Re: [p2pi] Information in an ALTO protocol

The cost of bandwidth testing isn't terrible, but the accuracy of measurements taken on 1K transfers is very low - you're really measuring latency and buffering, not sustained data transfer rate. Of course, there might be some relationship between them, but (at least in our measurements) it took much larger data transfers to get even remotely stable results. In addition, measuring actual throughput at a point in time can be affected by other traffic on the PC (e.g. PC's do many things when they're starting up), other nearby PC's (e.g. someone else in your neighborhood is busy right now), transient interference on the internet, etc. So while it's not unreasonable to do such testing, and we certainly do so, an ISP provided provisioned capacity would be much more accurate (when it's available).

- Laird Popkin, CTO, Pando Networks
  mobile: 646/465-0570

----- Original Message -----
From: "Nicholas Weaver" <nweaver@ICSI.Berkeley.EDU>
To: "Laird Popkin" <laird@pando.com>
Cc: "Nicholas Weaver" <nweaver@ICSI.Berkeley.EDU>, "Ye WANG" <wangye.thu@gmail.com>, p2pi@ietf.org
Sent: Monday, September 15, 2008 10:38:41 AM (GMT-0500) America/New_York
Subject: Re: [p2pi] Information in an ALTO protocol

On Sep 15, 2008, at 7:22 AM, Laird Popkin wrote:

> If a p2p application knows roughly what bandwidth the user is  
> provisioned with, it can set reasonable defaults.
>
> For example, Azureus' configuration wizard asks users what kind of  
> uplink speed they have (e.g. dial, 128K, 256K, ..., 1M) and uses  
> different application level values (e.g. max upload speed, number of  
> active transfers, number of active peer connections) appropriate for  
> that kind of network connection. If they could automatically  
> configure themselves rather than asking the user a rather  
> intimidating (to a non-technical consumer) question, that would be  
> better for the users and the network.

If you control a well-connected server, you can probably do this  
pretty easily (at least into the right ballpark) with just a few sets  
of back-to-back UDP packets to estimate uncongested-link bandwidth: at  
least to get

If you do this on startup and IP address change, use 8, 1000B packets  
(4 pairs), and have 10M clients a day doing a test, this would  
require, roughly, 80 GB of upload bandwidth a day.

Given the cost of EC2 as a host for such a test server, this would  
cost a company <$20/day in bandwidth to support 10M clients a day.

_______________________________________________
p2pi mailing list
p2pi@ietf.org
https://www.ietf.org/mailman/listinfo/p2pi