Firewall Wizards mailing list archives

RE: IPSEC over load-shared T1s (per packet)


From: "Ben Nagy" <ben () iagu net>
Date: Sat, 20 Sep 2003 05:51:54 +0200

I think this is pretty much solved now, but just for the sake of the
archives:

The problem was pretty much as I guessed (just lucky ;).

The packets were being sent over alternating links in strict round-robin,
which meant that the ESP packets sometimes arrived out of sequence. The
IPSec implementation was dropping all the ones with seq < currentseq, which
was causing retransmits in the tunneled TCP sessions.

One fix is to use "per destination" load balancing - but that is bad because
if all the traffic is VPN then only one link will get used (only one
destination).

What I suggested offlist is to look at either ppp-multilink, or MUX/DE-MUX -
both of those will make the link look like one big layer2 pipe, which will
fix the problem and preserve sequencing. PPP Multilink is software, and
simple. MUX stuff is more complicated but faster and can be more flexible.

I also got queries offlist about the E1/T1 RJ connectors. Yes, I did, OK? I
was curious. Ow.

ben

-----Original Message-----
From: R. DuFresne [mailto:dufresne () sysinfo com] 
Sent: Friday, September 19, 2003 5:32 PM
To: Ben Nagy
Cc: TSimons () Delphi-Tech com; firewall-wizards () honor icsalabs com


I might be reading Ben wrong, but, I get the impression he is talking
about session concerency?  A 'sticky' bit in the load 
balancer end, such
that a session started and sent via one t1 remains directed 
through that
path for the remainder of the session?

Thanks,

Ron DuFresne

On Thu, 18 Sep 2003, Ben Nagy wrote:

ObBOFH: One of the T1 RJ connectors must be dirty, which is 
causing packet
corruption. Give both the telco jacks a good clean (licking 
them works well)
and see if that fixes the problem. [1]

Seriously, I do have a theory ;)

Does this routing guarantee to preserve sequencing? 

If it's really as you described (packets send one for one 
via alternate
links) then you have some potential problems brewing, I think.

TCP will "work things out" when packets arrive out of 
sequence, but with
IPSec it's left up to the implementation. One security 
concern with most
crypto things is replay protection. IPSec addresses this by using a
mandatory sequence number in the ESP header. The receiveing 
IPSec doesn't
_have_ to take any notice, but most do. If your receiving 
IPSec has enabled
replay protection then if one link is going faster half the 
packets are
going to get dropped (sequence number < current).

This would make your tunneled protocol (say TCP) do the 
retransission thing,
so it would work itself out eventually, but the speed would 
indeed suffer
horribly.

See if you can convince your router to preserve "IP flows" 
and use the two
links in a more sensible manner. That might help.



Best of luck,

ben

PS: Let us know when you work it out? This is an interesting one.

[1] The RJ's are live, for non-network-engineer types. Not 
enough to kill
you, but it hurts. :)

-----Original Message-----
From: firewall-wizards-admin () honor icsalabs com 
[mailto:firewall-wizards-admin () honor icsalabs com] On Behalf 
[...]
We setup IP LOAD-SHARING PER-PACKET on each of the serial 
links on both
sides (NOC and Us) in order to get an aggregate 3.0mbit. 
PER-PACKET routing
alternates usage of the T1s, one for one...

Since then, VPN performance has taken a dive.  Sniffing out 
traffic, ESP
packets are sent 3-4 times before they can be properly decrypted. 

_______________________________________________
firewall-wizards mailing list
firewall-wizards () honor icsalabs com
http://honor.icsalabs.com/mailman/listinfo/firewall-wizards


Current thread: