Re: [SSI-devel] Process Loadleveling thoughts

[Aneesh K. K.V <ane...@di...>]

On Fri, 2003-08-15 at 01:38, Walker, Bruce J (HP) wrote:
> Demoing at LinuxWorld last week prompted me to think about a few things
> we might do to improve the process load balancing.
> 
> Things that weren't completely ideal: 
>    a. sometime several nodes would send processes to a node that was
> underloaded and perhaps swamp him; 
>    b. rexec load leveling doesn't kick in (for the demo) because it
> forks a lot of processes locally, they all sleep and wait for a signal
> and then all exec  (so at the time of the execs the node is not loaded.
> 


I guess the first thing (a) can be solved by allowing
the node to reject the process getting migrated into
that node depending on certain checks. But i am not
sure whether adding a delta to fork is really going to
help. The demo is a typical case where the intention
was  the demonstration of load balancing. I would say
the actual real world example to show rexec load
balancing would be a parallel build of huge source
code like linux kernel on a heavily loaded machine.
Each compile should happen on different nodes. 

I think making fork account for the load of the node
can result in lot of wrong load calculation. In the
typical case where all the forked process sleeps and
do exec later the load balancing should be actually be
done later by the load leveller and should not happen
at the exec time.  


>    C. perl or shell scripts that you want to load level can be a little
> tricky because the program name being run is actually perl or sh (the
> demo is actually a perl script).
> 

How is this going to be fixed. This seems to be tricky
?. 


-aneesh