Repository of scripts that can be used to conduct statistical analysis on performance of different benchmark algorithms in HClib, a C/C++ programming system for distributed systems.
This folder contains a script to run a suite of 2-sample t-tests to assess if the difference between the average performance of the different variants of an algorithm is statistically significant. It generates a PDF file with a table indicating the results of the two-sample t-tests along with a graph comparing the laptimes of the algorithms across different node configurations.
This folder contains the laptimes for the original and the modified (two-mailbox) variants of the HClib topological sorting algorithm.
In conclusion, even though the performance improvement is positive on average, and it is statistically significant (alpha = 0.05) for all but 1 node/PE combination. The samples were gathered using the following node/PE configurations, recommended by my mentor Dr. Akihiro Hayashi (with an matrix input size of 100,000 rows and 100,000 columns):
- 1 node, 24 PEs --> statistically insignificant
- 2 nodes, 48 PEs --> statistically significant
- 4 nodes, 96 PEs --> statistically significant
- 8 nodes, 192 PEs --> statistically significant
- 16 nodes, 384 PEs --> statistically significant
This folder contains the laptimes for the original and the modified (linear search and binary search) variants of the HClib triangle counting algorithm. For each pairwise comparison of the algorithm, for all node/PE combinations, the differences were statistically significant. These differences can be seen in the figure below.
