본문 바로가기
  • Home

Multi-Level Data Classification Based on Soft Trace-Driven Labeling for Write Amplification Optimization in SSDs

  • Journal of Internet of Things and Convergence
  • Abbr : JKIOTS
  • 2026, 12(2), pp.87~94
  • Publisher : The Korea Internet of Things Society
  • Research Area : Engineering > Computer Science > Internet Information Processing
  • Received : March 4, 2026
  • Accepted : April 21, 2026
  • Published : April 30, 2026

Seungwoo Lee 1

1영남이공대학교

Accredited

ABSTRACT

This paper proposes a self-adaptive multi-level data classification scheme based on softmax regression to optimize the Write Amplification Factor (WAF) in NAND flash SSDs. Conventional binary classification methods fail to capture the temperature information of "warm" data situated between hot and cold boundaries, leading to block pollution and increased WAF. To address this, we design a mechanism that converts raw workload frequencies into continuous temperature scores using a sigmoid function, which are then expanded into 3D probability distributions termed "soft targets." The proposed model is trained by minimizing the cross-entropy loss between the predicted probability distribution and the soft target, enabling precise tracking of subtle data temperature transitions. Experimental results using MSR Cambridge Traces demonstrate an average of 16.5% additional WAF reduction compared to conventional binary classification. Furthermore, the scheme proves exceptional self-adaptability even under concept drift conditions. By implementing high-resolution intelligent data placement within limited hardware resources, this study presents a new paradigm for maximizing SSD lifespan and performance.

Citation status

* References for papers published after 2024 are currently being built.