interesting expert problem - shaping over VPN
Here's a challenging problem for you experts to tackle:
I'm trying to shape traffic going into an IPSEC interface which then goes over a DSL PPPoE interface. I figure I need to shape the DSL interface to keep it's hardware queue mostly empty, and to prioritize between IPSEC and non-IPSEC traffic. I also have to shape going into the IPSEC, which carries VoIP (high pri), VNC (med pri) and other (email, etc, low pri).
I have it all set up and working, except that the IPSEC shaping doesn't seem to do any good whatsoever. Even if I allocate 99% of the bandwidth to the VoIP and 99% to IPSEC over PPPoE I still get break-ups in the VoIP signal when I do some heavy VNC. I tried such drastic things as reducing the "ceiling" to half of what the DSL line was spec'd (and tested) as supporting. I played with the numbers until they were really skewed (99%) in favor of VoIP, but still no joy.
So my question is, am I missing something fundamental conceptually regarding shaping traffic into an IPSEC/VPN interface and then shaping that along with non-VPN traffic out over the single internet connection? Is there some buffering/queue stuff in IPSEC or PPPoE that would prevent me shaping properly?
I am using FreeSWAN IPSEC on Fedora Core 1.
Thanks for your help.
_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
hi trevor,
Quoted 7 lines · Trevor Cordes
Quoted 5 lines · On 18 Sep, [email protected] wrote:
hi,
there was a thread on this recently -- please search the archives for
"traffic queueing and ipsec vpn"
Quoted 8 lines
Ya, I had seen that. I just reread the thread and it doesn't really help me with my problem. It's all conceptual with no specifics, and the concepts appear to agree with my knowledge and current configuration attempt.
The only thing that puzzles me a bit is this talk of INGRESS and EGRESS, which I don't recall being in the HOWTO's and I'm not really sure of what signifigance they are.
basically, ingress is more difficult to control and to granularly regulate traffic as we have no control over what's coming in and in what order. i have seen studies that indicate RED as an effective way of handling ingress.
I just wish I could be sure that that really is the case as I feel like
Quoted 8 lines
I'm real close to a solution. The filtering is working great, passing packets into the proper QDISC. It just doesn't appear to help the VoIP at all.
Of course, it doesn't help that there's that kernel panic bug in HTB into ipsec (https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id0172)
Thanks for your help.
egress on the other hand is completely under your control as you select in what order and with what speed packets are dequeued to the hardware. in slowing packets to be dequeued, tcp's AIMD comes into play -- Additive Increase Multiplicative Decrease, that is, tcp ramps up speed until a packet is lost or until an ACK takes longer than the congestion window. at that time, tcp multiplicatively decrease speed (cuts it in half) and then starts to ramp up again until such time as tcp feels that it has obtained optimum throughput.
Quoted 5 lines
I'm starting to think perhaps my problem is not necessarily in shaping stuff into the VPN, it's shaping everything out over the ADSL connection. I read somewhere that a 128k upload ADSL connection will take 40ms to transmit a max-size packet. So shaping becomes pointless if 40ms is too long for the VoIP to handle as a delay.
i think that you may be getting a bit confused -- in a simple lan/adsl environment, there are two ingresses and egresses: ingress coming in on ppp0 for example, and egress leaving eth0 towards the lan. similarly, there is ingress on eth0 as packets come in from the lan, and egress on ppp0 as packets are dequeued towards the adsl. i think that you should try placing egress on your ppp0 to classify packets and priorities in such a way that they are dequeued in a manner that corresponds with your needs.
<snip>
cheers
chalres
_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
whoops,
oh yea -- ipsec and that whole eth0 thingy :-/ no wakeup on sundays.
are you running your tunnel exclusively for voice, or are there other protocols as well ( i.e. contention).
cheers
c
Quoted 6 lines · Trevor Cordes
Quoted 4 lines · On 26 Sep, [email protected] wrote:
hi trevor,
perhaps others can chime in on their experiences and configs that have shown promise for voip.
Quoted 8 lines
OK, I'll keep watching/posting.
Quoted 3 lines
i would think that you would want to egress on both interfaces to insure that the packets you want pass first. i thought that you had both already.
I have egress on ipsec0 and ppp0 but not eth0. It's actually 3 interfaces total, which is why I hate PPPoE... it's so over complicated.
_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
hi trevor,
well, if you're controlling whats going over the vpn then there are several options:
i've been playing with racoon lately (well longer then with freeswan) so i'm not so sure with ipsec, but ... it appears that the meta-data ( i.e. packet marking) is perserved on packets that have not yet been encrypted but are going to be.
as a general strategy, i would mark packets with different marks depending on what the payload is -- maybe something like 0x1 for voice, 0x2 for smtp, etc.
then use these marks on the public interface to egress them towards the internet in the highest priority.
mark the inbound packets coming off the internet (once they've been decrypted) and place them in highest priority (depending on their type)
this wouldn't be too bad -- in fact it's about all you can do.
Alternatively, and with more complexity, open up several tunnels with different spi's -- pass traffic into tunnel by type -- this would allow you to know what an encrypted packet was carrying without having to decrypt it. cool, but i'm not sure that it would help much.
anyone else done this??? tcng files are great (hint :-)
cheers
charles
_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/