Runtime inter-process communication control

By controlling the interconnect of a cluster system as a dynamically controllable resource, we have been promoting research on technology and technique for optimization and streamlining of inter-process communication occured during distributed and parallel computing. In particular, this study focuses on MPI (Message Passing Interface), a de facto starndard library for distributed and prallel programming and aims to reduce the execution time of MPI collective communication. For this purpose, we have taken a strategy of introducing SDN (Software Defined Networking), a newly-emerged networking concept that enables dynamic control of network flows for MPI.

 

Research achievement(paper)

• Baatarsuren Munkhdorj, Keichi Takahashi, Dashdavaa Dashdavaa, Yasuhiro Watashiba, Yoshiyuki Kido, Susumu Date and Shinji Shimojo, “Design and Implementation of Control Sequence Generator for SDN-enhanced MPI”, The 5th International Workshop on Network-aware Data Management(NDM’15), Austin, Nov. 2015. (accpeted)
• Keichi Takahashi, Dashdavaa Khureltulga, Yasuhiro Watashiba, Yoshiyuki Kido, Susumu Date, Shinji Shimojo, “Performance Evaluation of SDN-enhanced MPI_Allreduce on a Cluster System with Fat-tree Interconnect”, The International Conference on High Performance Computing and Simulations (HPCS2014), pp. 784 – 792, Jul. 2014. (10.1109/HPCSim.2014.6903768)
• Khureltulga Dashdavaa, Susumu Date, Hiroaki Yamanaka, Eiji Kawai, Yasuhiro Watashiba, Kohei Ichikawa, Hirotake Abe, and Shinji Shimojo, ”Architecture of a High-speed MPI_Bcast Leveraging Software-Defined Network”, The 6th Workshop on UnConventional High Performance Computing　2013, Lecture Notes in Computer Science Volume 8374, 2014, pp 885-894, Aachen, Germany, Aug. 2013. (10.1007/978-3-642-54420-0_86)