To stride or not to stride the memory access?
Title
To stride or not stride the memory access?
Authors
Lennart Schmidt, Roland Kühn, Matti Krause, Jens Teubner, Wolfgang Lehner and Dirk Habich
Accepted at
3rd Workshop on Disruptive Memory Systems (DIMES), Seoul, Republic of Korea, October 2025
Abstract
Due to the increasing gap between CPU performance and memory bandwidth, memory access patterns play more and more a significant role for efficient data processing. The current core assumption is that a sequential access pattern delivers the best performance, especially when the data to be processed is stored in adjacent memory locations (contiguous memory). Given the continuous advances in memory technologies, it is of course questionable whether this assumption still holds true. To answer this question, we present a comprehensive experimental comparison of the sequential and the strided access pattern for data stored in contiguous memory on modern disruptive memory systems in this paper. As we are going to show, the core assumption must be revised, as the strided access pattern with a well-chosen stride size clearly outperforms the sequential access pattern. Even a SIMD-accelerated sequential access is considerably slower than the best-performing scalar strided access. In particular, we explain the differences, highlight further advantages, and present open challenges of the strided access pattern on disruptive memory systems.
Publication Log
September 2025
July 2025
Submission to DIMES 2025 (Accept)
- Submission
- Reviews (weak accept, weak accept, accept, weak accept)