Title: Summarization using MDL and Dynamic Programming
Speaker: Vanessa Cedeno
Abstract: Summarization is an analytical approach that develops a concise representation of the given structured data set. Unlike compression or HMM, the goal is to derive representations that provide a global view and a better understanding of a dataset. Having a global overview allows the user to explore different options of where to start a more detailed analysis. In this work, we address the summarization over log generated by monitors and sensors placed in high-performance computing systems such as Cab, a 431 TFLOPS/s Intel Xeon cluster with 1,296 nodes. The HPC ecosystem of applications, compute nodes, racks, networking fabric, job manager, scheduler, etc. form a large and complex non-linear system. In such systems, modeling and understanding the performance and fault tolerance becomes highly challenging. Since manual analysis is infeasible for such systems, interest has been towards developing automatic analyzers of the log data. Our approach is parameter-free and creates optimal global summaries using Minimum Description Length and Dynamic Programming. We model the monitor output and application schedule runs as discrete time events and discrete time intervals. Experiment results show that the summaries produced from our algorithms are optimal and the use of MDL has an intuitive interpretation.
Title: Effects of Network Structure on Propagation of Infectious Diseases
Speaker: Madhurima Nath
Abstract: The dynamics on a finite sized interacting system, modelled by a network, is affected by the underlying structure. The propagation of disease in a population as a diffusive process have been extensively studied using networks. There are several methods proposed in the literature for building equivalence classes or families of networks based on the structural aspects of the system. It is observed that similarities in the local statistics of two networks are not sufficient to predict the dynamics on them. A global statistic, the Moore and Shannon’s reliability polynomial, is suggested to explore the outbreak of an infectious disease on a network. It gives the probability that a system composed of many different interacting components has a desired property and depends on both the structure and the dynamics. The estimation of the reliability polynomial for networks with hundreds of millions of interactions is feasible using Monte-Carlo simulation as exact computations is often NP-hard. The network reliability as a function of the probability of transmission of a disease allows to map parameters of one network on to another using simple transformation which keeps the dynamics invariant.
Meeting number: 643 164 789
Host key: 749673
+1-855-749-4750 US TOLL FREE
+1-415-655-0001 US TOLL
Show toll-free dialing restrictions
Access code: 643 164 789