Senior Compute Platform Engineer, LSF

NVIDIA•Austin, TX

1d•$184,000 - $356,500•Hybrid

About The Position

NVIDIA's silicon does not tape out without the farm behind it. Our EDA compute environment runs millions of cores across federated LSF cells, and every simulation, synthesis run, and timing signoff on our roadmap passes through it. We are consolidating a dual-scheduler estate onto a single LSF platform, and we are looking for an engineer who knows LSF at the level of its internals — not just its configuration files. This is a deep-specialist role. You will be the person the team escalates to when scheduling latency creeps up and nothing in the logs explains why.

Requirements

BS or MS in Computer Science, Computer Engineering, or equivalent experience.
8+ years in HPC or large-scale batch compute, with 5+ years of that on IBM Spectrum LSF.
Demonstrated depth in LSF internals — you have debugged scheduler behavior beyond what the documentation covers, and you can explain a scheduling cycle from submission to dispatch.
Hands-on MultiCluster experience in a production, multi-site environment.
Strong Linux systems fundamentals, system programming languages and scripting in Python, Perl, and shell.

Nice To Haves

You have worked on LSF as a developer or in escalation engineering, rather than only as a consumer of it.
Experience with LSF integration points: esub, eexec, elim, submit wrappers, RTM, or the LSF APIs.
Background in semiconductor or EDA compute, where license constraints and job constraints compete.
You have migrated a production estate off Slurm, PBS, or Grid Engine without a scheduled outage users noticed.

Responsibilities

Owning scheduler behavior across 15–25 federated LSF cells, including mbatchd and mbschd tuning, scheduling cycle analysis, and the contention patterns that appear as cells approach host-count ceilings.
Diagnosing MultiCluster forwarding problems — remote queue sizing, forwarding policy, cross-cluster pend behavior — where the symptom reported by users is "the farm is slow" rather than a clear failure.
Setting the technical design for cell topology and federation as the farm grows, and deciding what belongs in a cell versus what belongs in a new one.
Working with our IaC engineer to encode scheduler policy into a config schema that survives contact with MultiCluster, rather than one that looks clean and breaks at scale.
Partnering directly with CAD and methodology teams on workloads that break normal assumptions: 500GB+ memory jobs, interactive-versus-batch contention, and tape-out crunch bursts.