Characterizing LLM Kernel Access and Memory Interaction in Multi-Partition NUMA GPUs
Published in arXiv preprint arXiv:2607.28824, 2026
This work examines how large language model workloads interact with multi-partition GPU architectures, where non-uniform memory access characteristics and inter-partition communication can amplify contention and degrade performance. We develop a memory trace analysis methodology to study data access patterns at the workgroup level, categorizing kernel operands into three sharing patterns — global, partial, and private — and recommend placement-aware kernel programming and smarter architectural support for work and data locality.
Recommended citation: Donghyeon Joo, Sooraj Puthoor, Nuwan Jayasena, Bahar Asgari. (2026). "Characterizing LLM Kernel Access and Memory Interaction in Multi-Partition NUMA GPUs." arXiv preprint arXiv:2607.28824.
Download Paper