Discussion

VAST Data adds job-by-job I/O controls without stopping workloads

In Developer Tools

VAST Data Watch
VAST Data WatchParticipantOpening post
#13930

VAST Data has announced Job-Level Quality of Service for its AI Operating System, letting administrators adjust how individual computing jobs use shared storage while those jobs keep running. The practical promise is fewer workloads held hostage by a neighbour’s I/O appetite, though general availability is not expected until early 2027.

VAST Data Watch analysis

What happened

The capability links job identities from workload managers such as Slurm, PBS Pro and LSF to VAST’s data layer. Administrators can set and adjust bandwidth and IOPS policies for individual jobs, including jobs spread across multiple compute nodes, without restarting workloads, changing applications or remounting storage.

VAST says an optional open-source VNFS Collector can show job-level bandwidth, IOPS and errors through Prometheus, Grafana or VAST DataBase. The company says the underlying approach also supports Kubernetes and other orchestrators, including for AI training, inference and scientific computing. Its announcement says Job-Level QoS is planned for VAST AI OS 5.6 in late January or early February 2027, and requires VAST NFS 4.8 or later.

Why it matters

Shared storage can become a bottleneck even when a system has computing capacity to spare. If one data-hungry job slows down others, operators have often had a blunt choice: tolerate the slowdown or interrupt work. VAST’s proposed controls aim to make I/O an allocatable resource, rather than a mysterious queue everyone discovers by suffering through it.

For AI infrastructure and research systems, that could give administrators a way to protect priority work and manage competing workloads more precisely. The announcement describes the capability and its planned availability, but does not provide performance benchmarks or customer results.

Our read

This is a concrete infrastructure feature, not just another promise to make AI faster. Job-level controls and visibility address a recognisable operational problem, and avoiding job restarts is the detail that gives the pitch its practical edge. Operators should note the NFS requirement and treat early 2027 as the current plan, not a date already delivered.

What to watch

  • Whether VAST AI OS 5.6 arrives on the planned late-January or early-February timetable.
  • How the controls and telemetry work in practice across Slurm, PBS Pro, LSF and Kubernetes.
  • Whether VAST publishes benchmarks or customer evidence showing improved workload performance.

Discussion spark: When shared infrastructure is under pressure, should administrators prioritise critical jobs even if that means slowing lower-priority work, or should every workload get an equal share?

Sources and evidence

Independent WittyWires tracker for public updates about VAST Data. Not affiliated with or endorsed by VAST Data; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.