GitLab

Simplified function for preparing data redistribution. New feature for intercomms where they communicate using intracomms when redistributing data. Other minor changes.

Mino refactor in Merge spawn + P2P redistribution to avoid possible blocking of some parents processes to send data waiting for other parents to start receiving it.

Bugfix for Data Redistribution with Merge Expand. Array sizes for parents should always be the total amount of particular processes between both groups. Also a minor refactor in distribution.c

Hotfix in CommDist for Merge Shrink. Refactor in block distribution, removed not needed variable, added error checking in alltoall and minor changes.

Various bugfixes over the changes in ADR and SDR. Also a fix in results for asynch iterations. Some warning corrections.

Refactored Manager and CommDist for non-blocking communications and MPI_Request usage. Two minor bug fixes. RMA asynchronous and Ibarrier strategy are the next steps to develop

Refactor for redistributing data with non-blocking funtions. Modified redistributing data constants names to -RED-. Fixed potential bug over requests as the number of requests used was not clear. WIP over CommDist and Manager to adapt non blocking calls. Seems to be working fine right now but more testing is required.

Added synchronous MPI-RMA redistributions. Works over Baseline and Merge spawn, but has not been tested with all the strategies

Added functionality for asynchronous redistributions with threads only. Support for non blocking calls is not yet fully implemented for Merge method.

Modificada salida de resultados, ahora tambien se indican los tiempos para cada etapa de una iteracion como locales

Arreglado funcionamiento de Regresion Lineal. Refactorizado codigo de Process_stage.c. Protegidos los archivos .h. En el tiempo final se tiene en cuenta el tiempo desperdiciado

WIP. Refactoring. Ahora se ha separado completamente la logica de la maleabilidad de la aplicación sintética

WIP. Encontrada condicion de carrera entorno a Alltoallv cuando el tiempo por iteracion es menor a 10ms

WIP. Añadido nuevo método de comunicación asíncrona. Ahora es posible elegir entre 4 métodos (Por defecto se usa uno con hebras). Se ha añadido a la salida de configuracion el numero de bytes de comunicación en cada iteración de computo.

Cambios en los archivos de ejecucion para permitir mas opciones y automaticamente relanzar trabajos con errores

Anadida maleabilidad asincrona y nuevo parametro para indicar como realizaran la espera de transmision de datos asincronos los padres