> > The first problem I have is that my chunks don't chunk right.....
> > Basically, I
> > get multiply responses, etc. This is more of a technical question than
> > one of
> > algorithms, but I figured I'd give it a chance.
>
> I'm not sure what you mean here...
It's a problem with my code, long story... doesn't really have anything to do
with this.
> You might want to investigate using a bit more of a client/server
> approach -- change your logic so that the server sends out an initial
> round of work to each client process, and then waits for anyone to
> reply with their results. When it gets a set of results from someone,
> it gives them the next round of work. In this way, naturally faster
> machines will request work faster, etc.
But in MPI_Recv, I need to indicate which node to recieve from. Is there a way
I can receive from any, and it will let me know which one, so I can later
tell which to send to?
> You can even adaptively set the chunksize on the fly, based on how long
> a process took to return results to the server, etc. You'll need a bit
> more instrumentation (e.g., record the start and end time for each
> chunk on each process), but it's certainly do-able.
Hey I'm still a newbie here ;-)
Thanks,
-Jack C
jack {at} crepinc.com
http://www.crepinc.com
|
