Capítulo de livro Revisado por pares

RDMA Communication Based on Rotating Buffers for Efficient Parallel Fine-Grain Computations

2004; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-540-24669-5_7

ISSN

1611-3349

Autores

Adam Smyk, Marek Tudruj,

Tópico(s)

Embedded Systems Design Techniques

Resumo

The paper presents a new method for efficient fine grain computations on distributed memory computers. RDMA (Remote Direct Memory Access) communication is applied which assures direct access to memories of remote processing nodes. To obtain high RDMA efficiency for fine-grain computations with very frequent transmissions of small size messages, a specially designed structure of RDMA rotating buffers (RB) is introduced. It allows to fully exploit available communication bandwidth by provision of a special communication control infrastructure prepared and activated in a program before effective computations start. As an example of a fine-grain problem implemented with the RDMA rotating buffers, the discrete Fast Fourier Transform (FFT) execution is presented. “The binary-exchange algorithm” of FFT is examined showing efficiency of the RB method in comparison to standard MPI communication.

Referência(s)