Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applications

GPUs lack fine-grained sharing primitives, so time-sharing and preemption between DL applications are expensive, and when one application cannot fill a GPU the device is wasted for everyone else.

Build fine-grained sharing from unmodified GPUs by interleaving kernels from concurrent applications at a granularity the stock scheduler already supports, with a monitor that decides when to interleave so each application still sees useful progress.

NOT RECORDED as a multiple — the arXiv abstract describes fine-grained GPU sharing primitives without a verified headline number carrying baseline and settings. Also note the scope mismatch: these are DL training/application-sharing results on V100-era GPUs, predating both MIG and LLM inference serving. Read the paper for the utilisation and fairness comparison against whole-GPU time sharing.

VENUE NOT VERIFIED — the arXiv comment is empty, USENIX site search returns no Salus paper, and ACM DL is antibot-blocked, so I record arxiv-preprint rather than guess EuroSys. Adoption 'abandoned' is a deliberate, evidence-based call: the software-sharing approach was superseded by hardware partitioning (NVIDIA MIG) and by CUDA MPS, which is the mechanism NVIDIA actually ships for GPU sharing today. Salus remains a well-cited systems paper whose premise the industry solved a different way. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.