Free NVIDIA NCA-AIIO Actual Exam Questions - Question 2 Discussion
being trained simultaneously on a shared GPU cluster. Some models require more GPU resources and
longer training times than others. Which orchestration strategy would best ensure that all models are
trained efficiently without causing delays for high-priority workloads?
Makes sense that without preemption, priority scheduling might not fully prevent delays. Still, random assignment (C) is clearly inefficient, so A remains the best option here by at least trying to allocate more resources to important jobs. A
A vs D? Giving equal resources (D) ignores different model needs, so some high-priority models might get delayed. A at least tries to match resources with priority, which should reduce overall wait times.
Option B avoids complexity and ensures no job starves, even if slower.
B, because FCFS at least ensures fairness and predictable order without complex resource juggling.
Option A seems best because it directly addresses the need to prioritize critical workloads without wasting resources on less important jobs. The others either ignore priorities or waste resources by treating all jobs the same.
A makes the most sense here. Prioritizing high-priority models with more GPUs prevents delays and keeps things efficient. Options B, C, and D don’t really handle differing needs well.