The ParPar Project
The goal of the ParPar project is to provide a platform for research
on parallel systems.
Its premise is that such a system can be built totaly in software,
based on off-the-shelf components.
Thus the hardware base is a set of PCs with a fast local area network,
and the software is based on a Unix kernel on each node.
The main software components and their interactions are depicted in
the following figure:
- Masterd
- The master daemon.
This is the heart of the system.
It is resopnsible for configuration management, resource management,
etc.
There is one masterd in the system, running on the host node.
- Noded
- The node daemon.
It is responsible for local activities on the node, e.g. the spawning
of sproc's.
There is one on each node.
- Master Control
- A GUI used to interact with the masterd, in order to control the
system (e.g. shutdown) and obtain information (e.g. about load conditions).
- Job Rep
- A GUI used to interact with the masterd in order to submit a job.
It also connects directly to the sproc's in order to display standard
I/O.
- sproc
- A "sub-process", i.e. a constituent of a parallel job (or xproc, for
"extended process").
For more information, or if you want to do some work on this project,
contact Dror Feitelson at
feit@cs.huji.ac.il.
You can also download the
full design document (currently in
version 0.2, at 77 pages and 140 KB; we're working on an updated version).
Publications
We also have a ParPar photo album (total
about 145 KB)
Acknowledgements
- Dror Feitelson - general design and miscelleneous pieces of
implementation
- Marc Volovic - initial design and implementation of infrastructure
- Anat Batat - job representative, memory considerations in gang scheduling
- Uri Lublin - gang scheduling support
- Dudi Er-El - RDGM design and implementation
- Avi Kavas - RDGM design and implementation, file transfer, study
of base kernels
- Gabi Benhanokh - general support, reimplementation of masterd in
C++
- Yoav Etsion - integration of FM, integration of communication and
gang scheduling
- Noam Lotner - general support
- Dan Tsafrir - gang scheduling extension for evolving jobs
- Niva Aldema - parallel debugger interface
- Hani Mazar - master control interface
- Mickael Raizman - FM routing and network topology
- Eitan Frachtenberg - flexible gang scheduling and dynamic
identification of gangs
- Tomer Klainer - parallel I/O to partitioned matrices with logical
caching
- Avraham Fraenkel - SGML-based parallel file system
To the Parallel
Systems Lab home page
To the Hebrew University Institute
of Computer Science home page