Learning Memory Access Patterns Using Machine Learning and Computer Vision

Abstract

Modern computing systems, from personal devices to powerful data centers, have witnessed a significant rise in processing power. This remarkable progress, however, creates a growing disparity between compute power and memory access speed, leading to performance bottlenecks. On a hardware level, this translates to wasted compute cycles while waiting for data retrieval. On a system level, handling more information necessitates more memory. Given the limitations of Dynamic Random Access Memory (DRAM), Non-Volatile Memory (NVM) is introduced, creating Hybrid Memory Solutions (HMS). Computer architecture utilizes prediction methods to mitigate the mentioned problems. At the CPU level, this involves data prefetching, where relevant data is preloaded in the cache to mask memory latency. On a system level, various page-related techniques (scheduling, migration, replacement, etc.) are employed to optimize memory management for HMS, ultimately improving overall performance. Currently implemented solutions in both cases are not too complex and focus on hardware or low-level software solutions. Evidently, accurate future page prediction could significantly enhance their performance. Fueled by the recent surge in machine learning, researchers are exploring novel applications in computer architecture, uncovering promising solutions.

Inspired by image-based solutions for financial time series forecasting, this thesis proposes a new approach that leverages similar image-based machine learning models to predict future page accesses. A complete pipeline is proposed that consists of a defined set of rules to visually represent the data and utilize image-based machine learning methods to predict future page accesses. This method aims to harness the strengths of both temporal and spatial information within the data. The research seeks to evaluate the effectiveness of this approach compared to traditional timeseries forecasting methods and current state-of-the-art LSTMs, and to explore the uncharted territory of predicting longer sequences of future page accesses. Although the proposed approach is promising, it is not intended to provide a directly implementable solution that surpasses or competes with existing hardware-based methods. Instead, it paves the way and lays the groundwork for further development of image-based machine learning techniques for page forecasting, potentially leading to significant performance improvements in both HMS and data prefetching.