Regular Paper Accepted at the Outstanding International Conference IEEE International Conference on Distributed Computing Systems (ICDCS) 2026
Files

The paper 'DUAL-BLADE: Dual-Path NVMe-Direct KV-Cache Offloading for Edge LLM Inference', written by Jung Bo-don (doctoral), Dr Byun Hong-su, Professor Kim Young-jae, Professor Weikuan Yu (FSU), Professor Yang Ji-hoon and Professor Park Sung-yong (corresponding author) of the Data-Centric AI Computing and Systems Laboratory (DISCOS), has been accepted for publication at the IEEE International Conference on Distributed Computing Systems (ICDCS) 2026 (acceptance rate: 18.59%).
When an LLM runs on a single-GPU edge AI system, the KV cache — which grows explosively with the length of the context to be processed — exceeds the device's available memory and creates a structural bottleneck. Existing systems have tried to resolve this by offloading to NVMe SSDs, but they rely excessively on the operating system's kernel page cache and so cause cache thrashing and severe software overhead.

Figure 1. High-level architecture of dual-path KV cache residency.
From this problem the research proposes DUAL-BLADE, a framework that operates in line with runtime memory availability. DUAL-BLADE dynamically splits KV tensors between a page cache path and an NVMe-Direct path. The NVMe-Direct path in particular bypasses the file system and maps data directly onto a contiguous logical block address (LBA) region, fundamentally eliminating unnecessary storage overhead. On top of this, the key point is that it maximizes overall data throughput by introducing an adaptive pipeline parallelization technique that overlaps storage I/O with GPU DMA operations.
Jung Bo-don, the paper's first author and a doctoral student, said of the significance of the research: "Accelerating LLM inference on memory-constrained edge systems requires breaking the indiscriminate reliance on the file system and overcoming I/O inefficiency. The dual-path design and kernel bypass techniques of this research will be a practical solution to that."
The IEEE International Conference on Distributed Computing Systems (ICDCS) is one of the flagship conferences of the IEEE Computer Society and a prestigious international conference in distributed computing systems. Now in its 46th edition, it will be held in Seoul, Korea, from 22 to 25 June this year, and peer-reviewed regular research papers will be presented in the conference's main technical sessions.