Sampling & Representativeness — SimPoint, SMARTS, ROI Discipline

Goal: Cut simulation time while bounding error and preserving phase behavior.

---

📋 Table of Contents

---

1) Why sampling works

Programs exhibit phases. If we identify representative regions and weight them correctly, we can estimate whole‑program metrics with bounded error.

---

2) SimPoint (phase clustering)

• Build Basic Block Vectors (BBVs) over fixed intervals (e.g., 10M instructions).
• Cluster intervals; pick one simulation point per cluster.
• Weight each point by cluster size; warm microarchitectural state before measurement.

Weights and reconstruction: For metric m, m_total ≈ Σ_i (w_i × m_i) with Σ_i w_i = 1.

Tips

• Keep BBVs architecture‑agnostic if possible (ease reuse).
• Re‑cluster when inputs change substantially.
• Visualize with t‑SNE/UMAP to sanity‑check clusters.

---

3) SMARTS (statistical sampling)

• Sample short windows periodically/randomly; checkpoint warmed state to reduce overhead (TurboSMARTS).
• Provides confidence intervals if windows cover the longest latencies.

Window sizing rule: window length > (worst‑case memory latency + queuing) × pipeline depth of dependency chains.

---

4) Multicore/multithread specifics

• Preserve timestamps & core IDs in traces to reconstruct interference.
• Use time‑based windows for barrier‑heavy workloads.

---

5) What to publish

• Number of points, interval sizes, warmup lengths, selection method, and measured error vs. full runs.
• Plots of phase stability across configs help reviewers (and future you).

---

References

• SimPoint (UCSD); SMARTS & TurboSMARTS papers.