Safe GPU Memory Overcommit for CUDA Jobs: Isolation, Active Admission, and Idle Reclamation

GPU sharing is useful only when additional admitted work remains productive and memory limits remain enforceable. We present a black-box evaluation of Nodus, a GPU memory virtualization system targeting arbitrary CUDA jobs. In the measured allocation suite, Nodus enforces exact per-tenant caps across 8 CUDA allocation APIs; the open-source comparator HAMi-core permits cap leakage on 2. Reported overhead ranges from −0.5% to +0.3%, and a vLLM serving experiment measures 0.3% overhead with unchanged time-per-output-token p95 of 11.4 ms. On an H100 PCIe GPU, active-job overcommit admits 1.5–2× as many jobs as strict packing, with useful throughput at least 0.90 of solo and no out-of-memory errors in valid arms. Unmanaged overcommit instead records 0.028–0.029 useful throughput and 0/4 completed jobs. A separate integrated serving test frees approximately 60.5 GiB while idle, then reactivates the base and adapter engines in 0.528 s and 0.646 s with exact greedy-output agreement. These results establish measured cap enforcement, admission gains, and reversible reclamation, not universal compatibility, adversarial security, or compute-share guarantees. The evidence is single-node, synthetic, and based on a small number of runs.