Two more possibilities:
Yet another possibility would be to modify the NAT / MASQ layer to
do the load balancing, doling out new connections over available
paths. At this point we're doing a routing project and not a modem
project. There are lots of points at which the policy could be
asserted, unless we're going to assert the policy with the tool that
combines two modem streams into a single connection, linux-ppp is the
wrong list: CC and reply-to set to linux-router.
Once ppp connections are up, their loadedness needs to be readable by
the route chooser, since they all have their own IP addresses (since
they're dialed to different ISPs, who have no clue that this is even
happening) so new connections need to get assigned to the less loaded
modems. This arrangement will break "sessions" with broken http sites
that use IP address as a session marker instead of cookie. And routing
cannot be done per packet it must be done per connection. End-to-end
load sharing works because the software on each end can break up the
traffic in smaller pieces and reassemble it at the other end.
With cable modems appearing at about the same cost as multiple regular
modems
there wouldn't be any big fiscal motivation in setting up a central
reassembling station somewhere with good network connectivity and
running massive equalization endpoints and more NAT there -- but you
could
do it on a small scale I guess, using multiple instances of existing
"roaming" software -- That wouldn't break IP-based session models.
Yet another other possibility, which would work only with "power users"
who are able to change the internal IP addresses on their own machines
themselves, would be to arrange three subnets and have each subnet
handled
by a ipchains rule involving a different modem, and let the users do
human politics concerning which of the three modems they are going to
be associated with -- if their boxes can do load balancing of multiple
interfaces based on response time they could even set their NIC to work
on their addresses in all three subnets, which would translate to use
of all three modems, and the loadbalancing would be pushed all the way
back into a "selection of IP address on machine with multiple IP
addresses"
situation. I don't know how different IP stacks and different OSes
handle
this one. Does the "default address" portion of the C<socket> call in
glibc2 already do this? Some software takes switches to select one of
multiple available IP numbers, but I'm not aware of any that will
load-balance
at the kernel level based on varying performance.
Joe Tennies wrote:
>
> I've been toying with that myself. One solution I found was
> running multiple SOCKS5 servers and then using the HUMMINGBIRD
> SOCKS5 Client and enable balancing on it. I haven't actually
> done it myself yet, but it may work. (It does in theory as long
> as you can get a socks5 server for each modem)
>
> Joe Tennies
>
> -----Original Message-----
> From: John Krivitsky [SMTP:[EMAIL PROTECTED]]
> Sent: Wednesday, November 03, 1999 2:16 AM
> To: [EMAIL PROTECTED]
> Subject: Multiple modems without eql?
>
> There are a few docs on how to do load balancing between two machines
> that run load balancing protocols, but does anybody know of a way of
> performing a sort of equivalent when connecting to a remote that doesn't
> support the protocol, or even to two different remote-ends?
>
> The application is for a group of people to access the Internet via a
> single dial-out box that holds multiple modems. The goal is to connect a
> larger number of users through the modems - doesn't matter that any one
> user still only has the maximum bandwidth that a single modem allows.
>
> In other words, say there are 9 people and 3 modems. It might mean that,
> on average, a single person is sharing the modem with 2 others, not with
> 8 others.
>
> Any ideas?
>
> Thanks!
>
> jk
______________________________________________________________
David Nicol 816.235.1187 [EMAIL PROTECTED]
End Daylight Savings Time in our lifetime
-
To unsubscribe from this list: send the line "unsubscribe linux-ppp" in
the body of a message to [EMAIL PROTECTED]
