Workload Characterization for Memory Management in Emerging Embedded Platforms - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Workload Characterization for Memory Management in Emerging Embedded Platforms

Biswadip Maity
  • Fonction : Auteur
  • PersonId : 1359665
Bryan Donyanavard
  • Fonction : Auteur
  • PersonId : 1359666
Nalini Venkatasubramanian
  • Fonction : Auteur
  • PersonId : 1093720
Nikil Dutt
  • Fonction : Auteur
  • PersonId : 1120098

Résumé

Memory has emerged as a primary performance and energy bottleneck for emerging embedded platforms that integrate heterogeneous compute units. Applications require a balance between performance and energy-efficiency and finding the optimal operating point on embedded platforms is challenging. There exist many opportunities to manage the memory subsystem efficiently at runtime to save energy without compromising quality in the face of dynamic workloads. Previous works have used memory bandwidth utilization to determine memory requirements and develop runtime policies to configure system knobs (e.g., memory controller frequency) accordingly. However, bandwidth utilization as a singular metric is not always sufficient: policies for a range of workload scenarios require insight into an application’s memory access pattern and working set size. Alternatively, memory profilers provide fine-grained information such as the memory access pattern for the entire virtual address space, and the load/store density of different regions of the memory. However, parsing this detailed information frequently at runtime induces excessive overhead. In this work, we propose a profiling mechanism that considers both (1) the working set size of running workloads and (2) memory bandwidth utilization to compute WBP (Working Set Size-Bandwidth Product). WBP can be estimated with low overhead, and the combined metric provides insights that runtime policies can use to determine desirable configurations for specific workload scenarios. Our early results show that a static configuration devised with this metric yields an optimal memory controller frequency for 8 out of 10 PARSEC workloads, demonstrating the promise of this approach.
Fichier sous embargo
Fichier sous embargo
1 8 5
Année Mois Jours
Avant la publication
jeudi 1 janvier 2026
Fichier sous embargo
jeudi 1 janvier 2026
Connectez-vous pour demander l'accès au fichier

Dates et versions

hal-04489020 , version 1 (04-03-2024)

Licence

Paternité

Identifiants

Citer

Biswadip Maity, Bryan Donyanavard, Nalini Venkatasubramanian, Nikil Dutt. Workload Characterization for Memory Management in Emerging Embedded Platforms. 6th International Embedded Systems Symposium (IESS), Sep 2019, Friedrichshafen, Germany. pp.65-76, ⟨10.1007/978-3-031-26500-6_6⟩. ⟨hal-04489020⟩
9 Consultations
1 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More