SCCS Colloquium - Sep 30, 2020

From Sccswiki
Jump to navigation Jump to search
Date: Sep 30, 2020
Room: Online (BBB)
Time: 15:00 - 16:00

Jeremy Harisch: Optimizing MPI Load Balancing in ls1 mardyn

Bachelor's thesis submission talk. Jeremy is advised by Steffen Seckler and Fabio Gratl.

Since molecular dynamic simulations are very calculation heavy tasks, load balancing is one of the key points which should be optimized to gain an optimal runtime. In this thesis, three load balancing strategies of the simulation framework ls1-MarDyn will be discussed due to optimization. Since ls1-MarDyn uses MPI as its communication protocol some time is spent inside of the MPI-routines and thus falsifies the time which is used as an input for the load balancing. A way will be shown how to reduce this time by the time spent inside of the MPI-routines. Furthermore, this implementation then is evaluated using three different simulations and in comparison, the newly reduced time leads to an overall better performance - but this only applies if it is used in combination with a diffusive domain decomposition.

Keywords: MPI Load-Balancing ls1-MarDyn AutoPas

Nicolas Witter: From Predictive to Reactive Load Balancing:Exploring Modern Load BalancingApproaches

Bachelor's thesis submission talk. Nicolas is advised by Philipp Samfass.

In the pursuit of increasing the total available processing power, modern systems are continuously growing more complex. As a result of this trend applications need to adapt to make full use of the increasing capacities, involving the identification of newly unfolding problems, such as the more recent focus on recognizing an rebalancing asymmetrical distribution in workload. To aid in the research and development of such solutions, this thesis attempts to by creating an application to deterministically obstructing another program, simulating such imbalances. In the progress MPI, OpenMP and Chameleon were explored, the first two were also utilized in the application itself to provide adaptability to the current environment. In the generation of the interference effect itself multiple different approaches considered, two of them implemented in the final solution. The effect of the created interference program was assessed on the CoolMUC2 massively parallel processor by running it alongside a benchmark application. While this confirmed the core functionality of the application that, with intelligent setup, is capable of creating more complex pattern, in practical use some iteration might be necessary, as the strength of the generated effect may differ depending on the targeted application and environment.

Keywords: High Performance Computing