This document outlines the final benchmarking results, the optimisation techniques applied and their performance effects.
The solver was benchmarked, obtaining the results shown in the following table:
| Benchmark name | Samples | Iterations | Estimated |
|---|---|---|---|
| mean | low mean | high mean | |
| std dev | low std dev | high std dev |
| Diffusion | Samples | Iterations | Estimated |
|---|---|---|---|
| 100 | 1 | 1.12728 s | |
| 11.2673 ms | 11.2522 ms | 11.2828 ms | |
| 78.1882 us | 69.031 us | 90.6525 us |
| Jacobi | Samples | Iterations | Estimated |
|---|---|---|---|
| 100 | 1 | 143.387 ms | |
| 1.43005 ms | 1.42343 ms | 1.4383 ms | |
| 37.3863 us | 31.027 us | 47.4977 us |
| Gauss-Seidel | Samples | Iterations | Estimated |
|---|---|---|---|
| 100 | 1 | 566.289 ms | |
| 5.62072 ms | 5.60853 ms | 5.63598 ms | |
| 69.3318 us | 54.8678 us | 93.2107 us |
*Note that the non-solver parts of the software have not been benchmarked, as they are dependent on user input and have a minimal impact in performance.
The code already followed several best practices, leading to a reasonable performance. However, several improvements were made in different areas:
- As shown in the benchmarking table, the algorithm choice greatly impacts performance and dominates over other factors. This resulted in improvements of several miliseconds.
- We replaced in-house functions and structures with those from the standard library when possible, as those are already optimised.
- Additionally, we chose to use the standard array instead of vector when possible. This resulted in small improvements of around 0.01 miliseconds.
- We precomputed variables which were constant throughout the loops, such as products and powers.
- We swapped pointers in substitution of the copy constructor, thus avoiding the creation of new objects.
- We simplified conditionals and loops, reducing the number of checks. This resulted in an improvement of around 1 milisecond.
Furthermore, we used the -O3 when compiling to further optimise the code. -O2 was also considered, but -O3 was kept as we did not encounter stability problems.