ABSTRACT. Traces of information on events executed from a discrete event simulation kernel to capture run time profile information can easily lead to very large trace files. While disk space is relatively inexpensive, manipulating and analyzing these large trace files can prove difficult. Further complicating the matter, some types of analysis must be performed in-core and they cannot be performed with the trace data exceeds the size of the physical RAM where the analysis is performed. Because of these limits, it is often necessary to strictly limit the simulation run time to satisfy the analysis time memory limits. Currently the in-memory tools to analyze event trace file are limited to files on the order of 10GB (on a machine with 24GB of RAM). Furthermore, even when it is possible to analyze large trace files, the run time costs of performing this analysis can take several days to complete. The work reported in this paper explores techniques to analyze much larger trace files than currently possible. While out-out-core analysis explorations have been examined as part of this work, the run time costs for out-of-core processing can increase processing time 10-fold. As a result, the work reported here will focus on an approach to sample the event trace file. The trace sample are then analyzed in anticipation that their analysis will show similar results to analysis results obtained from a full trace. The work reported in this paper will examine how closely the analysis from sampling matches the analysis from a full trace file. Techniques for comparison including visual inspection of results overlaid from the full trace to the sample traces will be presented. In addition, results with the Wasserstein, Directed Hausdorff, and Kolmogorov-Smirnov distance metrics will be reported. Finally, the ability to process a very large trace file of 80GB is reported.

ABSTRACT. Over the last years, Machine Learning (ML) has gained a significant momentum as an instrumental artefact for constructing new or improving existing knowledge. Reflecting this trend, the paper spurs a discussion on the potential integration of simulation models and ML. Subsequently, the paper develops a conceptual framework to guide the implementation of that integration. At its core, our approach is based on the premise that system knowledge can be (partially) captured and learned in an automated manner aided by ML. We conceive that the approach can help realise adaptive simulation models that can learn to change their behaviour in response to behavioural changes in the system of interest. To demonstrate its validity, a set of experiments are presented to serve as a proof-of-concept. The study is conceived to foster new ideas and speculative directions towards integrating the practice of M&S with data-driven knowledge learned by ML.

ABSTRACT. Assessing the impact of changes in a production network (e.g., new routing protocols or topologies) requires simulation or emulation tools capable of providing results as close as possible to those from a real-world experiment. Large traffic loads and complex control-data plane interactions both constitute significant challenges to these tools. To meet these challenges, in the specific context of the recently introduced paradigm of Software-Defined Networking (SDN), we propose a model for the fast and convenient evaluation of SDN as well as legacy networks. Our approach emulates the network's control plane and simulates the data plane, to achieve high fidelity necessary for control plane behaviour, while being capable of handling large traffic loads. We design and implement a proof of concept that realises the envisioned model. The initial results of the prototype, compared to a state-of-the-art solution, shows that it can increase the speed of network experiments by more than 95% with almost constant times regardless of the network size.

ABSTRACT. Markov models have a long tradition in modeling and simulation of dynamic systems. In this paper, we look at certain properties of a discrete time Markov chain including entropy, trace, and 2nd largest eigenvalue to better understand their role for time series analysis. We simulate a number of possible input signals, fit a discrete time Markov chain and explore properties with the help of Sobol indices. This research is motivated by recent results in the analysis of cell development for Xenopus laevis in cell biology that relied on the considered entropy measure to distinguish development stages from time series data of Calcium levels in cells.

ABSTRACT. Parallel application performance models provide valuable insight about the performance in real systems. Capable tools providing fast, accurate, and comprehensive prediction and evaluation of high-performance computing (HPC) applications and system architectures would have important value. This paper presents PyPassT, an analysis based modeling framework, which is built on static pro- gram analysis and integrated with simulation of target HPC architectures. The framework analyzes the HPC application source code written in C with OpenACC directives and transforms it into an application model describing its computation and communication behavior (including CPU and GPU workloads, memory accesses, and MPI transactions). The application model is then executed on a simulated HPC architecture for performance analysis. Our experiments show that the proposed framework can represent the runtime behavior of benchmark applications with good accuracy.

ABSTRACT. The growth in many-core CPUs has motivated development of shared-memory, multithreaded solutions to minimize communication and synchronization overheads in Parallel Discrete Event Simulations (PDES). Analogous capabilities, such as Cross Memory Attach (CMA) based approaches have been added to Message Passing Interface (MPI) libraries. CMA permits MPI-processes to directly read/write data from/to a different process's virtual memory space to exchange messages. This paper compares the performance of CMA capable, MPI-based version to our fine-tuned multithreaded version. The paper also discusses implementation and optimization of the multithreaded infrastructure to elucidate the design alternatives being compared and assessed. Our experiments conducted using 2--28 threads and a fine-grained (time per event < 0.7 us) version of PHOLD benchmark show that message-passing outperforms multithreading (by ~10%--20%) in many scenarios but underperforms in others. The complex performance landscape inferred from our experiments suggest that more in-depth analysis of model characteristics is needed to decide between shared-memory multithreading versus message-passing approaches.

ABSTRACT. We describe an Abstract Model for Diffusion Processes to simulate diffusion processes in multiplex dynamic networks using formal modeling and simulation (M&S) methodologies (in this case, the DEVS formalism). Once we have defined a network where a diffusion process occurs and its diffusion rules, we can translate this information into an Abstract Model for Diffusion Processes, which is formally defined in DEVS, and can be converted into a computerized model. Using the proposed Abstract Model for Diffusion Processes, we can study a diffusion process in multiplex networks with a formal simulation algorithm, improving the model’s definition. We present a case study using the CDBoost simulation engine.