A Fast MPI-Based Distributed Hash-Table as Surrogate Model for HPC Applications

Lübke, Max; De Lucia, Marco; Petri, Stefan; Schnor, Bettina

doi:10.1007/978-3-031-97635-3_28

Datensatz

DATENSATZ AKTIONENEXPORT

Zur Ablage hinzufügen

Lokale TagsFreigabegeschichteDetailsÜbersicht

Freigegeben

Konferenzbeitrag

A Fast MPI-Based Distributed Hash-Table as Surrogate Model for HPC Applications

Urheber*innen

Lübke, Max
Potsdam Institute for Climate Impact Research;

De Lucia, Marco
External Organizations;

/persons/resource/petri

Petri, Stefan
Potsdam Institute for Climate Impact Research;

Schnor, Bettina
External Organizations;

Externe Ressourcen

https://doi.org/10.48550/arXiv.2504.14374
(Preprint)

https://doi.org/10.5281/zenodo.15235554
(Ergänzendes Material)

Volltexte (beschränkter Zugriff)

Für Ihren IP-Bereich sind aktuell keine Volltexte freigegeben.

Volltexte (frei zugänglich)

32218oa.pdf
(Preprint), 653KB

Ergänzendes Material (frei zugänglich)

Es sind keine frei zugänglichen Ergänzenden Materialien verfügbar

Zitation

Lübke, M., De Lucia, M., Petri, S., Schnor, B. (2025): A Fast MPI-Based Distributed Hash-Table as Surrogate Model for HPC Applications. - In: Lees, M. H., Cai, W., Ann, C. S., Yi, S., Abramson, D., Dongarra, J. J., Sloot, P. M. A. (Eds.), - Computational Science – ICCS 2025: 25th International Conference, Singapore, July 7-9, 2025, Proceedings, (Lecture Notes in Computer Science ; 15906), 233-240.
https://doi.org/10.1007/978-3-031-97635-3_28

Zitierlink: https://publications.pik-potsdam.de/pubman/item/item_32218

Zusammenfassung

Surrogate models can enhance performance in HPC applications. Cache-based surrogates use pre-calculated simulation results for interpolation or extrapolation. However, this is only effective if retrieval is faster than simulation. This paper proposes an MPI-based distributed architecture where parallel processes share memory to build a distributed hash table. Three DHT approaches for HPC are presented. A lock-free design outperforms both coarse-grained and fine-grained locking DHTs, demonstrating good scaling for read and write performance. The lock-free DHT improved the runtime of a reactive transport simulation up to 42%.