[Info-vax] Accuweather new contract

[johnwallace4 at yahoo.co.uk]

On Monday, 30 March 2015 07:42:25 UTC+1, JF Mezei  wrote:
> On 15-03-29 23:25, Kerry Main wrote:
> >
> > The autoconfig issue is well known and has in various environments
> > impacted VAX, Alpha & Integrity. I have seen it first hand in a number
> > of different sites.
> 
> My alpha was NOT on autoconfg. port forced to 100mbps full both on alpha
> and on switch.
> 
> My mac is set to have autoconfig on port.
> 
> Mac was able to significantly outperform the Alpha for FTP transfer over
> the same DSL line.
> 
> I do not know when the last time engineering actually worked on the TCP
> login in the IP stack (windowing acks etc).  And I do not know if double
> buffering really makes such a huge difference.
> 
> Since my DSL speeds are slower than the original 10mbps ethernet, you'd
> think even a DSL10L would have enough CPU "umph!" to handle double
> buffering fast enough that it wouldn't slow the link down. (at 100mbps
> or 1gbps it probably makes a difference).

TCP (specifically, ie not UDP) has a number of performance issues
which can show up if two valid but not entirely compatible
implementations talk to each other.

TCP over a less-than-perfect medium such as a DSL line can show this
up particularly badly in the presence of line errors leading to packet
loss.

This is not an issue of ultimate performance of the underlying stack.

It is an issue which used to be (relatively) widely known back in the
early days of mass market DSL, and used to have a decent writeup on
DSLreports (and very probably elsewhere). I can't find that writeup
right now. Basically, Linux webservers and NT clients didn't play nice
together as a result. Presumably this is long since fixed in NT, but
may not have been fixed in other OSes.

The magic keyword is "TCP slow start".