gpu-orchestration

Gpu Orchestration

A deep learning training job requires GPU resources that are expensive and scarce. The orchestration pipeline needs to check GPU availability, allocate the right GPU type for the model architecture, launch the training job, monitor utilization, and release the GPU immediately when training completes.

Pipeline

[gpu_check_availability]
     |
     v
[gpu_allocate]
     |
     v
[gpu_submit_job]
     |
     v
[gpu_collect_results]
     |
     v
[gpu_release]

Workflow inputs: jobId, gpuType, modelPath

Workers

GpuAllocateWorker (task: gpu_allocate)

• Uses randomization
• Writes gpuId, allocated, memoryGb

GpuCheckAvailabilityWorker (task: gpu_check_availability)

• Writes available, gpuType, cluster

GpuCollectResultsWorker (task: gpu_collect_results)

• Writes collected, artifacts

GpuReleaseWorker (task: gpu_release)

• Writes released

GpuSubmitJobWorker (task: gpu_submit_job)

• Writes outputPath, epochs, lossVal

---

20 tests | Workflow: gpu_orchestration_demo | Timeout: 60s

See RUNNING.md for setup and usage.