these operators outline informations that help analyzing, debugging or tuning the application. Finally, these methods will have to be adapted to a multiscale analysis context, in which the user guide the aggregations performed by the visualization tool. Firstly, this implies that the design of aggragetion operations enable their composition (they can be applied on aggregates). Secondly, under the user guidance, data aggregation might have to be performed only to interesting subparts of the trace. This results in a data set made of several aggregated groups of information, aggregated at different levels, that have to be represented on the same visual space. The issues here are to maintain relevant information about aggregation levels and application scale.
- [JKVA11] Bahman Javadi, Derrick Kondo, Jean-Marc Vincent, and D.P. Anderson. Discovering Statistical Mo- dels of Availability in Large Distributed Systems : An Empirical Study of SETI@home. IEEE Tran- sactions on Parallel and Distributed Systems, 99(preprint), 2011.
- [Sch09] Lucas Mello Schnorr. Some Visualization Models applied to the Analysis of Parallel Applications. PhD thesis, INP Grenoble, oct 2009.
- [SLV11] Lucas Mello Schnorr, Arnaud Legrand, and Jean-Marc Vincent. Multi-scale analysis of large distribu- ted computing systems. In Proceedings of the third international workshop on Large-scale system and application performance, LSAP ’11, pages 27–34. ACM, 2011.