{"resourceId":"raser-slurm-agentic-workstealing-makespan-20260903","versions":[{"version":"external-a54edf6d5f648b68aa6cfc7cc28e23bb9ce270f499c6d1a3085a8b179e6eb1c2","resource":{"id":"raser-slurm-agentic-workstealing-makespan-20260903","title":"RASER user-space Slurm agent runtime cuts makespan ~39% vs static partitioning","organization":"TU Dresden / TU Chemnitz collaborators (arXiv preprint)","sector":"University research computing / HPC agent runtimes","geography":"Germany (author institutions); transferable to Slurm campuses with limits","publishedAt":"September 3, 2026 (arXiv v1)","publicationDate":"2026-09-03","eventDate":null,"sourceName":"arXiv","sourceLabel":"Academic preprint — RASER HPC agent scheduling runtime","sourceUrl":"https://arxiv.org/html/2609.03598","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["infrastructure","developers-agents","operating-model"],"finding":"RASER extends Slurm with agentic job arrays, filesystem work-stealing queues, application-level checkpointing plus requeue, and Apptainer isolation without image changes. Authors report nearly 39% makespan reduction versus static partitioning (38.8% in the detailed scenario) with near-full CPU utilization and estimated NVMe checkpoint/restore under 30 seconds, exercising resilience with injected preemption.","sledRelevance":"Why now: Prior research editions covered capacity awards and validation-gated agents; this still-unarchived Sep 2026 runtime paper supplies measured HPC scheduling benefits/constraints for deploying agentic scientific workflows on existing campus clusters.","evidence":"Comparative makespan vs static partitioning; CPU utilization claims; checkpoint overhead estimate; fault injection via scancel SIGTERM; user-space/no-kernel-privilege design.","architectureImplications":"Campus agent platforms should assume Slurm-native user-space orchestration with shared-FS queues and container isolation rather than bespoke privileged schedulers.","governanceImplications":"HPC centers must define who may run immortal agent arrays and how preemption/accounting apply.","securityPrivacyImplications":"Apptainer bind mounts with direct hardware access require project isolation and mount review.","caveats":"Scheduling metrics ≠ scientific outcomes; authors flag FS lock contention at higher concurrency; site Apptainer/Slurm assumptions vary.","streamIds":["research"],"roles":{"sales":"Sales framing: Campus research computing teams need to run long-horizon LLM agents on existing Slurm clusters without waiting for admin-privileged schedulers. RASER evidence supports a bounded engagement around user-space agent arrays, work stealing, and checkpoint/requeue—with measured ~39% makespan reduction versus static partitioning in the authors' evaluation—not a claim of scientific discovery gains. Ask whether Apptainer and shared filesystems already exist and whether preemption is routine. Unsupported: guaranteed utilization at extreme concurrency or kernel-free fit for every site.","engineering":"Engineering framing: Fit is a user-space layer on Slurm job arrays with filesystem queues, application-level state serialization plus requeue, and Apptainer bind-mounted isolation. Integration points are existing Slurm accounting and shared scratch; no external database claimed. Prerequisites: Apptainer, writable shared FS, agent state serialization hooks. Constraints: filesystem lock contention at high concurrency (authors' future work), site-specific preemption policies, and evaluation not covering scientific correctness. Security relies on container isolation while retaining direct hardware access—review campus container policy.","delivery":"Delivery framing: Pilot RASER-like agent arrays on a non-production partition with fault injection (preemption signals), measure makespan and CPU utilization against static partitioning, and document checkpoint RTO. Dependencies: HPC center cooperation and Apptainer support. Ownership: research computing owns runtime; research groups own agent logic. Governance checkpoints: preemption drills and queue-contention review. Acceptance: documented makespan improvement and restore under agreed seconds; risks include FS hotspots and confusing scheduling gains with science outcomes."},"roleDetails":{"sales":{"customerProblem":"Agentic research workflows fit poorly on static Slurm partitions, wasting capacity and failing under preemption.","stakeholders":["HPC center leadership","Research computing architects","Faculty running agent workflows","Security/container owners"],"discoveryQuestions":["Is Apptainer approved?","How often are jobs preempted?","What shared FS backs queues?","What static array inefficiencies exist today?"],"valueHypothesis":"User-space work-stealing agent arrays with checkpoint/requeue can cut makespan ~39% vs static partitioning when CPU utilization is the bottleneck—without claiming discovery quality gains.","potentialEngagement":"Non-production partition pilot comparing static arrays vs work-stealing agent arrays with fault injection.","unsupportedClaims":["Scientific accuracy improvements","Guaranteed performance under extreme FS lock contention","Works without Apptainer/Slurm"]},"engineering":{"fit":"User-space Slurm extension for dynamic agentic jobs with work stealing and checkpoint/requeue.","architectureIntegration":"Wrap agents in Apptainer; use shared FS queues; serialize agent state; requeue via Slurm primitives.","prerequisites":["Slurm","Apptainer","Shared filesystem","Agent state serialization API"],"constraints":["FS lock contention at high concurrency","Site preemption policies","Not a science-result validator"],"security":"Review container bind mounts and direct hardware access; enforce project isolation on shared queues.","proofOfValue":["Makespan vs static partitioning","CPU utilization","Checkpoint restore time","Survival under injected preemption"]},"delivery":{"work":"Deploy pilot plugins, run comparative workloads, inject preemption, and publish ops runbooks.","dependencies":["HPC center change window","Apptainer support","Research group agent workloads"],"ownership":"Research computing owns the runtime; PIs own agent application logic.","skillsAdoption":"Train users to submit agent arrays and recover from requeue rather than one-shot jobs.","governanceCheckpoints":["Preemption drill","FS contention review","Container policy check"],"acceptanceCriteria":["Documented makespan improvement","Restore within agreed seconds","Written non-goal: scientific correctness"],"risks":["Shared FS hotspots","Misread scheduling metrics as science progress","Container policy rejection"]}},"retrievedAt":"2026-09-19T03:12:25Z","enrichedAt":"2026-09-19T03:12:25Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"User-space design aims to democratize agentic HPC without admin privileges, but still needs Apptainer-fluent support staff.","procurementImplications":"Prefer solutions that integrate with existing Slurm/Apptainer estates over net-new privileged schedulers.","operatingModelImplications":"Moves agent workflow ops into research-computing runtimes while researchers retain agent logic ownership.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2609.03598","referenceExcerpt":"RASER reduces makespan by nearly 39% compared to static partitioning","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}