[p2p-hackers] UDP Behavior in the Real World

Pancrazio Auteri <Tvblob> pancrazio.auteri at tvblob.com
Sat Oct 21 19:01:55 EDT 2006


Hi,

we have some video streaming systems used over the Internet.
We use RTP over UDP transport then RTP headers gives us the ability to
collect info about packet loss, reordering and duplicates.

In the fiber-based Fastweb network we observed a very low rate of
reordering: the only statistic that I have is related to just one customer.
It says that we have 1 reordering affecting just 2 packets every about
100000 packets (the system is streaming at 250 packets/second).
We never observed duplicated packets.

In this case, the reordering has a very strong correlation with route change
(the ISP has multiple fibers and a fully meshed METRO network).

Regards,
Pancrazio




On 10/22/06, David Barrett <dbarrett at quinthar.com> wrote:
>
>  "No, I don't have the data and I really doubt that anyone has it. The
> reason being is that it'd been a very big project to collect raw
> *representative* stats and the data is also virtually impossible to
> estimate."
>
>
>
> Heh, now **there's** a challenge I'll take up.  Let me see if I can gather
> this unknowable number…
>
>
>   ------------------------------
>
> *From:* p2p-hackers-bounces at lists.zooko.com [mailto:
> p2p-hackers-bounces at lists.zooko.com] *On Behalf Of *Alex Pankratov
> *Sent:* Saturday, October 21, 2006 11:13 AM
> *To:* 'theory and practice of decentralized computer networks'
> *Subject:* RE: [p2p-hackers] UDP Behavior in the Real World
>
>
>
> No, I don't have the data and I really doubt that anyone has it. The
> reason being is that it'd been a very big project to collect raw
> *representative* stats and the data is also virtually impossible to
> estimate.
>
>
>
> I suspect the re-ordering chances are different between corporate, home
> and academic users; and they also probably change between different ISPs. As
> I mentioned below a multi-cpu router/firewall/utm box under load is likely
> to cause the re-ordering (unless it implements per-CPU stickiness of each
> network session it handles). These become increasingly popular now, but how
> many of them out there and how often they run close to their capacity is
> hard to tell.
>
>
>
> Having said this I know that IPsec guys in company I worked for had to
> extend IPsec replay protection window in their code from typical 32 bits to
> 1024 bits *exactly* because they ran into re-ordering issues with real-life
> customers. Interestingly enough the problem was that some packets were
> arriving way *earlier* than expected .. which is a bit odd if you think
> about it :)
>
>
>
> *From:* p2p-hackers-bounces at lists.zooko.com [mailto:
> p2p-hackers-bounces at lists.zooko.com] *On Behalf Of *David Barrett
> *Sent:* Friday, October 20, 2006 4:51 PM
> *To:* 'theory and practice of decentralized computer networks'
> *Subject:* RE: [p2p-hackers] UDP Behavior in the Real World
>
> Thanks.  By "does happen" do you mean "it happens for everyone every once
> and a while" or "it happens frequently for some users and not at all for
> others"?
>
>
>
> I realize the internet is a big place and there will always be some user
> that has any networking property in unusually high or unusually low
> frequency.  But I'm curious what the "usual" frequency is for this sort of
> thing, if you have any data on this.
>
>
>
> -david
>
>
>   ------------------------------
>
> *From:* p2p-hackers-bounces at lists.zooko.com [mailto:
> p2p-hackers-bounces at lists.zooko.com] *On Behalf Of *Alex Pankratov
> *Sent:* Friday, October 20, 2006 4:30 PM
> *To:* 'theory and practice of decentralized computer networks'
> *Subject:* RE: [p2p-hackers] UDP Behavior in the Real World
>
>
>
> Regarding (1) - reordering does happen.
>
>
>
> First of all it happens in L2 load-balanced scenarios, ie when the traffic
>
>
> goes through virtualized servers or server clusters. It also happens when
>
> the traffic goes through multi-CPU routers and finally It happens (though
>
> less frequently) due to route flips.
>
>
>
> There is an anecdotal evidence that reordering also happens more
>
> frequently over Gigabit links, but I suspect that this is hardware
> specific
>
> and it has something to do with NIC drivers flipping between interrupt and
>
> poll-based modes.
>
>
>
> Alex
>
>
>  ------------------------------
>
> *From:* p2p-hackers-bounces at lists.zooko.com [mailto:
> p2p-hackers-bounces at lists.zooko.com] *On Behalf Of *David Barrett
> *Sent:* Friday, October 20, 2006 4:14 PM
> *To:* 'theory and practice of decentralized computer networks'
> *Subject:* [p2p-hackers] UDP Behavior in the Real World
>
> So I'm working on a TCP over UDP layer and I'm wondering about the
> following; what have you seen?
>
>
>
> 1)       *How often does out-of-order packet delivery happen in the real
> world?*  My reading and experience indicates it's actually very rare, but
> I'm curious if you concur.
>
> 2)       *How often does duplicate packet delivery happen in the real
> world?*  Again, I'm of the impression it's even more rare than
> out-of-order, but I'd like your thoughts.
>
> Clearly, these happen and need to be accounted for.  But I'm curious to
> what degree they happen, as this affects performance characteristics of the
> implementation.
>
>
>
> -david
>
>
> _______________________________________________
> p2p-hackers mailing list
> p2p-hackers at lists.zooko.com
> http://lists.zooko.com/mailman/listinfo/p2p-hackers
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.zooko.com/pipermail/p2p-hackers/attachments/20061022/3b2fef14/attachment.htm


More information about the p2p-hackers mailing list