[ NNSquad ] Re: BT [UK] calls for action on net speeds

[Andrew Burnette <acb@acb.net>]

Interesting article, as I know BT peers at linx with numerous folks. I 
would like to hear more about the test criteria.

One thing that we (as techies and so on) even forget is that single 
stream (session, tcp or otherwise) measurements of bandwidth are not 
typically reliable.

MS windows up to and including windows XP, default IP stack tuning is 
best suited to a 10Mbps shared (CSMA/CD) Ethernet low latency LAN. 
Through a few minor settings changes, one can improve net performance by 
20% to as much as 50% depending on the usage profile of the user. (e.g. 
you live on the east coast, but more of the websites you visit are in 
the EU, or on the west coast, or you're 75% metric is less than 40ms 
round trip).

Also, another limiting factor is the asymmetric nature of xDSL (and 
cable) with "normal" (no p2p involved:-) ) activity, can lead to delay, 
jitter, and dropping of packets which control the overall sequence. If 
the sending end doesn't receive the next request for packets, they won't 
be sent.  I had unending problems with VoIP and comcast when I would 
simply send a 100k email while on the phone.

A decade ago, at AT&T worldnet, we built out a large dialup US 
nationwide network. In direct response to traffic gaps (sub 500ms) our 
own 'Customer QoS' measurements in the network indicated, we adjusted 
our servers to provide a different window size, and a smaller MTU (IIRC, 
we settled at 576bytes or so).  What this did for the typical dialup 
customer, was to fill the gaps in their download of email or usenet 
(remember usenet?) and effectively doubled throughput to our users, at 
no additional expense to us.  Interestingly, I do similar tuning for my 
own servers to either limit (according to bandwidth contract costs and 
conditions) or speed up responses.

Similarly, your chosen settings (tuned or not) and the default 
algorithms in control of your TCP stack significantly effect overall 
performance and throughput. (note the default TCP algorithm in linux 2.6 
kernel alone has been changed three times that I recall).

Try any of the speedtest sites out there from different computers (OS 
type, or IP stack tuning) and you'll see statistically significant 
performance differences between them on identical test criteria.

Best regards,
andy

Russell Smiley wrote:
> http://news.bbc.co.uk/go/rss/-/1/hi/technology/7292932.stm
>
> "BT Wholesale, which supplies eight million people, said many customers
> were disappointed by the mismatch between advertised and actual speeds.
>
> An independent survey found that 15% of people who bought eight megabit
> per second packages actually got the speed.
>
> The firm said regulators needed to agree rules about how broadband
> speeds could be sold to the public."