Publications
-
SC’26Low-Overhead Silent Error Detection for Large Language Model Inference at Scale[SC’26] International Conference for High Performance Computing, Networking, Storage, and Analysis · Accepted; to appear
Introduces TRACE, a selective activation-monitoring framework that detects silent data corruption across 15 large language models and 3 datasets with 95.3% average detection coverage and 0.56% runtime overhead.
-
SC’26Z-SPDK: Accelerating Lossy Compressed I/O on NVMe-SSDs[SC’26] International Conference for High Performance Computing, Networking, Storage, and Analysis · Accepted; to appear
Co-designs lossy compression and NVMe I/O using dual-pipeline parallelism and zero-copy transfers, delivering up to 12× speedup and 5.9× better bandwidth utilization.
-
SecDev’23Evaluating Container Debloaters[SecDev’23] 2023 IEEE Secure Development Conference
Introduces DebloatBench and evaluates container debloating tools across attack-surface reduction, CVE mitigation, image-size reduction, and execution correctness.
Manuscripts under review
- 2026
LLM Pretraining Reliability
Large-scale characterization of transient soft-error effects and resilience during Transformer pretraining.
- 2026
LLM Fault Resilience
Study of LLM behavior under increasing fault intensity, multi-bit corruption, and emerging high-error computing environments.
- 2026
HPC Reliability
Compiler-based, low-overhead detection of silent errors in scientific and HPC applications.