Performance Engineer

Aziro Technologies Llc

Santa Clara, CA, United States

about 1 month ago

Apply on www.dice.com

Prepare application

Role details

Contract type

Permanent contract

Employment type

Full-time (> 32 hours)

Working hours

Regular working hours

Job source

www.dice.com

Tech stack

Artificial Intelligence Amazon S3 Computer Clusters Software Debugging Distributed Systems InfiniBand Network File Systems Performance Tuning Remote Direct Memory Access Weka Rate Limiting

Job description

Set the performance architecture agenda. Bring deep, current expertise across file, block, and object storage protocols and translate it into concrete performance requirements for //EXA’s data path (NFSv3 direct-to-DN, pNFS layouts, S3 via MDN) informed by how HPC and hyperscale environments actually push storage systems (checkpointing, small-file metadata storms, GPU-starved read patterns, mixed-tenant burst I/O).
Track and act on the NeoCloud / sovereign-cloud shift. Maintain a living view of where //EXA’s highest-value deployments are heading GPU-cloud and sovereign-cloud operators (CoreWeave, Crusoe, Nscale, and similar) and make sure //EXA’s performance roadmap, reference architectures, and sizing guidance map to how these operators actually buy and operate infrastructure (multi-tenant GPU clusters, bursty training/inference mixes, strict SLAs to end customers).
Own competitive performance positioning. Build and maintain deep, technically substantiated comparisons against VAST Data, DDN, and WEKA not marketing bullet points, but real architectural analysis (metadata scaling model, erasure coding/durability tradeoffs, protocol support, GPU-direct paths, cost/performance at scale) that engineering and field teams can use to win technical evaluations and POCs.
Drive performance tuning for multi-tenant HPC/AI workloads. Lead tuning and validation work spanning the full stack a GPU cluster touches storage (MDN/DN geometry, pack groups, erasure coding layout), networking (RDMA, RoCE/InfiniBand fabric behavior, NIC/queue tuning), and compute (GPU-side I/O patterns, checkpoint/restore, data loader behavior) with particular focus on how these interact when multiple tenants/workloads share the same //EXA fleet.
Build and run the benchmark suite. Own //EXA’s benchmark framework and result credibility: MLPerf Storage (v2/v3), elbencho, IO500, fio/vdbench-class synthetic tests, and workload-representative benchmarks for AI training/inference and traditional HPC. Ensure results are reproducible, defensible in public disclosure, and directly comparable to published competitor numbers.
Define QoS, limits, and workload segmentation. Drive the technical requirements and validation for quality-of-service guarantees, per-tenant/per-workload throughput and IOPS limits, and workload isolation the mechanisms that let //EXA make hard SLA commitments in shared, multi-tenant NeoCloud deployments rather than best-effort performance.

Requirements

Deep, hands-on background in storage performance engineering across file, block, and object protocols, ideally with direct HPC or hyperscale exposure (parallel filesystems, pNFS/NFS at scale, S3-scale object stores).
Working knowledge of GPU cluster architecture RDMA fabrics, GPUDirect Storage, checkpoint/restore patterns for large model training and how storage bottlenecks manifest in mixed compute/network/storage systems.
Fluency with industry benchmark standards (MLPerf Storage, IO500) and load-generation tooling (elbencho, fio, vdbench), plus the judgment to design workload-representative tests beyond canned benchmarks.
Demonstrated ability to build rigorous, technically credible competitive analysis (not slideware) against systems like VAST, DDN, and WEKA architecture-level understanding, not just spec-sheet comparison.
Experience with multi-tenant resource management concepts (QoS, rate limiting, workload isolation) in a distributed systems context.
Comfortable operating across the stack and across audiences deep enough to debug an RDMA queue-pair stall or a metadata hot-partition, articulate enough to brief a NeoCloud customer’s technical evaluation team.