Carto - Topology Aware Processor Allocation with MPI
Carto is a framework for topology aware allocation of processors with OpenMPI. Currently Carto uses the physical topology as a basis for processor allocation in order to improve network performance. As there may be large differences between the physical topology and the logical topology as defined by the routing algorithm, performance can possible be further improved by making Carto routing aware, i.e. exploit the current routing to better utilise the network by proper placement of jobs.
Carto is a framework for topology aware allocation of processors with OpenMPI. Currently Carto uses the physical topology as a basis for processor allocation in order to improve network performance. As there may be large differences between the physical topology and the logical topology as defined by the routing algorithm, performance can possible be further improved by making Carto routing aware, i.e. exploit the current routing to better utilise the network by proper placement of jobs.
The purpose of the thesis will be to implement and evaluate routing awareness in Carto/OpenMPI.
What you will do:
- you will evaluate routing awareness in Carto
- you will make performance measurement in simulations or on real hardware
- you will do development in the C language as part of the OpenMPI framework and the OpenSM/OFED software stack
What you will learn:
- you will get to know the InfiniBand architecture, a state of the art high performance networking technology
- you will obtain a thorough understanding of the OpenMPI and the OFED/OpenFabrics software stack
- you will learn to do development in a large open source development project
- you will learn to analyse network performance
References
Paper accepted at ICC 2010
Invited talk by Kang Li 15. December
