[Swift-devel] Re: scheduler foo

[Mihael Hategan]

On Mon, 2008-07-21 at 11:58 -0500, Mihael Hategan wrote:
> On Mon, 2008-07-21 at 15:27 +0000, Ben Clifford wrote:
> > 
> > On Mon, 21 Jul 2008, Mihael Hategan wrote:
> > 
> > > > Secondly timer tasks are not killed, so a large number of timer tasks get 
> > > > started up for overloaded sites.
> > > 
> > > Oh, I see. They are started multiple times for the same delay.
> > 
> > yes.
> > 
> > I fiddled a little with cancelling timers before launching the next to see 
> > if it made the code better behaved. It does, though I'm still not 
> > convinced that its right. Also to do that in the present layout needs some 
> > kind of lookup table to figure out which timer to cancel which is 
> > increased complexity.
> 
> Right. The better strategy would be to keep track of waiting hosts using
> a O(1) op.
> 
> Let me poke around a bit today and tomorrow.

So one alternative is to have a separate thread that polls the sites for
changes in overloadedness. This is a certain type of constant time and
seems to work. I'll test this some more with various setups and see how
it pans out.

But I noticed that a number of times the site score is affected by
failure of tasks that we reasonably expect to fail. Such as the absence
of one of the stdout/err files. So I think it's an oversimplification to
change the score based on the type of a task. It should probably be
coupled with other things. Which I don't like because it breaks
abstraction. Which means it's a relatively broken abstraction.

> 
> > 
> 
> _______________________________________________
> Swift-devel mailing list
> Swift-devel at ci.uchicago.edu
> http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel