Quiz: AI Cluster Orchestration & Service Delivery (Part 10)

10 questions from Part 10 of the guide. Score instantly; every explanation links to the chapter that derives the answer.

The core scheduling-plane fork for an AI fleet is…
cron vs systemdone vendor's cloud consoleSlurm (HPC gang scheduling) vs Kubernetes + batch scheduler vs a hybrid of bothmanual queues vs spreadsheets
When should a scheduler reserve an intact rack-scale NVLink domain for one job?
Only when the rack has no failed nodesWhen the job's parallelism and communication profile require the domain; otherwise compatible smaller jobs may backfill without violating isolation or performance SLOsWhenever the job asks for more CPU memoryFor every job, including single-GPU inference
The 'golden stack' discipline means…
letting each team pick their stackalways running latest versionsbuying gold-tier supporta pinned, attested and compatibility-tested driver+runtime+collective-library+firmware tuple within each scheduling pool
Why does the guide call raw GPU utilization a metric that 'lies'?
a GPU can show fully busy while stalled on communication or recomputing lost work — goodput measures what actually advanced the jobvendors misreport itit can't be measuredit excludes idle nodes
The two latency targets that bind an inference SLO are…
TTFT (prefill-bound) and TPOT/ITL (decode-bound)p50 and p50 againping and jitterDNS and TLS time
GPU sharing across tenant classes should follow which spectrum?
everyone shares everythingrandom assignmentno sharing, everwhole-GPU for tightly-coupled training → hardware-partitioned (MIG) → time-sliced, matched to coupling and isolation needs
The value-stack ladder an operator chooses a rung on is…
on-prem vs cloudbare-metal → GPUaaS → managed platform → serverless tokensCPU → GPU → ASICbronze, silver, gold
Bare-metal lifecycle at fleet scale should be driven…
by the vendor's field techsdeclaratively, as infrastructure-as-code — the fleet's desired state is versioned and convergedmanually per nodeby runbooks a human follows
The parallelism-framework standardization choice (FSDP2 / DeepSpeed / Megatron-Core) matters because…
it decides which parallelism forms are first-class and how your checkpoint formats and topology mapping workonly one runs on NVIDIAthey license differentlyit sets the Python version
Before admitting a training corpus, which evidence belongs in the decision record?
A self-assigned bronze, silver, or gold labelOnly whether the source URL was publicRights or license basis by source class, jurisdiction and intended use, lineage and transformations, consent/privacy controls, restrictions, takedown path, and counsel-approved dispositionOnly whether the data is text, image, or audio

Check your understanding: Software, Orchestration & Service Delivery