Toward a Push-based Stream Programming Model with AIMSS: An Active In-Memory Storage System Approach - Archive ouverte HAL
Pré-Publication, Document De Travail (Working Paper) Année : 2024

Toward a Push-based Stream Programming Model with AIMSS: An Active In-Memory Storage System Approach

Résumé

This paper introduces the vision for an Active In-Memory Storage System (AIMSS), a novel architecture that shifts data movement responsibilities, such as source handling, sink management, and data shuffling, from applications like training large language models (LLMs) and big data streaming engines, directly to AIMSS. AIMSS will operate on a log-structured in-memory storage framework, leveraging immutable data access patterns, facilitating efficient real-time data movement. The AIMSS architecture deploys on CPU and GPU nodes, harnessing their memories and ensures efficient and transparent communication with disk-based file storage systems. We propose a push-based stream programming execution model that allows AIMSS to cost-effectively harness application-specific data (such as consumer offsets and data access patterns including read, write, and shuffle) and thereby enable a set of data-based optimizations. These include scalable data movement partitioning algorithms, faster stream storage recovery (speeding up application restarts), easy identification of application stragglers, and mitigation of power fluctuation issues during large-scale LLM training (e.g., by efficiently leveraging idle GPU resources for other computing tasks). Furthermore, AIMSS will minimize I/O interference in multi-CPU-GPU setups for multiple applications sharing a high-performance computing infrastructure, including CPUs, GPUs, and advanced interconnects. AIMSS promises significant performance improvements by actively handling data movement for data-intensive applications and by combining in-memory processing with a novel push-based stream programming model suitable for exascale computing.
Fichier principal
Vignette du fichier
aimss-push-model-system-vision-paper-final09.pdf (488.56 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04683563 , version 1 (02-09-2024)
hal-04683563 , version 2 (09-09-2024)

Licence

Identifiants

  • HAL Id : hal-04683563 , version 1

Citer

Ovidiu-Cristian Marcu, Grégoire Danoy, Pascal Bouvry. Toward a Push-based Stream Programming Model with AIMSS: An Active In-Memory Storage System Approach. 2024. ⟨hal-04683563v1⟩
32 Consultations
18 Téléchargements

Partager

More