Evaluating trace aggregation for performance visualization of large distributed systems

Abstract

Performance analysis through visualization techniques usually suffers semantic limitations due to the size of parallel applications. Most performance visualization tools rely on data aggregation to work at scale, without any attempt to evaluate the loss of information caused by such aggregations. This paper proposes a technique to evaluate the quality of aggregated representations -using measures from information theoryand to optimize such measures in order to build consistent multiresolution representations of large execution traces.

A Generic Algorithmic Framework to Solve Special Versions of the Set Partitioning Problem

“…Finding the optimal partition thus arises as soon as one wants to organize, to classify, or to abstract data. In this context, the SPP also relates to data aggregation problems, where data points are partitioned into homogeneous classes preserving the system's structure and optimizing a given compression rate, leading to applications in time series analysis [26,28,40], spacial analysis [20,28,30], multilevel community detection [43], database optimization, and image processing [35,38].…”

“…Although considerable results have been achieved for such settings, Sandholm et al [49] argue that, when some cost penalizes the coalition formation process itself (because of communication or anti-trust penalties), many applications of the coalition generation problem are neither superadditive nor subadditive. This is also the case in multilevel data analysis, where the information-theoretic measures are usually non-monotonous regarding the set inclusion, meaning that data points might be relatively homogeneous at some level, but heterogeneous at lower or higher levels [20,28,30,40,43].…”

“…However, constraints might also arise from semantics considerations when some subsets are not meaningful for the partition purposes. In data aggregation, for example, the partitioning should be consistent with the dataset's structural and topological properties so that the compressed data is usable by the domain expert [20,28,30,40]. In particularly, in image processing, the two-dimensional structure implies very strong constraints for compression [38].…”

“…For another example, in data analysis, multilevel aggregation consists in partitioning a set of data points into homogeneous classes. In order to be meaningfully interpreted by the analyst, the classes often need to preserve the system's semantic or topological structure, expressed by the set of admissible parts [28,30,38]. In this context, the cost function is usually defined according to a compression rate that should be maximized, leading to numerous applications such as time series analysis [26,28,40], database optimization and image processing [35,38], multilevel community detection [43], performance analysis of distributed systems [20,30], resource and task allocations for multiagent systems [50].…”

“…Such constraints, arising from global (semantic, structural, functional, or topological) properties of the modeled population, thus define tractable special versions of the SPP. For example, in data aggregation, the dataset's topological properties may allow to define a class of search spaces that are meaningful for the domain expert: e.g., assuming that admissible parts are intervals in the case of time series analysis [26,28,40], that they form a hierarchy in the case of spatial analysis of nested structures [28,30,43], or that they are rectangular tiles in the case of image processing, where the two-dimensional structure of the dataset implies very strong constraints for compression [38]. In this paper, we thus assume that the expert provides such global constraints to reduce the SPP computational complexity and provide tractable optimization algorithms.…”