Need to address the performance of certain algorithms in the nx-parallel whose parallel implementations are currently slower than their corresponding implementations in NetworkX. Some of these algorithms include:
Performance Comparison: What should we do with these functions that show slower performance in their parallel implementations? Should we consider reverting to or recommending the use of NetworkX's implementation in these cases?
Optimization: Is there a more optimized way that would improve the performance of these parallel implementations? And will that be way better the performance for all the algorithms? And how can we give the control of changing these "ways" to the end user?
Functionality: There are also some graph algorithms that cannot be parallelized due to their fundamentally non-parallelizable nature. How should we handle such algorithms in nx-parallel? Should we provide a clear indication in the documentation or in the code itself? Or not add them to nx-parallel at all?
Next Steps:
Gather insights and suggestions from the community.
Determine whether to optimize, remove, or replace the slower parallel implementations.
Identify any non-parallelizable functions and decide on how to best handle them.
Description:
Need to address the performance of certain algorithms in the
nx-parallelwhose parallel implementations are currently slower than their corresponding implementations in NetworkX. Some of these algorithms include:v_structuresin https://github.com/networkx/nx-parallel/pull/74collidersin https://github.com/networkx/nx-parallel/pull/74global_reaching_centralityin https://github.com/networkx/nx-parallel/pull/44local_reaching_centralityin https://github.com/networkx/nx-parallel/pull/44average_clusteringapproximate_average_clusteringnumber_of_isolatesIssues to Discuss:
Performance Comparison: What should we do with these functions that show slower performance in their parallel implementations? Should we consider reverting to or recommending the use of NetworkX's implementation in these cases?
Optimization: Is there a more optimized way that would improve the performance of these parallel implementations? And will that be way better the performance for all the algorithms? And how can we give the control of changing these "ways" to the end user?
Functionality: There are also some graph algorithms that cannot be parallelized due to their fundamentally non-parallelizable nature. How should we handle such algorithms in nx-parallel? Should we provide a clear indication in the documentation or in the code itself? Or not add them to nx-parallel at all?
Next Steps:
Thank you :)