Lecture 10 Exercises — Scaling Analysis (1-week workload)

Objective: measure strong and weak scaling on a small parallel benchmark, compute speedup/efficiency, and interpret results using Amdahl’s and Gustafson’s laws. Keep a lab log in codes/lab10/lab-notebook.md.

Exercise 1 — Prepare your workspace¶

1. In your sspa-sandbox, create codes/lab10/<name.surname>/ and copy the starter code from codes/lab10/python/ into it.

2. Verify the benchmark runs serially:

   • python bench_pi.py --processes 1 --samples 2000000

3. Record your environment in the lab log (CPU model, OS, python --version).

Exercise 2 — Strong scaling experiment¶

1. Fix the total work (e.g., --samples 20000000) and run with p = 1, 2, 4, 8, ... up to the core count you can access.

2. Use the provided runner to collect multiple trials and save a CSV:

   • python run_scaling.py strong --samples 20000000 --max-procs 16 --trials 5 --out strong.csv

3. In the lab log, note any anomalies (slow outliers, oversubscription effects, noisy runs).

Exercise 3 — Weak scaling experiment¶

1. Keep work per process constant (e.g., base samples per process) and scale the total samples with p.

2. Collect a second CSV:

   • python run_scaling.py weak --samples-per-proc 2000000 --max-procs 16 --trials 5 --out weak.csv

3. Discuss: does the wall time stay roughly constant? If not, why?

Exercise 4 — Analyze and interpret¶

1. Analyze both CSV files:

   • python analyze_scaling.py strong.csv

   • python analyze_scaling.py weak.csv

2. For strong scaling:

   • compute (S(p)) and (E(p))

   • estimate the serial fraction using the Karp–Flatt metric

   • compare the observed trend with Amdahl’s bound

3. For weak scaling:

   • report time growth (if any) and propose the likely bottleneck(s)

Exercise 5 — Short write-up¶

Write a short report (can be in lab-notebook.md) that includes:

• experimental setup (hardware, commands, parameters)

• strong scaling table (p, time, speedup, efficiency, estimated serial fraction)

• weak scaling table and interpretation

• at least 2 actionable performance ideas (e.g., reduce overhead, bigger chunks, different RNG strategy)

Submission checklist¶

• codes/lab10/<name.surname>/ containing:

  • your benchmark and scripts

  • strong.csv and weak.csv

  • lab-notebook.md with results and interpretation