arXiv subjects

Sunwoong Kim

Publications and source records attributed to Sunwoong Kim.

3 recordsLinked to original sources

arXiv2026

ITME: Inference Tiered Memory Expansion with Disaggregated CXL-Hybrid Memories

The rapid shift toward agentic and long-context workloads in Large Language Models (LLMs) is pushing the industry beyond the capacity of individual servers toward disaggregated shared storage to handle TB-scale context states. This movement has led to the emergence of specialized shared context layers designed to externalize and share cumulative inference states across distributed clusters. While offloading to a data processing unit (DPU) within just-a-bunch-of-flash (JBOF) architectures accelerates NVMe-over-fabrics (NVMe-oF) target processing, the need for sophisticated software-level optimization and cost-efficiency burdens remain significant. Consequently, the ideal architecture for scaling this shared context infrastructure is still an active area of exploration. In this paper, we propose ITME (Inference Tiered Memory Expansion), which leverages a CXL-hybrid memory to present a massive, TB-scale byte-addressable remote memory expansion. This approach enables cost-efficient scaling and simplifies the software stack through direct byte-addressability, effectively addressing the challenges of shared context infrastructure. Our key insight is that the deterministic access patterns of voluminous model weights and prefix caches enable the system to proactively manage data movement across the memory-storage hierarchy. We validate ITME by evaluating its performance potential with production-grade SK Hynix CMM and PCIe Gen5 NVMe SSDs, while further demonstrating its functional feasibility through an FPGA-based hardware prototype. Overall, ITME enhances conventional CPU-offloading by providing additional remote memory expansion to accommodate large KV cache footprints beyond host memory limits, achieving up to a 35.7\% throughput improvement.

cs.DC↗

arXiv2025

Leveraging FPGAs for Homomorphic Matrix-Vector Multiplication in Oblivious Message Retrieval

While end-to-end encryption protects the content of messages, it does not secure metadata, which exposes sender and receiver information through traffic analysis. A plausible approach to protecting this metadata is to have senders post encrypted messages on a public bulletin board and receivers scan it for relevant messages. Oblivious message retrieval (OMR) leverages homomorphic encryption (HE) to improve user experience in this solution by delegating the scan to a resource-rich server while preserving privacy. A key process in OMR is the homomorphic detection of pertinent messages for the receiver from the bulletin board. It relies on a specialized matrix-vector multiplication algorithm, which involves extensive multiplications between ciphertext vectors and plaintext matrices, as well as homomorphic rotations. The computationally intensive nature of this process limits the practicality of OMR. To address this challenge, this paper proposes a hardware architecture to accelerate the matrix-vector multiplication algorithm. The building homomorphic operators in this algorithm are implemented using high-level synthesis, with design parameters for different parallelism levels. These operators are then deployed on a field-programmable gate array platform using an efficient design space exploration strategy to accelerate homomorphic matrix-vector multiplication. Compared to a software implementation, the proposed hardware accelerator achieves a 13.86x speedup.

cs.CR↗

arXiv2024

LLM-Mirror: A Generated-Persona Approach for Survey Pre-Testing

Surveys are widely used in social sciences to understand human behavior, but their implementation often involves iterative adjustments that demand significant effort and resources. To this end, researchers have increasingly turned to large language models (LLMs) to simulate human behavior. While existing studies have focused on distributional similarities, individual-level comparisons remain underexplored. Building upon prior work, we investigate whether providing LLMs with respondents' prior information can replicate both statistical distributions and individual decision-making patterns using Partial Least Squares Structural Equation Modeling (PLS-SEM), a well-established causal analysis method. We also introduce the concept of the LLM-Mirror, user personas generated by supplying respondent-specific information to the LLM. By comparing responses generated by the LLM-Mirror with actual individual survey responses, we assess its effectiveness in replicating individual-level outcomes. Our findings show that: (1) PLS-SEM analysis shows LLM-generated responses align with human responses, (2) LLMs, when provided with respondent-specific information, are capable of reproducing individual human responses, and (3) LLM-Mirror responses closely follow human responses at the individual level. These findings highlight the potential of LLMs as a complementary tool for pre-testing surveys and optimizing research design.

cs.CY↗