Changes by: Jason Hickey (jyh at cs.caltech.edu)
Date: 1998-08-21 15:22:55 -0700 (Fri, 21 Aug 1998)
Revision: 2442
Log message:
Added distributed refinement using Ensemble.
This is an initial version. The scheduler needs some performance
tuning, and Ensemble needs a little work on blocking IO.
By default, Ensemble support is not compiled into Nuprl-Light.
To enable distributed refinement, you need a copy of Ensemble,
which is available at:
http://www.cs.cornell.edu/Info/Projects/Ensemble/
When Ensemble is installed, set the environment variable
ENSROOT to the root Ensemble directory. On mojave, I've precompiled
a version of Ensemble in ~jyh/nuprl/src/ensemble. Once this
variable is set, make will build the distributed refiner.
Ensemble uses a separate "gossip" daemon to get processes to form groups.
The program editor/ml/nlgossip starts this daemon. Once the daemon
is started, multiple copies of nl will know about each other, and ship
refinement jobs to each other. So if you want faster refinement, start
multiple copies of nl. These copies can be started and killed at
any time; Ensemble provides support for failure detection and recovery.