Date: Tue, 21 May 2002 13:52:08 -0700 From: Terry Lambert <tlambert2@mindspring.com> To: "Vladimir B. Grebenschikov" <vova@sw.ru> Cc: Oleg Chebotarev <chebotarev@yahoo.com>, freebsd-current@FreeBSD.org Subject: Re: multi default routes in freebsd !? Message-ID: <3CEAB378.D4B7F9E@mindspring.com> References: <20020521032746.14287.qmail@web11607.mail.yahoo.com> <3CE9DD37.88DDCABC@mindspring.com> <1021965925.557.9.camel@vbook.express.ru> <3CEA0FC0.E71EC770@mindspring.com> <1021977292.24428.30.camel@vbook.express.ru>
next in thread | previous in thread | raw e-mail | index | archive | help
"Vladimir B. Grebenschikov" wrote:
> > Multipath routing is not as useful as you imply. Neither is
> > round-robin'ing between a set of paths. It assumes that the
> > pool retention time on the router is longer than the drain time
> > for a single path, such that you end up with a higher aggregate
> > throughput than you would otherwise get. Most of the time,
> > with what you are suggesting, you will get the same throughput,
> > you will just get differential pipe utilization (using B == !A).
> > When this isn't the case, the amount of latency for a single
> > path is such that you end up with only a small fractional
> > improvement, when there is any improvement at all.
>
> Lets imagine - we have 3 links 2Mbit/s on different interfaces.
> I want to join them all, but I have no control of other end (provider)
> so I can't build netgraph-joiner.
>
> Solution with installing 3 routes (through BGP of course, one BGP
> session per link) solves problem.
>
> I have 6 Mbit/s summary bandwith.
No, you don't.
Without the cooperation of the tother end, you don't have
control of the symmetry of the return route. So maybe
your packets are round-robin'ed out interfaces, but they
all come back through the same interface, because you have
no control of the other end.
So if you have interfaces, all with equal throughput, and
your traffic load is bidirectionally symmetric, then you
get:
total unidirectional
interfaces throughput
1 .5
2 .66
3 .75
...
The only way around this is to have the default routes outbound
be totally seperate from the inbound (don't eat your inbound
bandwidth witho outbound load). The unidirectional troughput
goes to "1", and your total throughput doubles. But it never
gets beyond double.
This is why BGP is a better deal: it implicitly enlists the
cooperation of the other end of the link.
> > The primary failure of this is that it can't detect when a
> > route goes down, so you are screwed when that happens.
>
> If interface goes down route will be DOWN by kernel.
> So it is not problem.
No.
,---------------.
| BOX WITH TWO |
,-------| DEF. ROUTES |-------.
route #1 | `---------------' | route #2
| |
,---------------. ,---------------.
| ROUTER "A" | | ROUTER "B" |
`---------------' `---------------'
| |
| |
good link dead link
The link between the box and "router B" remains up. Therefore the
box fails to note that packets sent via "route #2" never get to
their destination.
It's ridiculous to think that the interface for "route #2" will
somehow "magically" down itself on "box" because, several hops
down the line, the line is dead.
> Anyway if problem happens without downing interface BGP will detect
> problem and down routes.
Only if you are using BGP. And if you are using BGP, you don't
need the hack you want, BGP will take care of it for you.
> > You are much better off using BGP.
> >
> > If you absolutely refuse to use BGP for some reason which you
> > absolutely refuse to post to the list, you should consider using
> > PPPOE and multilink PPP in combination (both are Netgraph nodes).
>
> It is usual practice to use, say OSPF for internal routing (inside one
> AS). Yes I understand that some netgraph solutions can help
> (multilink PPP or ng_one2many, I am extensively use netgraph)
> But if routing protocol there solution for it with alid link down
> detection and so why we need to use some "workarounds" to emulate
> protocol behavior ?
You need to read the "README":
ftp://ftp.flirble.org/pub/unix/hacks/FreeBSD/README.MPATH
Specifically, you need to read:
WHAT IT DOES NOT DO
It doesn't detect when remote hosts are down. This is not
the job of the kernel. It's not a routing protocol, it's
not an automatic failover system.
So it is not a "routing protocol solution" in any sense of things.
> > There is also a VRRP implementation for FreeBSD. I've posted
> > the URL for it before. In combination, Virtual Router Redundancy
> > Protocol *and* multipath are, together, roughly equivalent to
> > using BGP (assuming both your routers are running the VRRP code).
>
> No, VRRP can't help if you want to use summary bandwidth, but helps a
> lot if you are want to get redundancy (I think so because I am using
> VRRP on my core routers since 4.2).
I don't think you can get what you want without the cooperation
of the other end of the link. If you want FEC, then use Bill Paul's
FEC code, which does the channel bonding that you seem to want. But
you will need the other end to cooperate.
> BGP can use multipath, as well as OSPF.
> Possibility of kernel to store some number of routes for one prefix in
> FIB can't replace BGP and, on other hand, BGP itself can't replace such
> kernel feature.
I already said that I think the code should be committed to the
main line kernel, and preserved.
> These two things almost orthogonal.
Yes. They are. BGP solves the problem you are trying to solve,
while the multipath is useful, but doesn't solve your problem.
> PS:
>
> My opinion - it is useful feature for FreeBSD kernel, often used now
> as good routing platform.
I already said that I think the code should be committed to the
main line kernel, and preserved.
I just don't think you're going to get out of it what you think
you will get out of it.
-- Terry
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3CEAB378.D4B7F9E>