DragonFly kernel List (threaded) for 2003-09
Re: checkpoint/restart milestone 1
MPI programs that run on clusters can have Fault Tolerance
in the form of checkpoint/restart. This technology has been around
for a little bit but I don't think I have seen it at a general
level before :).
I guess I am just pointing out that this sort of thing can be done
in a distributed manner.
Kip, could you brief us on what this `checkpoint/restart' stuff
exactly means? I did not go to the Con, so I my guess is that
this is like a software suspend kind of a thing, where the state
of the system is saved, and it can be later resumed in the exact
condition -- although I maybe wrong. :-)
You're right, except its only at the application level, not the whole
I'll give you an example:
You have a compute-bound program that can run for weeks. The program,
like many scientific applications was not well structured, i.e. it has
a lot of implicit state sitting around in various globals and statics
here and there. Hence, having the application programmer save its state
is out of the question. However, you don't trust your hardware, so
like to be able checkpoint the program in an application-independent
way. This provides you with functionality for doing that. So if your
program/computer crashes the program can just restart from the last
it did a checkpoint.
This could also be useful for debugging as well. If you have an
application that starts doing something weird, you can just checkpoint
it and send the checkpoint off to the developer.