https://github.com/kchristin22/graph_minor
Compute the graph minor when given the cluster configuration. Implemented in three versions using OpenMPI, pthreads and OpenCilk.
https://github.com/kchristin22/graph_minor
csr graph-algorithms graph-minor opencilk openmpi openmpi-benchmark parallel-computing pthreads
Last synced: about 1 year ago
JSON representation
Compute the graph minor when given the cluster configuration. Implemented in three versions using OpenMPI, pthreads and OpenCilk.
- Host: GitHub
- URL: https://github.com/kchristin22/graph_minor
- Owner: kchristin22
- License: gpl-3.0
- Created: 2023-11-05T12:31:41.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-12-15T18:48:16.000Z (over 1 year ago)
- Last Synced: 2025-01-22T07:45:33.356Z (over 1 year ago)
- Topics: csr, graph-algorithms, graph-minor, opencilk, openmpi, openmpi-benchmark, parallel-computing, pthreads
- Language: C++
- Homepage:
- Size: 627 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Graph_minor
## (!!!) There's an error in the benchmarks in the report: instead of pthreads, OpenMP was run. The fixed tests are below.
## You may need to remove the condition check on using sparse or dense matrix implementation for the program to run for matrices where n $^2$ > n + 2*nz.
## After validating the sequential algorithm you can check the validity of the parallel implementation by using the function`areCSRVectorsEqual` for larger datasets as well.
### **About:**
The purpose of this project is to calculate the graph minor of a given graph, based on an input vector denoting the clusters each original node belongs to. This vector contains the ids of the
clusters, ranging from 1 to the amount of discrete clusters, in the equivalent positions of the corresponding nodes. No assumptions on the nature of the input graph are made. However, since most of the big networks are sparse, the primary emphasis lies on sparse matrix representations, and partic-ularly in the Compressed Sparse Row (CSR) format. It is important to note that this report does not address any preprocessing steps required to convert the input adjacency matrix into CSR format.
### **Benchmarks:**
Comparison of implementations using the median time in us:
| Dataset | Matlab | Sequential | OpenMP (4 threads) | Pthreads (4 threads) | OpenCilk (4 threads)|
|---------|-----------|-----------|----------|---------| -----------|
| ri2010 | 4500 | 34289 | 9767 | 33618 | 19133 |
| nj2010 | 56600 |1584908 | 339045 | 545741 | 544979 |
| pa2010 | 110000 |6037295 | 2060738 | 2739690 | 2103721 |
| ca2010 | 187400 |15898973 | 5423717 | 6440966 | 4838647 |
| tx2010 | 231300 |25460247 | 8036885 | 9773966 | 7118827 |
OpenCilk manages to outperform OpenMP in larger data sets, despite the use of a common atomic vector. This is assumed to stem from its work-stealing properties.
The version of Matlab was measured using its fastest implementation of the two.
The `us2010` dataset was never managed to be parsed correctly or quicky enough.