real-world data about fragmentation
Hi all, It's common wisdom that a datagram that needs to be fragmented between endpoints (because it is bigger than the path MTU) will demonstrate less reliable delivery and reassembly than a datagram that doesn't need to be fragmented, because math, firewall, other, take your pick. Is anybody aware of any wide-scale studies that examine the probability of fragmentation of datagrams of different sizes? For example, I could reasonable expect an IPv4 packet of 576 bytes not to be fragmented very often (to choose a size not at random). The probability of a 10,000 octet IPv4 packet getting fragmented seems likely to be 100%, if we're talking about arbitrary paths across the Internet. What does the curve look like between 576 bytes and 10,000 bytes? I might expect exciting curve action around 1500 bytes (because ethernet), 1492 (PPPoE), 1480 (GRE), etc. But I'm interested in actual data. Anybody have any pointers? IPv4 and IPv6 are both interesting. Joe
I can send you a copy of an invited presentation at AINTEC from 2009. /bill On Wed, Apr 02, 2014 at 02:14:22PM -0400, Joe Abley wrote:
Hi all,
It's common wisdom that a datagram that needs to be fragmented between endpoints (because it is bigger than the path MTU) will demonstrate less reliable delivery and reassembly than a datagram that doesn't need to be fragmented, because math, firewall, other, take your pick.
Is anybody aware of any wide-scale studies that examine the probability of fragmentation of datagrams of different sizes?
For example, I could reasonable expect an IPv4 packet of 576 bytes not to be fragmented very often (to choose a size not at random). The probability of a 10,000 octet IPv4 packet getting fragmented seems likely to be 100%, if we're talking about arbitrary paths across the Internet.
What does the curve look like between 576 bytes and 10,000 bytes?
I might expect exciting curve action around 1500 bytes (because ethernet), 1492 (PPPoE), 1480 (GRE), etc. But I'm interested in actual data.
Anybody have any pointers? IPv4 and IPv6 are both interesting.
Joe
This isn't a direct answer to the question, but I find this paper pretty useful (even though it is dated now): Beyond Folklore: Observations on Fragmented Traffic by Colleen Shannon, David Moore, and k claffy IEEE/ACM Transactions on Networking, December 2002 http://www.caida.org/publications/papers/2002/Frag/frag.pdf (Bill, I'd be curious to see your AINTEC slides, too.) -- Jen On Apr 2, 2014, at 2:50 PM, bmanning@vacation.karoshi.com wrote:
I can send you a copy of an invited presentation at AINTEC from 2009.
/bill
On Wed, Apr 02, 2014 at 02:14:22PM -0400, Joe Abley wrote:
Hi all,
It's common wisdom that a datagram that needs to be fragmented between endpoints (because it is bigger than the path MTU) will demonstrate less reliable delivery and reassembly than a datagram that doesn't need to be fragmented, because math, firewall, other, take your pick.
Is anybody aware of any wide-scale studies that examine the probability of fragmentation of datagrams of different sizes?
For example, I could reasonable expect an IPv4 packet of 576 bytes not to be fragmented very often (to choose a size not at random). The probability of a 10,000 octet IPv4 packet getting fragmented seems likely to be 100%, if we're talking about arbitrary paths across the Internet.
What does the curve look like between 576 bytes and 10,000 bytes?
I might expect exciting curve action around 1500 bytes (because ethernet), 1492 (PPPoE), 1480 (GRE), etc. But I'm interested in actual data.
Anybody have any pointers? IPv4 and IPv6 are both interesting.
Joe
Hi, Joe, On 04/02/2014 03:14 PM, Joe Abley wrote:
Is anybody aware of any wide-scale studies that examine the probability of fragmentation of datagrams of different sizes?
We're in the process of measuring some (kind of related stuff). If you're interested in this data, we might be able to provide something along these lines in 1 month or so... It seems to be mostly about measuring the MTU to as many destinations as possible, so to speak...
For example, I could reasonable expect an IPv4 packet of 576 bytes not to be fragmented very often (to choose a size not at random).
Note: there shouldn't be any special magic around this number (usualy mistakenly interpreted as the minimum IPv6 MTU, but rather being the minimum IPv4 reassembly buffer size).
The probability of a 10,000 octet IPv4 packet getting fragmented seems likely to be 100%, if we're talking about arbitrary paths across the Internet.
What does the curve look like between 576 bytes and 10,000 bytes?
I might expect exciting curve action around 1500 bytes (because ethernet), 1492 (PPPoE), 1480 (GRE), etc. But I'm interested in actual data.
Anybody have any pointers? IPv4 and IPv6 are both interesting.
Probably off-topic, but since you mentioned reliability of IPv6 fragmentation: * <http://www.iepg.org/2013-11-ietf88/fgont-iepg-ietf88-ipv6-frag-and-eh.pdf> * <http://www.iepg.org/2014-03-02-ietf89/fgont-iepg-ietf89-eh-update.pdf> Thanks! Cheers, -- Fernando Gont e-mail: fernando@gont.com.ar || fgont@si6networks.com PGP Fingerprint: 7809 84F5 322E 45C7 F1C9 3945 96EE A9EF D076 FFF1
On Apr 2, 2014, at 11:14 AM, Joe Abley <jabley@hopcount.ca> wrote:
Hi all,
It's common wisdom that a datagram that needs to be fragmented between endpoints (because it is bigger than the path MTU) will demonstrate less reliable delivery and reassembly than a datagram that doesn't need to be fragmented, because math, firewall, other, take your pick.
Is anybody aware of any wide-scale studies that examine the probability of fragmentation of datagrams of different sizes?
For example, I could reasonable expect an IPv4 packet of 576 bytes not to be fragmented very often (to choose a size not at random). The probability of a 10,000 octet IPv4 packet getting fragmented seems likely to be 100%, if we're talking about arbitrary paths across the Internet.
What does the curve look like between 576 bytes and 10,000 bytes?
I might expect exciting curve action around 1500 bytes (because ethernet), 1492 (PPPoE), 1480 (GRE), etc. But I'm interested in actual data.
Anybody have any pointers? IPv4 and IPv6 are both interesting.
Seems a good thing for RIPE Atlas probes to measure. But they are probably not generally connected to representative networks (read: poor networks). -d
participants (5)
-
bmanning＠vacation.karoshi.com -
Dan Wing -
Fernando Gont -
Jennifer Rexford -
Joe Abley