τ-Lop is a new parallel performance model aimed to help in the design and optimization of parallel algorithms in multi-core clusters. It represents a parallel algorithm and predicts accurately its costs through the concept of concurrent transfers. By now, τ-Lop has been applied to underlying algorithms in MPICH and Open MPI mainstream implementations of some MPI collectives in shared memory. Current work is in the application of the model to collective operations when deployed in networks of multicore nodes.
Stars
4
Forks
4
Watchers
4
Open Issues
1
Overall repository health assessment
No package.json found
This might not be a Node.js project
Suara2D/3D: included an algorithm for evaluating numerically the allreduce in a grid (Suara) considering the mapping of the processes to the platform.
3d819c0View on GitHubBug: this is an algorithm for a flat platform, and mapping is not considered. However, how to deal with the platform characteristics? Yet an open problem.
2e090c7View on GitHubBugs fix: In a grid, the algorithm executed in rows runs with Pc processes, while the algorithm executed in columns runs with Pr processes. In the 3D case, the algorithm executed in depth runs with Pd processes. It is a bit confusing.
7fed270View on GitHubSuara2D and Suara3D stage 1 algorithm implementation for the MPI allreduce collective operation.
252aec8View on GitHubMapping creation bug fixed. Mapping() constructor takes parameters P, M, Q and Map. M is included.
b1a7a81View on GitHubNew parameters and transform new option (-vv: highly verbose).
e9da277View on GitHubTransform parameters: adding a warning to show if times lower than 0.0 are obtained in the process of transforming mpiblib values to taulop params in a channel.
2902457View on GitHubTransform parameters: adding a warning to show if times lower than 0.0 are obtained in the process of transforming mpiblib values to taulop params in a channel.
80488f9View on GitHubTransform parameters code changed to adapt to the new mpiblib parameters files format.
4d27897View on GitHubChanges in mpiblib requires new form of transforming: introduced k=semgnetsize, generating overhead eager and rendezvous (both).
e8395a4View on GitHubTransform params from TCP: fix bug in computation of the L.
682019aView on GitHub