arXiv：cs.LG· Ehsan Yousefzadeh-Asl-Miandoab, B\"u\c{s}ra Karatay Demiray, Florina M. Ciorba, Pamela Delgado, P{\i}nar T\"oz\"un·· 6 小时前AI 评分37
AEGIS：面向多租户深度学习训练的运行时 GPU 共置调度系统
AEGIS: Runtime-Guided GPU Collocation for Multi-Tenant Deep Learning Training
AI 导读
AEGIS 是一个面向共享多 GPU 服务器的运行时调度系统，通过内存可行性判断、放置后观测与运行时压力过滤实现深度学习训练任务的受控共置。
正文
Abstract:Deep learning training commonly runs on shared multi-tenant GPU servers, where exclusive allocation provides isolation but can leave resources underutilized and increase queueing time. Collocation can improve efficiency, but interference-agnostic placement may cause severe slowdowns, while inaccurate memory information can lead to out-of-memory (OOM) failures.
We present AEGIS, a server-scale runtime scheduling system for controlled collocation of deep learning training workloads on shared multi-GPU servers. AEGIS integrates memory feasibility, post-placement observation, runtime-pressure filtering, placement, and OOM-aware recovery in a single scheduling loop. After placement, AEGIS observes workload activity before permitting further collocation, then uses low-overhead telemetry to determine whether a GPU can safely accept additional work. OOM failures trigger retries under progressively safer memory conditions, eventually falling back to exclusive execution. This online approach avoids costly offline pairwise compatibility profiling.
We evaluate AEGIS using vision, Transformer, recommendation, and LLM-style workloads across three production-derived traces. AEGIS reduces geometric-mean makespan by 16% relative to Lucid, 21% relative to Horus, and 27% relative to exclusive allocation. Sensitivity studies show that activity-anchored observation and runtime-pressure filtering balance conservative isolation against interference-agnostic collocation, improving makespan while limiting sharing-induced per-task slowdown.
| Subjects: | Distributed, Parallel, and Cluster Computing (cs.DC); Machine Learning (cs.LG); Performance (cs.PF) |
| Cite as: | arXiv:2508.19073 [cs.DC] |
| (or arXiv:2508.19073v4 [cs.DC] for this version) | |
| https://doi.org/10.48550/arXiv.2508.19073 arXiv-issued DOI via DataCite |
Submission history
From: Ehsan Yousefzadeh-Asl-Miandoab [view email]
[v1]
Tue, 26 Aug 2025 14:29:34 UTC (3,479 KB)
[v2]
Sat, 1 Nov 2025 16:13:11 UTC (4,293 KB)
[v3]
Mon, 23 Feb 2026 11:03:09 UTC (6,896 KB)
[v4]
Tue, 6 Oct 2026 14:30:27 UTC (7,308 KB)
来源：arXiv：cs.LG · arxiv.org