Use layered evidence to identify whether a seeded regression is compute, placement, EFA, storage, scheduler, or data-staging related, then apply the narrow remediation.
1.Runtime regressed and MPI all-reduce time doubled while compute time is unchanged. First next step?
2.Jobs remain pending with reason `Resources`. First next step?
3.Thousands of tiny files make staging slow while bulk throughput is low. Likely bottleneck?
4.Single-node runtime regressed after an image update. Best action?
5.EFA interface exists but `fi_info -p efa` returns no provider. Likely issue?
6.I/O benchmark scales, but the application reads through rank 0 only. Best remediation?
7.Spot job restarts from the beginning after interruption. Missing capability?
8.CPU utilization is low while memory bandwidth is saturated. Best response?
9.Queue wait increased because every job requests 48 hours but usually runs 30 minutes. Remediation?
10.A benchmark changed instance type, MPI version, compiler, and data simultaneously. What is wrong?
- Choose the next diagnostic signal rather than guessing.
- Separate communication, I/O, scheduling, and compute symptoms.
- Select the smallest evidence-based remediation.
- Score at least 80%.