Skip to content
Open access

S³-Agent: Causal and Contract-Checked Scheduling for Abstract GPU Partitioning

2026 · IEEE Access · Vol 14, pp. 166419-166436 · 0 citations · 31 references

Abstract

We evaluate S3-Agent as a causal, contract-checked interface for abstract GPU partitioning. Online policies choose among three layouts of a four-unit simulated resource using arrived jobs. A stateful executor checks allocation, drained reconfiguration, and completion, committing simulator state only after validation. Across 20 independent seeds and six workload cells, queue-fit scheduling reduced the mean profile-fit gap from 0.393 to 0.217. Its mean completion-time difference from a fixed layout was −0.89 simulated seconds (95% confidence interval −2.27 to 0.49), which does not resolve an improvement. A separate 12-seed sequential study connects diagnostic output to a later Qwen2.5-1.5B-Instruct planner, with optional read-only retrieval. The current CPU repeat made 188 model calls in 60 simulated runs and charged measured callback time as scheduler delay. The complete path averaged 29.88 simulated seconds versus 23.71 for queue-fit; its paired difference was + 6.17 s (95% paired interval + 2.90 to + 9.45). Diagnostic labels agreed with the declared queue-count reference in 18/62 calls. Earlier controlled and neutral-prompt diagnosis failures are retained. These studies do not establish the necessity of all three proposed modules. A separate simulation uses published CERN particle-job runtimes as service-time inputs. Safety results assume atomic, single-threaded simulated transitions; physical MIG scheduling and rollback remain unmeasured.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.