Computer Architecture · All levels
Performance Analysis Tricky Q&A
20+ senior Performance Analysis interview questions.
Q&A bank
Answer with mechanism, pitfall, validation check, and product judgment. At 10+ years, a correct definition is not enough.
Senior answer rubric
Start with the failing metric and analysis context.
Explain the microarchitectural or system mechanism.
Name the cheapest evidence-gathering experiment.
Choose a bounded fix and state what it can regress.
Close with product tradeoff, validation, or escalation criteria.
What makes a performance counter trustworthy for architecture decisions?
diagram
[INT][ARCH]
Q: What makes a performance counter trustworthy for architecture decisions?
A:
Stable semantics, validated collection path, and repeatable correlation with independent traces.
FOLLOW-UP TRAP: Using counters without validating overflow, aliasing, or collection window bias.How do you separate primary from secondary bottlenecks?
diagram
[INT][ARCH]
Q: How do you separate primary from secondary bottlenecks?
A:
Use constrained-resource evidence and falsification experiments to isolate the limiter with highest global impact.
FOLLOW-UP TRAP: Optimizing whichever block reports the largest local utilization.When does roofline analysis mislead teams?
diagram
[INT][ARCH]
Q: When does roofline analysis mislead teams?
A:
When it uses peak theoretical bandwidth or unrealistic operational intensity instead of measured production behavior.
FOLLOW-UP TRAP: Treating a slide-level roofline as signoff proof.Why can IPC be an unsafe headline metric?
diagram
[INT][ARCH]
Q: Why can IPC be an unsafe headline metric?
A:
It hides stall composition and workload mix; equal IPC can mask very different resource stress patterns.
FOLLOW-UP TRAP: Declaring architecture success from IPC alone.What is your first action after seeing a sudden throughput drop?
diagram
[INT][ARCH]
Q: What is your first action after seeing a sudden throughput drop?
A:
Freeze workload revision and runtime envelope, then compare counter and trace deltas against the last known-good run.
FOLLOW-UP TRAP: Changing multiple knobs before baseline alignment.How do queue metrics help bottleneck analysis?
diagram
[INT][ARCH]
Q: How do queue metrics help bottleneck analysis?
A:
Queue depth, service rate, and backpressure timing reveal where throughput is truly capped.
FOLLOW-UP TRAP: Using queue occupancy without service-rate context.When should you reject a proposed compute-width increase?
diagram
[INT][ARCH]
Q: When should you reject a proposed compute-width increase?
A:
Reject when workload placement is bandwidth- or synchronization-bound and extra width raises area/power without KPI gain.
FOLLOW-UP TRAP: Approving width increases because peak TOPS improves.How do you keep tuning decisions senior-level?
diagram
[INT][ARCH]
Q: How do you keep tuning decisions senior-level?
A:
Document expected gain, regression surface, rollout guardrails, and rollback triggers before enabling changes.
FOLLOW-UP TRAP: Shipping a lab-only tuning tweak with no production guardrails.What indicates counter instrumentation debt?
diagram
[INT][ARCH]
Q: What indicates counter instrumentation debt?
A:
Frequent root-cause debates caused by ambiguous events and missing phase-aware telemetry.
FOLLOW-UP TRAP: Assuming more counters automatically means better observability.How do you evaluate workload representativeness?
diagram
[INT][ARCH]
Q: How do you evaluate workload representativeness?
A:
Check traffic shape, phase behavior, concurrency mix, and data locality against field deployment profiles.
FOLLOW-UP TRAP: Tuning only on a single popular benchmark.Why is p99 latency essential in architecture tuning?
diagram
[INT][ARCH]
Q: Why is p99 latency essential in architecture tuning?
A:
Tail behavior exposes queueing and interference risks hidden by average throughput metrics.
FOLLOW-UP TRAP: Optimizing means while ignoring tail SLA failures.What is a strong bottleneck review memo?
diagram
[INT][ARCH]
Q: What is a strong bottleneck review memo?
A:
It states dominant limiter, competing hypothesis, falsification evidence, chosen fix, and regression checks.
FOLLOW-UP TRAP: A report dump without a decision and owner.How do memory hierarchy effects show up in roofline placement?
diagram
[INT][ARCH]
Q: How do memory hierarchy effects show up in roofline placement?
A:
Kernels shift between bandwidth ceilings as locality changes across cache, SRAM, and DRAM levels.
FOLLOW-UP TRAP: Using one flat bandwidth number for all kernels.When do you escalate a performance issue cross-functionally?
diagram
[INT][ARCH]
Q: When do you escalate a performance issue cross-functionally?
A:
Escalate when fixes cross architecture, compiler, firmware, or PD ownership boundaries and local optimizations stall.
FOLLOW-UP TRAP: Escalating before isolating a concrete mechanism.How do you guard against tuning overfit?
diagram
[INT][ARCH]
Q: How do you guard against tuning overfit?
A:
Use holdout workloads, long-duration runs, and multi-KPI acceptance gates before rollout.
FOLLOW-UP TRAP: Accepting gains from short benchmark loops only.Why track joules per operation with throughput?
diagram
[INT][ARCH]
Q: Why track joules per operation with throughput?
A:
Efficiency trends reveal unsustainable wins that may violate thermal and battery or TCO targets.
FOLLOW-UP TRAP: Treating power as a separate later-stage concern.What is the value of counter-to-trace correlation?
diagram
[INT][ARCH]
Q: What is the value of counter-to-trace correlation?
A:
It validates that aggregate telemetry reflects real timeline behavior rather than measurement artifacts.
FOLLOW-UP TRAP: Trusting counters that cannot be explained on traces.How do you choose the next optimization after a successful fix?
diagram
[INT][ARCH]
Q: How do you choose the next optimization after a successful fix?
A:
Recompute bottleneck ranking because removing one limiter often exposes the next dominant constraint.
FOLLOW-UP TRAP: Repeating the same fix pattern without re-ranking constraints.What does a release-ready workload tuning checklist include?
diagram
[INT][ARCH]
Q: What does a release-ready workload tuning checklist include?
A:
Baseline, representative mix, KPI deltas, regression budget, owner approvals, and rollback plan.
FOLLOW-UP TRAP: Only benchmark score improvements.How should senior candidates answer performance debug questions?
diagram
[INT][ARCH]
Q: How should senior candidates answer performance debug questions?
A:
Frame answers as METRIC -> MECHANISM -> EXPERIMENT -> FIX -> REGRESSION -> RELEASE JUDGMENT.
FOLLOW-UP TRAP: Listing tools and commands without decision logic.Q&A drill guide
diagram
MECHANISM → METRIC → EXPERIMENT → TRADEOFF → VALIDATIONSketch while answering
diagram
TOP-DOWN PERFORMANCE METHOD
Total cycles
├─ Retiring useful work
├─ Frontend bound
├─ Bad speculation
├─ Backend core bound
└─ Backend memory bound
Only after classification should you propose cache, branch, pipeline, or NoC changes.Key takeaways
Connect every architecture claim to a workload and measurable metric.
State verification and PPA impact before proposing design changes.
Common pitfalls
Feature-driven design without MPKI/IPC/bandwidth evidence.
Ignoring coherency and NoC traffic in cache and accelerator sizing.