[Swift-devel] karajan scheduler hack

[Mihael Hategan]

On Thu, 2007-11-01 at 23:23 +0000, Ben Clifford wrote:
> Recently, I've been making runs with Andrew with a scheduler hack to stop 
> karajan's site score go below -10.
> 
> This has been useful in stopped clustered job failures from causing 
> catastrophic slow down.
> 
> Its not clear what if any easy change can be made that isn't so hackish to 
> achieve this same benefit.

I think that would be a reasonable hack for now.

However, I do think that the problem is the details of the algorithm not
being well thought out. In principle, I think it should remain a
feedback system (also given that the opportunistic scheduling for VDS
paper pretty much did the same). Given that, and given that with the
current assumptions all the feedback inputs are accounted for, I'm led
to believe that this is a matter of properly specifying the feedback
function. But I'm fuzzy on many things.

>