Slides which contain outdated info are marked [old]
Table of Contents of the presentation:Hydra (title page only)
Author: Pawel Artymowicz
Home Page: http://planets.utsc.utoronto.ca/~pawel
Note on the reliability of Hydra (May 2001):
There are 7 Ultra 5 nodes w/o keyboards or monitors plus one regular workstation. Each unit has a power cord and a short twisted pair cable to the Fast Ethernet switch. It was uncertain in the beginning if we needed a little monitor and keyboard next to hydra for monitoring/repairs.They were simply unnecessary. On one or two occasions over the past 2.5 yr hydra needed attention: notably when the main fan mounted on a pizza box of Ultra 5 failed, this not only went unnoticed for x days, but even after detection turned out not to have caused any noticeable rise in the temperature of the box itself. The cpu fan seems to have provided sufficient cooling for the entire box, although it must be mentioned that the system is in an airconditioned closet. Failure rate of ~2/8/2.5 times/node/yr ~ 1/(10 yr)/node is remarkably low. In addition, on the software side the whole system is prectically zero maintainance. It only hanged and had to be restated several times in its life (except for scheduled upgrades). Can you beat that?
Recent work done with Hydra (gzipped postscript,
"External perturbations and structure of the beta Pictoris disk" ,
B.Sc. thesis by Anders Jeneskog (April 2001). A million-particle simulation of the famous dust disk's dynamics. Contains chapters on MPI, MPI vs. PVM, and more.
Additional reading on: Low-budget parallel supercomputation. Competing Linux clusters etc. (1999)