Minsung Jang 장민성

PAISE IEEE HPCA 2025 Peer-reviewed

Problem. Decoder-only LLM inference is memory-bound: attention re-reads a growing KV cache and stalls the GPU on HBM bandwidth.

Contribution. A per-layer GPU-vs-PIM offloading scheduler with data-layout adjustment and an interleaved-batched GEMM kernel on real HBM-PIM — up to 48.3% faster than GPU-only.

Role. Senior (last) author; set the direction and the offloading-scheduling algorithm and cost model.

Publication →

Omni-MP MCCSys @ ACM ICS 2026 · to appear Peer-reviewed

Problem. Real PIM accepts only fixed matrix tiles, so prior PIM-for-LLM work accelerated only feed-forward layers, not attention.

Contribution. Generalized batched PIM-GEMV mapping (interlacing/cutting), zero-tile skipping, and PIM-aware KV concatenation — 40.2% average end-to-end speedup on real hardware.

Role. Senior co-author and real-PIM program lead; set the generalization goal.

JABAS EuroSys 2025 · with UNIST Peer-reviewed

Problem. Training on heterogeneous GPU clusters wastes resources, and the usual remedy breaks the i.i.d. assumption and hurts convergence.

Contribution. IIDP preserves convergence via virtual stream workers while batching and GPU allocation are tuned jointly — −33.3% time, −54.2% cost, no accuracy loss.

Role. Industry co-author (collaboration with UNIST); supplied the production heterogeneous-GPU problem and real-cluster validation context — did not lead the algorithm design.

Publication → Code →

CORN IEEE IPCCC 2023 Peer-reviewed

Problem. RoCEv2 leans on switch-level PFC, which causes head-of-line blocking and PFC storms at the scale of AI GPU fabrics.

Contribution. End-host, RTT/BDP-driven congestion control with selective-ACK recovery makes RDMA run PFC-free and switch-transparent — zero PFC events in evaluation.

Role. Senior (last) author; set the cloud-RDMA direction and the RTT-based congestion-detection idea. Implemented as a hardware-NIC prototype in collaboration with AMD and demonstrated at SC25.

Publication →

DSDE IEEE BigData 2025 Peer-reviewed

Problem. A fixed speculation length in speculative decoding wastes compute and creates batch stragglers under diverse serving load.

Contribution. A training-free, KLD-variance-driven per-sequence speculation length with an adaptive cap, implemented in vLLM — robust in low-acceptance-rate regimes.

Role. Senior co-author and research lead; drove the adaptive speculation-length cap and the vLLM integration.

Paper →