[PLUG] Looking at bandwidth optimization schemes...

[Kyle Hayes]

On Wednesday 27 August 2003 11:14, Michael C. Robinson wrote:
> What is the advantage of a fixed packet size as used in ATM?

There's a whole set of "lore" about this.  ATM was supposed to be able to be 
much faster than other kinds of transport.  You can see some of what they're 
doing from a few things:

1) small packet size (Europeans wanted 32 bytes, US wanted 64 bytes, so we all 
got 48 bytes).  This is something of a rumor (48 vs 64/32), but I'm fairly 
sure I read this in Tanenbaum's networking book.  Could have been somewhere 
else, but I think it was in Tanenbaum.  They wanted ATM to be used for 
telephony as well.  With that, you don't want big packets.  You want packets 
that are as small as you can get and you want latency that is as low as you 
can get.

2) fields in the packet are designed for "on the fly" switching.  An ATM 
fabric can be built with very dumb, but very fast, switches.  Each address 
comes in most significant bit order and can be used to trip binary (i.e. 
left/right) switches as the packet data flows past (seen the hardware to do 
it, this was the original intent).  The idea was that you'd be able to switch 
from network to network at least as fast as the data was transmitted.  I.e. 
there was less than a packet latency (sometimes much less) through a switch.  
This makes for very low latency compared to IP-based networks where you have 
to buffer the whole packet (to an extent) before you send it on.

3) the fixed packet ("cell" I believe is the ATM terminology) also apparently 
helps keep latencies low.  Every packet is the same size, so buffer sizes are 
just multiples of a packet size.  I only have vague memories of why this was 
a good thing.  Something to do with buffers and cheaper hardware as well as 
helping with latency.  Not sure if I remember this correctly.

ATM has a lot of really interesting ideas in it, but it never hit the volume 
needed to make it cheap.  With gigabit ethernet at a buck a port, there's 
really no point to ATM any more.  Remember FDDI?

> For real time computing would you only accept inbound connections
> at a certain rate of speed for certain protocols or do you have
> to control both ends of the link?  If packets are dropped they
> get retransmitted wasting more bandwidth, right?  What causes
> tcp, icmp, or udp thrashing to occur?

TCP packets are retransmitted.  All the others are "fire and forget".

> In gnome ftp you can lower the transfer rate where I've noticed
> that dropping it to the speed of my ADSL when I'm going online
> reduces the number of collisions on the hub.  I'm going
> to get a switch at some point :-)

Switches are good.  I run most of my internal network at 10Mbps simply because 
I've got the hardware (i.e. paid for) and I rarely have much traffic between 
machines.  Nearly all network traffic is out the cable modem.  Thus, I'm 
limited to whatever Comcast decides to give me today.  That's a lot less than 
10Mbps and will be for a while.

> The purposes for ISPs are: distributed management of ip addresses,
> theoretically worm virus and spam control, and lastly sharing of
> high cost TI/E1/OC3/fiber etc. lines between multiple subscribers.
> IP addresses, even blocks, used to be free.  Nowadays a single IP
> is $5 or more a month.  

IP addresses are a scarce resource now.  There's less than 2^32 total (some 
ranges are completely removed).  With IPv6, this problem will go away until 
someone figures out how to use sparse addressing for some really interesting 
problem.  Then we'll be back to scarcity.

>ISP's are not required to advertise
> compatibility information about their email offerings or even
> provide spam control that their end users can take advantage of
> which will actually reduce real spam.  That's not all, the cost
> of bandwidth is wildly variable even from a single provider
> over the areas that provider serves.  

Wholesale bandwidth prices are often below cost.  When the Internet dot.bomb 
hit, a lot of companies folded and those that remained picked up dark fiber 
and network centers at pennies on the dollar.  A friend of mine bought a 
fully equiped NOC room (racks, raised floor, network patch panel, huge UPS, 
AC etc.) for $25 from Worldcom because they were abandoning the building his 
office was in.  It was $25 because that is what he had in his pocket when he 
talked to the Worldcom rep.  Really.  Reality is stranger than fiction.

So, don't take the weird bandwidth costs as dogma at this point.  A lot of 
stuff is filling up, but there's still a lot of dark fiber left.  The low 
hanging fruit's probably gone though.

> The bandwidth cap seems
> to be wildly variable also, I believe 800k downstream 512k up
> is the maximum a single ADSL line in Scappoose can do.  St.
> Helens residents evidently can go faster on lines that are
> of arguably lower data quality.  If you could get just DSL
> on a second line it wouldn't be so bad, but DSL isn't sold
> as a seperate service from POTS.  I know they could sell it
> that way in Scappoose because they've turned POTS off before
> without turning off DSL.  The only thing I'm wondering is
> if there is a bottleneck from everyone coming in from as far
> away as Astoria artificially lowering the speeds that can be
> offered in Scappoose.

I doubt it.  Often the larger cities have worse bandwidth because the 
equipment is older, there's less of it, and things were not wired for the 
needed capacity.  In the rural areas (if you can get broadband), you might be 
one of three people on a fairly large pipe to a POP that was put in one or 
two years ago.

For DSL, you have all kinds of factors.  When was the copper put in?  What 
kind of DSLAM do they have?  When was the POP built?  What is the backhaul 
line to the main trunk?  Bad coppy will cause DSL to be crappy.  I had a DSL 
line once that was really good, but my neighbor had poor service.  I had new 
copper run less than five years before and he was on one of the original copy 
lines from 1957 when the development was built.  His DSL modem was constantly 
dropping and retraining.

There are so many factors involved that it is pretty difficult to say that any 
one thing is the key factor.

Best,
Kyle