[Swift-devel] Re: replication vs site score

[Ben Clifford]

On Mon, 6 Apr 2009, Mihael Hategan wrote:

> You are saying this with the awareness of the fact that replicas are
> only sent after the prototype job sat in the queue (and didn't start
> running) for what is deemed _too_much_?

I'm not suggesting that we reduce any submission load to remote sites. I'm 
suggesting a different order for those submissions.

The queue delay is not so _too_much_ that we cancel the original on 
replication; and it appears (though I don't have stats on real runs) that 
the originals do run sometimes (though it would be interesting to know in 
real situations how often)

Given that, I'm suggesting that a better use of our load capacity is to do 
it with the ordering I suggested.

As far as I can tell, it will not result in slower runs. In the case where 
originals do run eventually, this should results in faster runs.

Thinking about it more, I can see a situation where a site is pretty fully 
loaded queuewise by swift yet never actually runs a job, because by the 
time a job gets near the front of the queue it has been replicated and run 
elsewhere. That's an extreme, but I think its the extreme of the same 
situation I talk about in my original message.

--