Research-grade
scheduling, operated.
Slurm is the right scheduler for training work and the wrong thing for a research team to maintain. We stand it up on your dedicated cluster, tune it to how your people actually queue jobs, and keep it running.
- Scheduler
- Slurm
- Containers
- Pyxis / Enroot / Apptainer
- Accounting
- slurmdbd, fair-share
- Runs on
- Your dedicated cluster
- Upgrades
- Managed, tested first
What you get.
Cluster stood up and tuned
Controller and database nodes, high availability where the cluster warrants it, topology-aware scheduling so multi-node jobs land on the right rails, and prolog/epilog hooks that clean up after failed runs.
Partitions, QoS, and fair-share
Queue policy modelled on your team: reserved partitions for production runs, preemptible partitions for experiments, per-group fair-share so no single user starves the cluster.
Containers and environments
Pyxis and Enroot, or Apptainer, wired in so srun pulls a container without root. A curated set of base images for PyTorch, JAX, NCCL, and the CUDA track your cluster runs.
Job telemetry and accounting
Per-job GPU utilization, memory, and interconnect metrics in Grafana, plus slurmdbd accounting so you can see who used what and where throughput was left on the table.
Health checks and node draining
Automated node health checks catch a bad GPU or a flapping link and drain the node before it silently kills a multi-day run. Failed nodes are replaced from spares, not debugged in the queue.
The technical sheet.
Configured per deployment. These are the ranges we build within; your quote names the exact parts.
- Version track
- Current stable, tested before rollout
- Controller
- HA pair available
- Accounting DB
- slurmdbd with backups
- Topology
- Rail-aware placement
- Preemption
- Configured per partition policy
- Containers
- Pyxis / Enroot / Apptainer
- MPI
- OpenMPI, NCCL-tuned
- Modules
- Lmod or container-only, your call
- Shared FS
- Home, scratch, checkpoints
- Metrics
- Prometheus / Grafana
- GPU
- DCGM per-job utilization
- Logs
- Centralized, retained
- Reporting
- Monthly utilization summary
- Auth
- LDAP / OIDC / SSH keys
- Login nodes
- Dedicated, hardened
- Quotas
- Per user and per group
- Health checks
- Automated drain on failure
Where this fits.
Groups running many multi-node jobs where queue policy and fair-share matter more than a web console.
Organizations whose scientists would otherwise be maintaining the scheduler themselves, badly, at night.
Labs and enterprises already fluent in Slurm who want the same workflow on modern GPU hardware without running it.
Who carries the pager.
We do. Every item below is our responsibility for the life of the engagement.
- Scheduler, controller, and accounting database monitored and backed up 24/7.
- Version upgrades staged and tested before they touch your production queue.
- Queue policy reviewed with you as team size and workload mix change.
- Node health checks, automatic draining, and spare-node replacement.
- Direct escalation to the engineers who configured your cluster.
Tell us the shape of the cluster.
An engineer replies with a configuration and a deployment window. No pricing on the website, no sales sequence.
contact@crystalcloud.ai / Los Angeles, CA