Neumann, Philipp

Many combinatorial optimization methods and related optimization software, particularly those for mixed-integer programming, exhibit limited scalability when utilizing parallel computing resources, whether across multiple cores or multiple nodes. Nevertheless, high-performance computing (HPC) systems continue to grow in size, with increasing core counts, memory capacity, and power consumption. Rather than dedicating all available resources to a single problem instance, HPC systems can be leveraged to solve multiple optimization instances concurrently – a common requirement in applications such as stochastic optimization, policy design for sequential decision making, parameter tuning, and optimization-as-a-service. In this work, we study strategies for efficiently allocating solver jobs across compute nodes, exploring how to schedule multiple optimization jobs across a given number of cores or nodes. Using metrics from performance monitoring and benchmarking tools as well as metered PDUs, we analyze trade-offs between energy consumption and runtime, providing insights into how to balance computational efficiency and sustainability in large-scale optimization workflows.

In the scope of the dtec.bw project hpc.bw, innovative HPC hardware resources were procured to investigate their performance for HSU-relevant compute-intensive software. Benchmarks for different software packages were conducted, and respective results are reported and documented in the following, considering the Intel Xeon architecture used in the HPC cluster HSUper, AMD EPYC 7763 and ARM FX700.

Benchmarking the performance of one’s application in high performance computing (HPC) systems is critically important for reducing runtime and energy costs. Yet, accessing the plethora of relevant metrics that impact performance is often challenging, particularly for users without hardware experience. In this paper, we introduce the novel benchmarking tool xbat developed by MEGWARE GmbH. xbat requires no setup from the user side, and it allows the user to run, monitor and evaluate their application from the tool’s web interface, consolidating the entire benchmarking process in an approachable, intuitive workflow. We demonstrate the capabilities of the tool using benchmark applications of varying complexity and show that it can manage all aspects of the benchmarking workflow in a seamless manner. In particular, we focus on the open-source molecular dynamics research software ls1 mardyn, and the closed-source optimisation package Gurobi. Both packages present unique challenges. Mixed-integer programming solvers, such as those integrated in the Gurobi software, exhibit significant performance variability, so that seemingly innocuous parameter changes and machine characteristics can affect the runtime drastically, and ls1 mardyn comes with an auto-tuning library AutoPas, that enables the selection of various node-level algorithms to compute molecular trajectories. Focusing on these two packages, we showcase the practicality, versatility and utility of xbat, and share its current and future developments.

There exists an extensive literature of algorithms for short-range pairwise interactions in particle simulations, however, there is no single algorithm that performs the most optimally in every scenario, motivating the use of auto-tuning to select the optimal pairwise interaction algorithm. Previous efforts to auto-tune Molecular Dynamics have focused on Single-Site Molecular Dynamics, where the computational cost for the intermolecular force calculation is constant. Alternatively, for Multi-Site Molecular Dynamics, the cost of this calculation varies with the number of sites, which, as we show in this paper, can result in different optimal algorithms. Despite this further benefit for auto-tuning, it has yet to be applied to Multi-Site Molecules. In this paper, we introduce an implementation of Multi-Site Molecular Dynamics that is integrated with AutoPas. Using this implementation, we analyse how the relative performance between these algorithms varies as the number of sites varies, for both homogeneous and heterogeneous molecule distributions, and for two different hardware. Furthermore, we demonstrate the advantage of auto-tuning in the context of Multi-Site Molecular Dynamics using the node-level short-range particle simulation library, AutoPas.