Satellite latency (fwd)

Wed Feb 27 17:09:01 UTC 2002

Ukyo Kuonji wrote (on Feb 27):
> Looking at a spam I just received from satcast.com, it looks like they are 
> considering the distance from the dish to the satellite to be roughly 44000 
> miles.  That's roughly  70000 km.  Assuming that they are stating that from 
> the dish to the bird, that would account for the larger RTT times that 
> people see.

We ran a similar service a number of years ago - which we managed to finally
put to bed about a year ago. We had an issue that because most peoples use
was HTTP, with short-lived connections, they never had a chance to open
up their windows and make full use of the satellite link when empty and
wouldn't shrink fast enough when busy. Dropped packets, retry galore,
only making matters worse.

I overcame this with a really nasty hack to a BSD kernel. Ran a HTTP proxy
on it and made the TCP stack bias the start window size of a connection
 and how quickly it changed depending on the measured load of the 
uplink to the bird.

People using this proxy had far fewer dropped packets and hung TCP 
sessions. It needed a ton of memory to do it and a lot of this had to
be assigned to the kernel for network buffers, but it worked. We
had window sizes in the region of 512k when the uplink was empty and
atmospheric conditions didn't drop packets.

> My real problem with these directpc type systems are two fold.  The first is 
> that these systems use a USB connection.  I use wireless at home since I 
> have a number of laptops.  I certainly am not willing to tie myself down to 
> a usb cable.

This was a DVB card in a PCI slot. Hauppage do similar these days I think.

> The second problem is that the users are addressed using rfc1918 address 
> space and are translated via nat to the Internet.  This may mess up any 
> number of applications that I use, including VPN software.  Add to this the 
> first problem, and suddenly I have two levels of NAT that I have to contend 
> with, one on my windows machine connected to the USB cable (with Internet 
> sharing), and one at earthlink's edge.

IMO, 1918 sucks at the best of times and you shouldn't be using it for
customer addressing unless requested.

Our system was cute in that you could control whether to use the satellite
link for the toward-subscriber path or not, plus you could dial any number
in our RAS pool, PSTN or ISDN, that you were authorised to use and it would
work.

> Those are what's keeping me at dialup, even though I only get 28.8 at best.  
> Well... that and the fact that I can't get cable where I live, and the CO is 
> sonething like 15 miles away (not that they can spell DSL out here).

I think I'd move. :)

Regards,
Chris.
-- 
== chris at easynet.net                                    T: +44 845 333 0122
== Global IP Network Engineering, Easynet Group PLC     F: +44 845 333 0122