[mat-wg] Ambient measurements

On Fri, 18 Oct 2013, Richard Barnes wrote:

> Nonetheless, I continue to think that collecting ambient measurement 
> data is a really interesting idea, so it would be cool to hear what 
> others here think.

Hi!

I thought I'd try to write down my thinking on the matter so far...

I first started thinking about this when I was fault finding VoIP problems 
10 years ago. Back then, as an extra feature on the VoIP system there was 
available to add to the CDR (call data record) some IP characteristics for 
the duration of the call, packet loss, reordering, PDV buffer misses due 
to excessive delay, and such. This meant that if the user was expericing 
quality problems, the transport network could either be confirmed to be 
causing the problem, or ruled out. This was excellent.

Now, on to what I mentioned yesterday at the mic. On every TCP session, 
the host TCP stack has a lot of information available to it, it's got RTT 
measurements, it knows packet loss, reordering, PDV and other packet 
behaviour characteristics. So instead of having the user run active tests 
and waste capacity in the network (and potentially have ISPs game the 
system by giving higher priority in the network to the test traffic), why 
not keep track of the regular traffic instead?

My idea was to continously collect TCP packet characteristics on existing 
user traffic, and try to aggregate the information as some kind of quality 
measure, that can be used for fault reporting or other statistics.

There are several questions, thanks Richard for bringing some of them up 
yesterday in the bus that I didn't originally think to bring up initially, 
but let's do it anyway!

1. What data should be collected per TCP session.

Let's start with the basic ones and name a few, which would be 
source/destination IP address, number of packets transferred, #packets 
lost, #packets reordered, average/max/min RTT.

2. How should the data be aggregated?

Considering what paths the packets take, the will experience different 
network behaviour. So destinations need to somehow be put together. By 
AS-number/address block at the other end might be a good way.

3. What should be done with the data?

Presented to the user somehow? Sent to a statistics analysis organisation 
if the user permits it? To the customer ISP in case the customer wants to?


Anyone got some bright ideas or can point to work already being done? I 
would like to see this going into every TCP/IP stack on the planet in a 
uniform way.

-- 
Mikael Abrahamsson    email: swmike at swm.pp.se