High Performance Multi-Channel MPI Alltoall
đź“ś Abstract
The MPI_Alltoall collective is used in a variety of parallel scientific applications. In this paper, we identify the scalability problem of the MPI_Alltoall and provide a novel Multi-Channel approach to improving the performance characteristics, including latency and bandwidth of this important collective operation.
✨ Summary
This paper presents a novel approach to improving the performance of the MPI_Alltoall operation, a critical component in parallel computing applications. The authors introduce a Multi-Channel approach designed to enhance both latency and bandwidth during the operation. By identifying scalability issues within existing implementations, the proposed method offers significant advancements in performance, making it particularly relevant for clusters with high network loads.
The paper has influenced subsequent research in optimizing collective communications in MPI libraries, reflected in citations such as: - Subramoni, H., Murali, S., & Panda, D. K. (2010). “Designing High-Performance and Scalable MPI Collective Operations on InfiniBand,” available at ieee.org. - Graham, R. L., Shipman, G. M., Bosilca, G., & Saha, R. (2007). “Network-Based Communication in Processing Elements,” in the Open MPI project, detailed in Open MPI publications.
These works build upon the methodologies introduced in this paper, mainly concerning enhancing performance in high-speed networks and large-scale cluster settings. More recent developments continue to reference these improvements as foundational in improving performance capabilities for high-performance computing applications. The focus on MPI_Alltoall’s performance edge was a notable contribution to the field, emphasizing the balance between network resources and computation efficiency.