Path Based Optimization of MPI Collective Communication Operation in Cloud

There has been a considerable research in collective communication operations, especially in MPI Broadcast and Gather operations, on message passing platforms. Majority of the research work is done, to improve efficiency of the collective communication operations for specific architectures by considering either their communication network or platform parameters. In this work, a simple and general approach to optimize legacy MPI collective communication algorithms for cloud environment, is proposed. Cloud Computing is one of the prominent technologies as a new platform for distributed, large scale applications. Cloud services reduce investment in hardware cost and make application development and deployment faster. Because of the use of hardware virtualization, it is flexible and elastic in terms of resource provisioning. This virtualization hides the network topology information. Scalable parallel programs can be efficiently and effectively expressed using MPI library, which is powerful and portable. System performance is mostly affected by its collective communication operations. The use of hardware virtualization, makes topology aware communication algorithm ineffective. Thus the main focus in this paper is on improving the efficiency of MPI collective communication operations, especially on MPI Broadcast, Gather and exploring cost optimizations for MPI. Therefore, a new approach to improve broadcast and gather operation based on network performance matrices, is developed. The proposed approach is tested on LAN environment and a HPC cluster. The experimental results are improved compared to the existing MPICH2 library.