A Study of Low-Confidence Prefetches for Enhancing DRAM Row Buffer Locality

As high-performance computing (HPC) systems continue to grow in scale and complexity, memory performance has become a critical bottleneck. The widening gap between processor speed and DRAM access latency, commonly known as the memory wall, limits system throughput, while increasing data movement con...

Descripción completa

Detalles Bibliográficos
Autor: Migliari, Lorenzo
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/449738
Acceso en línea:https://hdl.handle.net/2117/449738
Access Level:acceso abierto
Palabra clave:Computer storage devices
Parallel processing (Electronic computers)
memory hierarchy
hardware prefetching
DRAM
row-buffer locality
Ordinadors--Dispositius de memòria
Processament en paral·lel (Ordinadors)
Àrees temàtiques de la UPC::Informàtica::Arquitectura de computadors
Descripción
Sumario:As high-performance computing (HPC) systems continue to grow in scale and complexity, memory performance has become a critical bottleneck. The widening gap between processor speed and DRAM access latency, commonly known as the memory wall, limits system throughput, while increasing data movement contributes significantly to overall energy consumption. Hardware prefetching is a well-established and essential feature of modern high-performance processors, aiming to reduce memory latency. However, current designs typically discard low-confidence prefetches to avoid cache pollution and bandwidth contention, which means that potentially valuable predictive information is overlooked. This thesis proposes a novel architectural technique that repurposes these discarded low-confidence prefetches as speculative hints to guide memory controller behaviour. Rather than triggering cache fills, these predictions are used to improve DRAM row-buffer locality by preemptively activating memory rows likely to be accessed in the near future. The goal is to enhance memory access efficiency and increase prefetch coverage, without incurring the typical side effects of aggressive prefetching, such as bandwidth saturation or cache pollution. The proposed approach was evaluated using a trace-based, cycle-accurate simulator through both a simplified prototype and a realistic simulation model. Results show that low-confidence prefetches frequently correspond to future memory accesses and can be exploited to reduce memory access latency and improve overall efficiency. This work reveals that low-confidence prefetches can offer significant utility, introducing a new perspective on how hardware prefetching capabilities could be extended.