
VAST Data Brings Job-Level QoS to Storage for HPC and AI
VAST Data will ship Job-Level QoS in VAST AI OS 5.6 by early February 2027, letting Slurm, PBS Pro and LFS schedulers enforce per-job bandwidth and IOPS policies on shared storage.
- By
- Grace Kim
- Filed
- Channel
- Memory & Storage
- Read
- 4 min read
VAST Data has made bandwidth and IOPS controllable at the individual job level, so that long-running compute jobs no longer slow down when other jobs monopolize shared storage. The feature, called Job-Level QoS, will reach general availability with VAST AI OS 5.6 in late January or early February 2027. It requires VAST NFS 4.8 or later.
The move targets a chronic pain point in large-scale computing. A single I/O-intensive job can degrade the performance of dozens of others running on the same system, and administrators have historically had few ways to intervene without disrupting work in progress. On the world's largest computing systems, a critical simulation can be delayed by an unrelated analysis workload even when ample compute capacity sits idle. For research institutions and AI infrastructure providers, that competition produces unpredictable performance, operational disruption and inefficient use of expensive infrastructure.
Glenn Lockwood, Principal Technical Strategist at VAST Data, described the problem in direct terms. "Anyone who's used a supercomputer knows the frustration of unpredictable performance when another job consumes more than its share of shared resources," he said. "The scheduler knows exactly which jobs are running, but that understanding has historically stopped at the storage system. With Job-Level QoS, we're closing that gap. The storage system itself can now manage the I/O performance it delivers to individual jobs competing for resources, allowing resource managers and schedulers to prioritize bandwidth and IOPS without having to cancel jobs or change how applications run."
How does job identity reach the storage layer?
Job-Level QoS extends the job identity defined by workload managers such as Slurm, PBS Pro and LSF into the VAST data layer. Once the storage system knows which job issued which I/O request, organizations can treat I/O performance as an allocatable resource alongside compute. Administrators can enforce individual bandwidth and IOPS policies on a job-by-job basis and adjust allocations in real time while workloads continue running.
The design principle is simple: the scheduler already holds the authoritative map of running jobs, and the storage array already controls the physical bandwidth. VAST connects the two, so a queue position or priority level set in the workload manager translates directly into a performance envelope enforced by the storage system rather than left to contention.
This closes a gap that has persisted across generations of parallel file systems. Protecting priority workloads no longer requires cancelling jobs, rewriting applications or changing how code interacts with storage.
What controls ship in VAST AI OS?
VAST AI OS gains several new mechanisms to enable job-level enforcement.
- Enforce QoS at the individual job level. The VAST NFS client attaches configurable job identifiers to I/O requests. VAST can then apply independent bandwidth and IOPS policies to multiple jobs sharing a single compute node, or manage the collective I/O of one job spanning multiple nodes under a single policy.
- Adjust performance without interruption. Administrators can change read and write performance policies in real time, with no job restarts, no application modifications and no storage remounts.
- See which jobs are consuming resources. An optional, open-source component called the VNFS Collector exposes job-level telemetry for bandwidth, IOPS and errors, queryable through Prometheus, Grafana or the VAST DataBase.
- Extend controls beyond batch. The underlying architecture also supports Kubernetes and other orchestrators, enabling workload-level performance management across AI training, inference and scientific computing environments.
The last point matters commercially. Batch schedulers dominate traditional HPC, but AI infrastructure increasingly runs on Kubernetes, where training and inference workloads compete for the same storage pools. Building the mechanism at the architecture level rather than as a Slurm-specific hook positions the feature for both markets.
Real-time policy adjustment is the operationally significant piece. An administrator watching a priority simulation fall behind can raise its bandwidth allocation and throttle the offending analysis job on the fly, without touching either workload. That converts storage performance from a fixed contention problem into a managed, schedulable resource.
When can customers deploy it?
VAST will demonstrate Job-Level QoS in action at SC26, showing how administrators identify competing jobs, dynamically adjust bandwidth allocations and enforce independent performance policies across shared infrastructure while workloads continue running uninterrupted.
General availability is planned with VAST AI OS 5.6 in late January or early February 2027. Deployment requires VAST NFS 4.8 or later on the client side, which sets the upgrade path for existing installations.
The competitive stakes extend beyond HPC centers. VAST has positioned its platform as the storage layer for AI factories, where GPU clusters spend tens of millions of dollars and any I/O bottleneck idles that capital. If job-level QoS works as described, VAST gives operators a lever that competing storage platforms currently lack, and it sets a benchmark that rival all-flash and parallel file system vendors will likely have to answer as AI infrastructure spending accelerates through 2027.
Original: slurm.schedmd.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,116 articles
Related articles
netapp-s-novus-targets-100-tb-s-storage-for-ai-factories-1a4b1ef0
NetApp's Novus Targets 100 TB/s Storage for AI Factories
power-not-accelerators-now-caps-data-center-ai-scaling-6bd78ff3
Power, Not Accelerators, Now Caps Data Center AI Scaling
ai-inference-turns-the-memory-wall-into-a-system-problem
AI Inference Turns the Memory Wall Into a System Problem
fractile-ceo-ai-bottleneck-is-memory-bandwidth-not-compute-d458a481
Fractile CEO: AI Bottleneck Is Memory Bandwidth, Not Compute



