Supercomputing Frontiers and Innovations

Bacterial Mini Microtubule as a Minimal Model System for Exploring Dynamic Instability Using Molecular Dynamics Simulations

Wed, 08 Oct 2025 00:00:00 +0500

Large scale computational modeling has been fruitfully applied to explore microtubules – an essential component of the cellular skeleton – for over two decades. In this paper, we describe simulations of a yet computationally unexplored minimalistic system of the bacterial mini microtubule, using the high performance resources of Lomonosov Moscow State University. We highlight similarities between the eukaryotic and bacterial microtubules at the protofilament level, the size and stability of the entire mini microtubule system and the computational benefits of using the bacterial mini microtubule as a minimal model to understand dynamic instability. Our results are discussed in the context of a bigger picture of the evolution of molecular dynamics simulations, aiming to understand microtubules, illustrating how the sophistication and scale of the computational efforts increased over the years.

GPU Implementation of Zippel Method for Feynman Integral Reconstruction

Alexander V. Smirnov, Boris I. Rozhnov, Vadim V. Voevodin — Wed, 08 Oct 2025 00:00:00 +0500

The Zippel algorithm performs a rational reconstruction of multivariate polynomials and aims specifically at the sparse case, where other approaches, such as iterative Newton and Thiele reconstructions, have a significantly higher complexity. It is applied in different fields of science, lately becoming an important step in Feynman integral reduction in elementary particle physics within the modular approach to reduction. For some cases with multiple variables the Zippel reconstruction might become a bottleneck for the whole evaluation so that different optimizations are required. In this paper, we describe how we ported the classical Zippel algorithm for polynomials together with its balanced version for rational functions to graphical processor units (GPUs), as well as carried out its performance evaluation on several types of GPUs. According to our information, this is the first publically available implementation of this algorithm on GPUs, and the results show speedup up to 14.5 times compared to CPU-based version.

Mastering 3D-detection of Extensive Air Showers in Cherenkov Light

Wed, 08 Oct 2025 00:00:00 +0500

A new SPHERE series complex extensive air shower detector is under development. The main goal of its mission is to study the mass composition of cosmic ray nuclei in the 1–100 PeV energy range at a new level. The helium-filled balloon will be substituted by an unmanned aerial vehicle as a carrier. This change opens the upper hemisphere for observations. The already well-established telescope of Cherenkov light reflected from the snow-covered ice surface of Lake Baikal from an altitude of 500–1000 m will be supported by a detector of direct light pointed upward. Since the two detectors will study the same shower at different stages of its development, it could be called a 3D-detection, which is completely new for the EAS method. The development is based on an extensive MC modeling of the shower and the detection process using the Supercomputer Complex of the Lomonosov Moscow State University.

Prospects for Improving Computational Efficiency of Hydrodynamic Simulations on Supercomputers by Increasing the Number of GPUs per Compute Node

Sergei S. Khrapov, Ekaterina O. Agafonnikova, Alexander V. Khoperskov — Wed, 08 Oct 2025 00:00:00 +0500

Hydrodynamic models for studying surface water dynamics on realistic topography place special demands on computational performance. Such simulations must cover large areas to ensure hydrological connectivity of the territory due to the influence of catchment areas. On the other hand, small topographic inhomogeneities on the scale of a meter are often the determining factors of fluid dynamics. Our analysis is based on a model of surface water and sediment dynamics for a large mountainous area of the Krasnodar region under rainfall/runoff conditions. The results of such large-scale models can be provided by parallel OpenMP-CUDA codes for computing systems with multi-GPU. We focus on different ways of transferring data between GPUs using both GPUDirect and HostCopy technologies on computing systems with one to eight GPUs. The parallel code with HostCopy is on average several times slower and less efficient compared to the GPUDirect approach. We propose to use auxiliary characteristics to analyze the efficiency of parallel implementation of a numerical algorithm. These values are calculated based on the average processing time of one computational cell and allow us to determine the optimal grid resolution in terms of performance.

Comparison of Quantum-Chemical Programs and Methods for the Calculation of Enthalpies of Formation of High-Energy Tetracyclic Compounds

Wed, 08 Oct 2025 00:00:00 +0500

A comparative study was carried out on the thermochemical properties of a series of high-energy tetracyclic compounds containing amino, cyano, azido, and dinitrophenyl groups. Various quantum-chemical methods were employed to calculate the gas-phase enthalpies of formation, including the B3LYP functional with 6-311+G(2d,p) and cc-pVTZ basis sets, the composite G4MP2 method implemented in Gaussian 09, and a G4MP2-based scheme adapted for implementation in NWChem. The G4MP2 method was used as a reference for accuracy, against which the results of other approaches were evaluated. It is shown that the use of NWChem and reaction-based schemes yields enthalpy values close to those obtained by G4MP2, while significantly reducing computational costs. Structural factors affecting the enthalpy of formation are analyzed, along with differences in the IR absorption spectra. The results confirm the applicability of various theoretical levels for the thermochemical evaluation of promising high-energy materials.

Dynamic Content-Oriented Indexing and Replication for High-Performance Storage and Analysis of Big Data in the IPFS Network

Maxim V. Shevarev, Stanislav V. Suvorov — Wed, 08 Oct 2025 00:00:00 +0500

This paper presents an architecture for dynamic, content-oriented indexing and adaptive replication that enables high-performance storage and analysis of big data on IPFS. We first outline key gaps of vanilla IPFS for analytics – no global content search, non-guaranteed persistence without coordinated pinning, static replication, and highly variable retrieval latency – and address them with two components: (1) a two-tier distributed index (per-attribute/keyword inverted lists as IPFS objects plus a lightweight catalog that maps search keys to index CIDs via DHT/IPNS or CRDT-based dissemination); and (2) an adaptive replication service that aggregates access telemetry and adjusts replica counts and placement using hysteresis thresholds and topology-aware selection. The contribution is a theoretical proposal and architectural blueprint; no prototype or experimental results are reported here. We discuss integration with analytical engines through two paths: a pragmatic FUSE mount that exposes IPFS content as a local filesystem to Spark/Flink, and prospective native connectors that parallelize block reads over the IPFS API. For tabular datasets, dataset metadata (schema, partitioning, file CIDs) is maintained in IPFS to support versioning and reproducibility. A plan for comparative evaluation versus HDFS, Ceph, and S3 (e.g., TPC-DS and subsets of Common Crawl) is outlined. Expected benefits are faster content discovery, higher throughput under skew and multi-tenant load, and improved resilience, with modest index/coordination overheads. The approach combines the openness of a decentralized P2P substrate with the manageability required by enterprise-scale analytics.