Maisto, Vincenzo (2024) Harnessing Energy Across the Computing Landscape: Cross-domain Approaches for Energy Efficiency and Hardware Consolidation. [Tesi di dottorato]

Documento PDF MaistoVincenzo_ITEE_PhDThesis.pdf Visibile a [TBR] Amministratori dell'archivio Download (21MB) | Richiedi una copia

Abstract

Computing technologies are nowadays ubiquitous and their demand is expanding at an ever-increasing rate. The environmental footprint of such immense growth is a severe environmental issue, with computation alone expected to account for 20% of the global energy demand by 2030. If on one hand, the interest in green computing approaches is rising, on the other, the modern digital computing landscape is extremely diverse and heterogeneous. High-end cloud systems are cluster-based and business-oriented, edge and IoT devices are concerned with field deployment and battery power, while low-power micro-architectural designs focus on low-level optimizations. Nevertheless, all domains share the urge for lower energy consumption, either to save on the datacenter energy bill or prolong battery lifetime. Unfortunately, most techniques are domain-specific and non-generalizable, and there seems to be no silver bullet to save energy. In the quest for energy consumption reduction across computing domains, this thesis presents and validates common approaches for energy efficiency and hardware consolidation across the digital computing landscape. We target server-class HPC clusters, edge MPSoC platforms, and low-power micro-architectural designs. We enable hardware consolidation with advanced hardware/software co-design, scalability, and heterogeneity, and avoid domain-specific methodologies. In the HPC domain, we enable efficient acceleration of erasure coding in distributed file systems with low-power and scalable PCIe-attached accelerators. In edge computing, we focus on multi-tenant edge-AI on MPSoC platforms and propose our heterogeneous deep learning acceleration platform. Finally, we address low-power RISC-V vector consolidation, enabling and validating multi-threaded vector computation on open-source micro-processors.

Downloads

Actions (login required)

Modifica documento