|Home -> Research -> Older Projects -> CollOpt|
MPI Collective Optimization|
MPI Collective Operations are crucial for the running time of many
parallel applications and algorithms. Our group tries to enhance the
running time of Collective Operations on particular networks (e.g.
InfiniBand, Ethernet). We are also testing the impact of non-blocking
collectives on current applications (e.g. Linpack, Ab Initio Codes).
InfiniBand optimized collective operationsA collective component which enables optimized collective comunication over InfiniBand is under development at the TU Chemnitz. Currently, only an optimized InfiniBand barrier (see "Fast Barrier Synchronization for InfiniBand") is available. Other collective operations are currently under development by students under my advisory. We use different hardware sepcialities of InfiniBand, as the different transport types, Multicast and the hardware parallelism (see "LogfP - A Model for small Messages in InfiniBand") to optimize collective operations. The main goal is to prove that all collective operations can be implemented in a non-blocking manner without a performance penalty. Non-blocking collective operations can be used to overlap communication and computation efficiently.
Adaptive collective operationsWe investigated also the possibility for self-optimization techniques for collective operations. A result of this is an adaptive barrier which benchmarks a set of different barrier algorithms during the creation of each communicator and chooses the fastest one to perform the operation afterwards. We used this framework to verify LogP predictions for all operations, see "A practical approach to the rating of barrier algorithms using the LogP model and Open-MPI".
|serving: 126.96.36.199:56500||© Torsten Hoefler|