Systems Research

High Performance and Scalable MPI Implementation on InfiniBand

Principal Investigators: D.K. Panda and P. Wyckoff
Funding Source: Sandia National Labs, CA
Duration: 5/2/2002 - 3/30/2003

Description: The recently proposed InfiniBand (IBA) standard provides many new features and mechanisms which are not available in other contemporary and popular cluster interconnect technologies such as Myrinet and Gigabit Ethernet. These new features raise interesting challenges regarding how to take advantage of them to design a high performance and scalable MPI implementation for large clusters. We are investigating aspects associated with implementing both point-to-point and collective communication primitives of the MPI standard. Specifically, we focus on the following topics: optimal usage of IBA transport services, efficient rendezvous protocol with RDMA and service level support, using atomic operations for collective communication, using multicast support for collective communication, and performance and scalability benefits of the MPI implementations on IBA compared to other networks.