Putting the Context back into Memory

Fuente: arXiv
Saved in:
Bibliographic Details
Main Author: Roberts, David A.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912662991929344
author Roberts, David A.
author_facet Roberts, David A.
contents Requests arriving at main memory are often different from what programmers can observe or estimate by using CPU-based monitoring. Hardware cache prefetching, memory request scheduling and interleaving cause a loss of observability that limits potential data movement and tiering optimizations. In response, memory-side telemetry hardware like page access heat map units (HMU) and page prefetchers were proposed to inform Operating Systems with accurate usage data. However, it is still hard to map memory activity to software program functions and objects because of the decoupled nature of host processors and memory devices. Valuable program context is stripped out from the memory bus, leaving only commands, addresses and data. Programmers have expert knowledge of future data accesses, priorities, and access to processor state, which could be useful hints for runtime memory device optimization. This paper makes context visible at memory devices by encoding any user-visible state as detectable packets in the memory read address stream, in a nondestructive manner without significant capacity overhead, drivers or special access privileges. We prototyped an end-to-end system with metadata injection that can be reliably detected and decoded from a memory address trace, either by a host processor, or a memory module. We illustrate a use case with precise code execution markers and object address range tracking. In the future, real time metadata decoding with near-memory computing (NMC) could provide customized telemetry and statistics to users, or act on application hints to perform functions like prioritizing requests, remapping data and reconfiguring devices.
format Preprint
id arxiv_https___arxiv_org_abs_2510_15878
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Putting the Context back into Memory
Roberts, David A.
Hardware Architecture
Operating Systems
Performance
Requests arriving at main memory are often different from what programmers can observe or estimate by using CPU-based monitoring. Hardware cache prefetching, memory request scheduling and interleaving cause a loss of observability that limits potential data movement and tiering optimizations. In response, memory-side telemetry hardware like page access heat map units (HMU) and page prefetchers were proposed to inform Operating Systems with accurate usage data. However, it is still hard to map memory activity to software program functions and objects because of the decoupled nature of host processors and memory devices. Valuable program context is stripped out from the memory bus, leaving only commands, addresses and data. Programmers have expert knowledge of future data accesses, priorities, and access to processor state, which could be useful hints for runtime memory device optimization. This paper makes context visible at memory devices by encoding any user-visible state as detectable packets in the memory read address stream, in a nondestructive manner without significant capacity overhead, drivers or special access privileges. We prototyped an end-to-end system with metadata injection that can be reliably detected and decoded from a memory address trace, either by a host processor, or a memory module. We illustrate a use case with precise code execution markers and object address range tracking. In the future, real time metadata decoding with near-memory computing (NMC) could provide customized telemetry and statistics to users, or act on application hints to perform functions like prioritizing requests, remapping data and reconfiguring devices.
title Putting the Context back into Memory
topic Hardware Architecture
Operating Systems
Performance
url https://arxiv.org/abs/2510.15878