Access Bit Scanning for Page Replacement Policies
Event details
| Date | 22.06.2026 |
| Hour | 10:00 › 12:00 |
| Speaker |
Arunkrishna Annai Madalam Sivasubramanian |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Thomas Bourgeat
Thesis advisor: Prof. Babak Falsafi
Co-examiner: Prof. Sanidhya Kashyap
Abstract
Modern workloads with memory footprints ranging from hundreds of gigabytes to several terabytes have become widespread over the last decade. With memory accounting for nearly 50% of server costs in modern data centres, provisioning large DRAM capacities significantly increases total cost of ownership (TCO). Tiered memory architectures combine memories with different capacities, latencies, and cost characteristics to improve cost efficiency while satisfying application performance requirements. Application performance in these architectures is highly sensitive to page placement decisions. Consequently, making informed placement decisions requires precise tracking of data access patterns and accurate identification of hot pages.
Existing operating systems commonly rely on hardware-maintained access bits in page table entries to estimate memory access behaviour and identify hot data regions. They periodically scan the application's entire virtual address space to construct workload access profiles. However, linear scanning of page table access bits does not scale for modern applications with terabyte-scale memory footprints. Since scanning overhead scales with the memory footprint, a single full scan can take several seconds for workloads with large memory footprints. Reducing scanning frequency lowers monitoring overhead but also decreases profiling accuracy and responsiveness to changing access patterns, resulting in a fundamental tradeoff between profiling accuracy and monitoring overhead. Therefore, we need scalable, low-overhead mechanisms for access monitoring in next-generation memory management systems.
In this talk, I will discuss recent approaches to scalable memory access profiling in large-memory systems. Specifically, I will present works that use techniques such as adaptive sampling, region-based monitoring, and hierarchical profiling of page table trees to reduce and bound scanning overhead.
Selected papers
coming soon
Exam president: Prof. Thomas Bourgeat
Thesis advisor: Prof. Babak Falsafi
Co-examiner: Prof. Sanidhya Kashyap
Abstract
Modern workloads with memory footprints ranging from hundreds of gigabytes to several terabytes have become widespread over the last decade. With memory accounting for nearly 50% of server costs in modern data centres, provisioning large DRAM capacities significantly increases total cost of ownership (TCO). Tiered memory architectures combine memories with different capacities, latencies, and cost characteristics to improve cost efficiency while satisfying application performance requirements. Application performance in these architectures is highly sensitive to page placement decisions. Consequently, making informed placement decisions requires precise tracking of data access patterns and accurate identification of hot pages.
Existing operating systems commonly rely on hardware-maintained access bits in page table entries to estimate memory access behaviour and identify hot data regions. They periodically scan the application's entire virtual address space to construct workload access profiles. However, linear scanning of page table access bits does not scale for modern applications with terabyte-scale memory footprints. Since scanning overhead scales with the memory footprint, a single full scan can take several seconds for workloads with large memory footprints. Reducing scanning frequency lowers monitoring overhead but also decreases profiling accuracy and responsiveness to changing access patterns, resulting in a fundamental tradeoff between profiling accuracy and monitoring overhead. Therefore, we need scalable, low-overhead mechanisms for access monitoring in next-generation memory management systems.
In this talk, I will discuss recent approaches to scalable memory access profiling in large-memory systems. Specifically, I will present works that use techniques such as adaptive sampling, region-based monitoring, and hierarchical profiling of page table trees to reduce and bound scanning overhead.
Selected papers
coming soon
Practical information
- General public
- Free