[CentOS] Extending Network - OT -- PCs are poor wire-speed routers

Fri Jan 6 14:48:36 UTC 2006
Bryan J. Smith <thebs413 at earthlink.net>

On Thu, 2006-01-05 at 23:06 -0500, Alain Reguera wrote:
> Thanks for replay Bryan. Excuse me for my low knowledge level. I'll
> try to explain it.

I just didn't know what you meant by your terms.
Now I see you mean the subnet.

> Imagine you need to give service(mail, web, browsing etc.) to
> different institutions. Some institution connect using commuted lines
> and others directly through the main ISP router. The location of the
> node where all the servers and the main local router are, is inside
> one of these institutions. In this moment, the network of the node
> have a subneted C class range  and the local institution (where is the
> node place) is connected using a PC with 2 interfaces that connects
> both networks.
> At this time all is working, but new workstations are planed to arrive
> and we need to increase the number of stations in the local
> institution, so 254 PCs actually are not enough. So we are looking a
> way to extend or increase the number of possibles workstation.

You _could_ "supernet" Class Cs and increase your subnet mask.  E.g. /23
(255.255.254.0) will give you 510 usable addresses, /22 (255.255.252.0)
will give you 1022 usable, etc...

> I proposed the idea of create various networks and separate the local
> institutional services from the node, to make them independent one of
> another. So, connected to the main local router will be a switch, this
> will be the top level local switch where the node and the local
> institution will be.

If you want to segment, that will give you separate broadcast domains.
If you do that, you either want to have a very fast router on a GbE
port, or a layer-3 switch that does direct port-to-port after the IP
route has been established between 2 nodes (as well as offering a
dynamic routing protocol such as RIPv2 or OSPF).

The best, entry-level layer-3 switch I've seen is the Netgear
FSM7328S:  
  http://www.netgear.com/products/details/FSM7328S.php  

4xGbE, 24xFE for about $400 list.  They have a 52-port version (4xGbE,
48xFE) in the FSM7352S as well, but at that point ($800), you'd probably
want to look to a GSM7312 (12xGbE) instead for about the same cost
($900).

> The node is formed by various servers that will be connected directly
> to the switch. The main objective of the node is to administer mail
> accounts and RADIUS service (don't know it at all) and control
> browsing for the users connected.
> The institution is formed by a PC with various eth interfaces, one to
> connect to the router, and a serie of 192.168.1-2-3-...n.0 that permit
> us to connect 254 workstation for each one. Maybe will be needed more
> than 1 box here, think that the number of eth interfaces in a PC is
> limited. The main objective here are browsing, mail and web
> publishing.

You really want to _avoid_ using a PC as a router at wire-speeds.  It's
going to be very slow, unless you spend a _lot_ of money on a powerful
system, PCI-X/PCIe cards/channels, etc...

You're far better off going with a dedicated piece of equipment.  Not
just a router, but a layer-3 switch, which does direct layer-2 switching
at the MAC level after routes have been established between two ports.

If you still want to use a PC as a router, be sure to build your kernel
so it is optimized as a router (this is a selection in the networking
subsystem), and not as a host (which is going to be the default of any
kernel build).

But I really would recommend _against_ that, _unless_ you can guarantee
that 95% of the traffic is local to the subnet.

[ For the naysayers that might say that several layer-3 switches use
Linux, remember that these layer-3 switches have ASIC hardware that is
driven by the Linux OS.  A PC does _not_.  A PC can_not_ do what a
layer-3 switch can anywhere near as fast. ]

> That's it, don't know if I explain my self. Don't know if my idea is
> correct, just an idea of what I've been reading on Douglas E. Comer
> TCP/IP (my first reading about networks).
> Again, thanks Bryan for replaying.
> Any suggestion or idea of how make this will be strongly appreciated.


1)  Unless 95% of the traffic stays local to the same subnet, I would
_not_ use a PC as a router.

2)  If you don't need segmentation, then "supernet."  _All_ systems can
keep the same IP addresses, just their subnet masks need to change.

3)  If you want segmentation, but more than 5% of your traffic crosses
subnets, get a Layer-3 switch.


-- 
Bryan J. Smith   mailto:b.j.smith at ieee.org
http://thebs413.blogspot.com
------------------------------------------
Some things (or athletes) money can't buy.
For everything else there's "ManningCard."