High performance RDMA-based MPI implementation over InfiniBand
Jiuxing Liu, Jiesheng Wu, Sushmitha P. Kini, Pete Wyckoff, Dhabaleswar K. DK Panda
Abstract
Jiuxing Liu, Jiesheng Wu, Sushmitha P. Kini, Pete Wyckoff, Dhabaleswar K. DK Panda
Abstract
Although InfiniBand Architecture is relatively new in the high performance computing area, it offers many features which help us to improve the performance of communication subsystems. One of these features is Remote Direct Memory Access (RDMA) operations. In this paper, we propose a new design of MPI over InfiniBand which brings the benefit of RDMA to not only large messages, but also small and control messages. We also achieve better scalability by exploiting application communication pattern and combining send/receive operations with RDMA operations. Our RDMA-based MPI implementation currently delivers a latency of 6.8 microseconds for small messages and a peak bandwidth of 871 Million Bytes (831 Mega Bytes) per second. Performance evaluation at the MPI level shows that for small messages, our RDMA-based design can reduce the latency by 24%, increase the bandwidth by over 104%, and reduce the host overhead by up to 22%. For large messages, we improve performance by reducing the time for transferring control messages. We have also shown that our new design is beneficial to MPI collective communication and NAS Parallel Benchmarks.
OpenAlex reports 358 citations for this work. Citation counts describe recorded attention and do not establish research quality.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Although InfiniBand Architecture is relatively new in the high performance computing area, it offers many features which help us to improve the performance of communication subsystems. One of these features is Remote Direct Memory Access (RDMA) operations. In this paper, we propose a new design of MPI over InfiniBand which brings the benefit of RDMA to not only large messages, but also small and control messages. We also achieve better scalability by exploiting application communication pattern and combining send/receive operations with RDMA operations. Our RDMA-based MPI implementation currently delivers a latency of 6.8 microseconds for small messages and a peak bandwidth of 871 Million Bytes (831 Mega Bytes) per second. Performance evaluation at the MPI level shows that for small messages, our RDMA-based design can reduce the latency by 24%, increase the bandwidth by over 104%, and reduce the host overhead by up to 22%. For large messages, we improve performance by reducing the time for transferring control messages. We have also shown that our new design is beneficial to MPI collective communication and NAS Parallel Benchmarks.
Key concepts: Remote direct memory access, InfiniBand, Computer science, Scalability, Operating system, Byte, Computer network, Low latency (capital markets)