Results & Findings

This section provides a consolidated view of benchmark results across tasks. Detailed reproduction steps and full result analysis can be found on each task page.

Task 2 — LINPACK Performance (GFLOPS)

Peak result: 11.81 GFLOPS at N=18000 on 8 Pi 3 nodes (2×4 grid, verified, PASSED). A full strong-scaling sweep was run across N ∈ {5000, 8000, 18000} × {1, 2, 4, 8} workers (5 repeats per cell); the headline configuration is shown below.

N	Workers	P×Q	GFLOPS	Residual	Status
18000	1	1×1	—	—	CRASH (OOM)
18000	2	1×2	—	—	OOM
18000	4	2×2	7.61	—	PASSED
18000	8	2×4	11.81	3.74e-03	PASSED

Full reproduction guide →

---

Task 3 — MPI Scalability

24.45× π strong speedup at 32 ranks — 16 ranks fastest stencil strong configuration — 6.23x Stencil speedup at 16 ranks

Full reproduction guide →

---

Task 4 — Task Distributor (Amdahl/Gustafson)

Amdahl's Law explains why small images scale adversely as you add cores, while Gustafson's Law explains why the exact same architecture successfully achieves positive speedup once you scale the image size up to 3200x2400.

The experiment used a fixed resolution per batch and varied the worker count across 1, 2, 4, and 8 workers.

Workers	Width	Height	Runs	Average Runtime [s]	Speedup	Efficiency	Serial Fraction Estimate
1	800	600	3	4.057	1.000	1.000	0.003
2	800	600	3	4.190	0.968	0.484	0.034
4	800	600	3	6.881	0.590	0.147	0.022
8	800	600	3	9.243	0.439	0.055	0.016
1	1600	600	3	5.057	1.000	1.000	0.002
2	1600	600	3	5.616	0.900	0.450	0.041
4	1600	600	3	6.970	0.725	0.181	0.034
8	1600	600	3	7.989	0.633	0.079	0.029
1	1600	1200	3	7.060	1.000	1.000	0.001
2	1600	1200	3	8.728	0.809	0.404	0.038
4	1600	1200	3	8.832	0.799	0.200	0.049
8	1600	1200	3	11.535	0.612	0.077	0.037
1	3200	2400	3	30.121	1.000	1.000	0.000
2	3200	2400	3	23.806	1.265	0.633	0.044
4	3200	2400	3	18.914	1.593	0.398	0.061
8	3200	2400	3	17.900	1.683	0.210	0.081

Full reproduction guide →