How about:

(Probability of picking peer X) := (Prov.BW of peer X) / (Sum prov.BW) 

Or here's a totally different idea. The client states a bandwidth floor and the 
server/service eliminates peers without at least that much provisioned 
bandwidth??? And no response from server/service if too few peers?

-- Rich


----- Original Message -----
From: Sebastian Kiesel <[email protected]>
To: Woundy, Richard
Cc: Sebastian Kiesel <[email protected]>; [email protected] <[email protected]>
Sent: Thu Mar 18 18:20:44 2010
Subject: Re: [alto] Comments on provisioned bandwidth and ALTO

Rich,

On Thu, Mar 18, 2010 at 05:48:04PM -0400, Woundy, Richard wrote:
> I agree.
> 
> Here's a slightly different example. The common tier for ISP A may have
> provisioned bandwidth of 12 Mbps, and 10 Mbps for ISP B. Assuming that
> an application doesn't need 11 Mbps, it may make more sense for the app
> to pick an ISP A client 12 times out of 22 (55%), and an ISP B client 10
> times out of 22 (45%), 

what you describe here sounds very reasonable. But how would you
implement it as an algorithm that works for arbitrary BW distributions
in the swarm?

My first thought,

(Probability of picking peer X) := (Prov.BW of peer X) / (Average prov.BW) 

would exactly cause the problem I described, if there were very many
peers with a BW slightly below the average and very few ones with a BW
far above the average.



I'm not saying that the provisioned bandwidth attribute is useles,
but I think we must be very careful in designing algorithms that make
use of it.


-- Sebastian
_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto

Reply via email to