Staff Slurm Cluster & HPC Engineer

Bitdeer Technologies Group•San Jose, CA

13d•$180,000 - $260,000

About The Position

We are seeking a Staff Slurm Cluster & HPC Scheduling Engineer to own Slurm as a first-class, productized scheduling layer across that fleet. This person is the single technical owner of Slurm cluster architecture, multi-tenant scheduling policy, and cluster reliability on both bare-metal and VM-based GPU nodes, and will lead our adoption of the Slinky operator stack (slurm-operator, and slurm-bridge where it fits) so that Slurm and Kubernetes workloads can share the same GPU pool. The role is deeply hands-on, customer-facing during onboarding and escalations, and sets the engineering standard the rest of the platform team builds on.

Requirements

8+ years in HPC, systems, or cloud infrastructure engineering, including 4+ years operating production Slurm clusters at 100+ GPU-node scale with real users and service-level commitments.
Deep hands-on Slurm expertise: slurm.conf, gres.conf, topology.conf, cgroup.conf, partitions/QOS/fairshare/preemption/reservations, slurmdbd accounting, slurmrestd, MUNGE/SACK and JWT authentication, and version upgrades performed on live clusters.
Strong GPU and fabric fundamentals: NVIDIA drivers and Fabric Manager, DCGM, MIG, InfiniBand/RoCEv2 (subnet manager/UFM, rail-optimized topology), GPUDirect RDMA, and practical NCCL tuning and failure diagnosis.
Production Kubernetes experience and working knowledge of the operator/CRD pattern, plus hands-on exposure to at least one Slurm-on-Kubernetes stack — Slinky slurm-operator or slurm-bridge, CoreWeave SUNK, or Nebius Soperator — with an informed view of the tradeoffs between them.
Experience delivering both bare-metal and virtualized compute: bare-metal provisioning and firmware/BIOS lifecycle management, hypervisor or VM-based clusters (KVM/QEMU or a public-cloud equivalent), and Terraform/Ansible-driven automation.
Working knowledge of parallel and shared storage for AI workloads — Lustre, GPFS/Spectrum Scale, WEKA, VAST, or NFS — and of how storage behavior shapes job performance and failure modes.
Proficient in Python and Bash for cluster automation; Go experience is a plus for integrating with Bitdeer AI's platform control plane and with Slurm/Slinky REST client code.
Multi-tenant security discipline: derives tenant scope from a verified identity rather than client-supplied fields, designs authorization to fail closed, and treats isolation across accounts, namespaces, storage, and networks as a hard requirement.
Clear written and verbal communication in English, with the maturity to work directly with enterprise customers and to translate scheduling and reliability tradeoffs for product, sales, and executive stakeholders

Responsibilities

Slurm cluster architecture and lifecycle — Design, deploy, and operate production Slurm clusters on bare metal and VMs: slurmctld/slurmdbd high availability, slurmrestd, configless slurmd, SACK/MUNGE and JWT authentication, and rolling version upgrades on live clusters without losing running jobs.
Topology-aware scheduling for GPU fabrics — Model the physical fabric in topology.conf — topology/tree for rail-optimized InfiniBand/RoCE designs and topology/block for NVLink domains such as GB200/GB300 NVL72 — and prove placement quality with NCCL bandwidth and multi-node training validation rather than assumption.
Multi-tenant scheduling policy — Own the account/association tree, partitions, QOS, fairshare, preemption, reservations, and per-tenant TRES limits. Enforce fail-closed defaults: an unresolved tenant identity or an empty entitlement set must deny, never degrade into unrestricted access.
Slinky on Kubernetes — Lead implementation of the Slinky slurm-operator, including its NodeSet, LoginSet, Accounting, RestAPI, and Token custom resources, cert-manager and Helm-based delivery, shared parallel-storage mounts, and login pods running sackd/sshd. Evaluate and pilot slurm-bridge for co-scheduling Kubernetes Pods, PodGroups, Jobs, JobSets, and LeaderWorkerSets through the Slurm scheduler, and document its constraints — notably exclusive whole-node allocation — before any customer exposure.
Elastic capacity between Slurm and Kubernetes — Use Slurm cloud and power-save mechanisms (ResumeProgram/SuspendProgram, SuspendTime, ResumeTimeout) together with fleet automation to shift GPU nodes between batch training queues and Kubernetes inference capacity as demand moves.
Container and job runtime — Operate Pyxis/Enroot and OCI/containerd job paths with correct gres.conf, cgroup v2 device constraints, and CUDA_VISIBLE_DEVICES behavior; support MPI/PMIx, module/Spack environments, and customer-supplied images.
Cluster health and reliability engineering — Build the passive and active health-check system expected of a top-tier GPU cloud: prolog/epilog checks, LBNL NHC or equivalent, DCGM diagnostics, and detection of XID/SXID errors, ECC faults, PCIe errors, GPUs falling off the bus, IB/RoCE link flaps, and NCCL stalls — with automatic drain and job requeue. Own burn-in and acceptance testing for every new rack before it carries paid work.
Automation and infrastructure as code — Deliver clusters through Terraform/Ansible, golden images, and bare-metal provisioning (PXE, Redfish, IPMI) so that a cluster build is reproducible, reviewable, and auditable rather than hand-tuned.
Observability, accounting, and billing integration — Instrument queue wait time, allocation efficiency, GPU utilization, and job failure taxonomy through a Slurm exporter into Prometheus/Grafana; configure AccountingStorageTRES and TRESBillingWeights, and reconcile sacct/sreport GPU-hours against the platform's metering and invoicing pipeline.
Technical leadership and customer engagement — Write runbooks and tenant-facing documentation, onboard and support enterprise customers, act as escalation point for cluster incidents, and mentor platform engineers on Slurm and HPC scheduling practice.