ASIM team
LIP6 Laboratory
Paris, France

Parallel Virtual Machine

The  MPC  parallel machine offers high performance communications  over an  HSL  network  through the use of  the direct deposit  protocol (also called remote write). The objective of a PVM implementation in a multiprocessor like MPC is to explore the potential of this network to improve the performance of PVM systems.

As the MPC  multicomputer is a distributed memory architecture, the use of a message passing library is mandatory  to ease the development of user applications. In a first moment our choice was  PVM  because of the wide acceptance it gained and the fact that it become a de facto standard  for message-passing  programming. So a PVM implementation in the MPC parallel computer is provided.

The implementation is based on the 3.3 version and is optimized for the MPC architecture. One special node is chosen to run the PVM daemon. It is called  service  node. The other nodes are used for number crunching  (compute nodes). This schema follows PVM implementations in several other parallel machines. The PVM Daemon is essentially responsible for load balancing, task creation and management, and  serves as the gateway to the outside world. The use of only one daemon allows external workstations to recognize the MPC parallel computer as a unique, high performance machine.

In this implementation, a task can be started on the service node as a Unix process and enrolls in PVM by establishing a Unix socket connection  to the daemon. The only way to start PVM tasks on the compute nodes via a pvm_spawn() through the task started in the service node. When the daemon receives a request to spawn new tasks, it allocates a set of nodes, and load the executable in the specified number of nodes. The library to be linked with the task that will run on the service node is the standard PVM library; for compute nodes there is another library that implements message passing functions in terms of SLR/V system calls. This means that all communication between the daemon and compute node tasks and between compute nodes tasks themselves are made through the high performance HSL network.

You can find more general informations about PVM on The documentation about the PVM port to MPC (in french) is available here.

Contact F. Silva, I. Demeure or K. Mana to get more informations about this implementation of PVM - this www page was made by K. Mana and F. Silva.
HTML-4.0 compliant Made With Cascading Style Sheets Powered by FreeBSD Powered by Apache

Server design A. Fenyö - contact people
About this Web Site
$Date: 1998/03/02 23:26:20 $
Copyright © 1997-1998 UPMC/LIP6
All rights reserved