| 摘要: | 隨著對高記憶體密度和快速記憶體讀寫性能的需求不斷增加，斯格明子賽道 式記憶體(Skyrmion Racetrack Memory,SKRM)正逐漸成為替代當前 DRAM 技術在 主記憶體應用中的主要競爭者。然而，斯格明子賽道式記憶體具有一些獨特的特 點，包括其能夠進行元素移位、生成和銷毀的能力，這些特點與傳統的 DRAM 有 很大的不同。如果不重新評估傳統的演算法和資料結構，以使其能夠適應基於斯 格明子(Skyrmion)的主記憶體的獨特性質，則系統性能可能會受到嚴重影響。在 各種演算法中，例如排序、搜索、圖形處理和機器學習等，交換操作都是基本的， 然而，直接應用於基於斯格明子的主記憶體時，會面臨到一些挑戰。傳統的方法 可能會導致高能源消耗，因為需要頻繁地生成斯格明子元素，或者會因經常進行 移位操作而導致系統性能下降。為了應對這些挑戰，我們的研究提出了一種斯格 明子友善的方法，引入了一種適合交換操作的賽道式記憶體架構，以同時降低能 耗和存取延遲。此外，我們將這一解決方案與基於決策樹的機器學習框架結合， 認識到在決策樹構建過程中，交換操作在頻繁執行中的重要作用。我們的實驗結 果令人振奮，結果顯示了能源消耗和存取延遲的減少。;With the demand of the high memory density and fast memory load/store performance, Skyrmion racetrack memory emerges as a leading contender to replace current DRAM technology in main memory applications. However, the distinctive features of skyrmion racetrack memory, including its ability to shift, generate, and destroy skyrmion elements, diverge significantly from traditional DRAM. Without reevaluating conventional algorithms and data structures to align with the unique properties of skyrmion-based main memory, system performance becomes poor. The swapping operation, fundamental in various algorithms like sorting, searching, graph, and machine learning, poses challenges when directly applied to skyrmion-based main memory. The conventional approach can lead to heightened power consumption from frequent skyrmion element creation or suboptimal system performance due to frequent shifting operations. To address this, our paper proposes a skyrmion-friendly approach, introducing a swapping friendly racetrack memory architecture to concurrently minimize energy consumption and access latency. Furthermore, we integrate this solution with a decision tree-based machine learning framework, recognizing the pivotal role of the swapping process in frequent execution during decision tree construction. Our experiments yield encouraging results, demonstrating reductions in energy consumption and access latency. |
