next up previous contents
Next: BIP message-passing model and Up: BIP Messages User Manual Previous: Contents

Introduction

Our objective is to provide for the Myrinet [BCFtex2html_wrap63395, Myr95, SBSS93, LS92, FDCF94] network a basic interface for message-passing, targeted towards parallel computing. Of course there are standard API well established: MPI [SOHLtex2html_wrap63395] and PVM [GBDtex2html_wrap63394], but to begin with we choose to implement our own simpler interface: BIP (for Basic Interface for Parallelism). One of BIP main goal is to really achieve the maximal performance of the hardware, but with a sufficient abstraction level, to hide the hardware details.

Currently BIP messages has only been implemented for cluster of x86/Linux workstations, linked by Myrinet boards with the LANAI4.1 processor.

The base principle of BIP messages is to implement all communication in a user-level library with zero-memory copies and direct access the network hardware without system calls. Our current implementation achieves less than 5 microseconds latency for small messages and more than 125Mbytes/s bandwidth (so more than 1Gigabit/s if you take Giga=tex2html_wrap_inline21).



Loic Prylli
Mon Jun 8 09:38:30 CEST 1998