Grid'5000 user report for
User information( user)
More user information in the user management interface.
- Optimization of MPI alltoallv operation for fat and long networks (Networking) [in progress]
Description: There exists diverse optimizations for total exchange operations. The idea is to optimize the alltoallv operation in the case of inter-cluster exchange (with fat and long links involved).
- Analysis of the AllToAllV in cluster of cluster environments (Networking) [in progress]
Description: The context of this work is the total exchange of data between processes running in different clusters linked by a wide-area link (backbone). We can experiment such an environment using several sites of the Grid5000 testbed. Starting from the work done on the AllToAll operation we have extended the study to the AllToAllV operation. AllToAll in the MPI terminology means a total exchange of pieces of data of the same size between participant processes. AllToAllV implies that each piece of data may be of a different size. Unlike AllToAll, it is difficult to choose a good routing algorithm (e.g binomial tree, Bruck algorithm) depending on the data size for AllToAllV since the processes do not know the amount of data sent by the others. MPI implementations (OpenMPI, MPICH-2, GridMPI) have today a very simple implementation for AllToAllV, in which all processes send and receive asynchronously their data to all other processes and wait for the completion of all communications. Yet simple, this strategy performs better than any optimized routing scheme used in the other collective operations. We have tried more sophitiscated approaches based on message aggregation, congestion regulation and load-balance. In this kind of strategy, we select forwarders processes at each cluster, whose task is to aggregate single messages into a larger one sent over the backbone. The number of forwarders allows to control how many TCP streams simultaneously compete for the backbone. We can balance the size of the aggregated messages with an extra step at the begining where process first exchange the information about how much data they send. However, our proposal have only equaled the original AllToAllV implementation so far, bringing no significant improvement. One of the key finding of this work is that message aggregation does not improve the overall throughtput of the streams over the wide area link. Work is currently undergoing to model the behavior of the various routing strategies in the PLogP model.