InfiniBand networks are commonly used in the high performance computing area. They offer RDMA-based operations that help to improve the performance of communication subsystems. In this paper, we propose a minimal message-passing communication layer providing the programmer with a point-to-point communication channel implemented by way of InfiniBand RDMA features. Differently from other libraries exploiting the InfiniBand features, such as the well-known Message Passing Interface (MPI), the proposed library is a communication layer only rather than a programming model, and can be easily used as building block for high-level parallel programming frameworks. Evaluated on micro-benchmarks, the proposed RDMA-based communication channel implementation achieves a comparable performance with highly optimised MPI/InfiniBand implementations. Eventually, the flexibility of the communication layer is evaluated by integrating it within the FastFlow parallel framework, currently supporting TCP/IP networks (via the ZeroMQ communication library). © 2014 IEEE.
Message passing on InfiniBand RDMA for parallel run-time supports
TORQUATI, MASSIMO
2014-01-01
Abstract
InfiniBand networks are commonly used in the high performance computing area. They offer RDMA-based operations that help to improve the performance of communication subsystems. In this paper, we propose a minimal message-passing communication layer providing the programmer with a point-to-point communication channel implemented by way of InfiniBand RDMA features. Differently from other libraries exploiting the InfiniBand features, such as the well-known Message Passing Interface (MPI), the proposed library is a communication layer only rather than a programming model, and can be easily used as building block for high-level parallel programming frameworks. Evaluated on micro-benchmarks, the proposed RDMA-based communication channel implementation achieves a comparable performance with highly optimised MPI/InfiniBand implementations. Eventually, the flexibility of the communication layer is evaluated by integrating it within the FastFlow parallel framework, currently supporting TCP/IP networks (via the ZeroMQ communication library). © 2014 IEEE.File | Dimensione | Formato | |
---|---|---|---|
779953_PDP_postprint.pdf
accesso aperto
Tipologia:
Documento in Post-print
Licenza:
Tutti i diritti riservati (All rights reserved)
Dimensione
168.23 kB
Formato
Adobe PDF
|
168.23 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.