[rssac-caucus] MTU, MSS and Frag
I propose RSSAC caucus explore an understanding of the MTU, MSS and fragmentation behaviour of all instances of the roots, in IPv6 and IPv4. If we had this data, we would be able to understand if what we see with any given instance of query to a root was local, or more likely along the path. Some evidence from random query tests suggests both that there is no consistency between the root operators, and some inconsistency within a given root letter, depending on where you query goes, or what protocol you use. Because of the increase in TCP, and packetsizes in responses, this is of interest because of the exposure to fragmentation, or loss of pMTU signalling, or the variances in MSS/TCP against interface MTU and UDP. This question has a quality which goes to "only the root operators can say" which is why it feels to me like a well-formed question for the caucus to consider, which implies RSSAC taking it on board to ask the root operators about. -George
On Mar 26, 2017, at 5:03 PM, George Michaelson <ggm@algebras.org> wrote:
Some evidence from random query tests suggests both that there is no consistency between the root operators, and some inconsistency within a given root letter, depending on where you query goes, or what protocol you use.
George, I'm getting the sense here that you might think such inconsistency is bad, or unintentional. Should this be a topic of study, perhaps one component should be whether it is beneficial or harmful to have such differences, both inter- and intra-operator. DW
Its an open point. I certainly do think its bad, but I absolutely concede others might think it's either benign, or even good. I'm struggling to hypothesise how it could be good to have distinct fragmentation outcomes which are essentially unknowable in advance for a specific client, and the debug qualities for a large provider facing reports of problem from point A which are unrepeatable from point B stand out to me, but I suppose it is in that set of things which fall to "diversity is good". If it is intentional, wouldn't it be better to have it documented? -G On Sun, Mar 26, 2017 at 5:23 PM, Wessels, Duane <dwessels@verisign.com> wrote:
On Mar 26, 2017, at 5:03 PM, George Michaelson <ggm@algebras.org> wrote:
Some evidence from random query tests suggests both that there is no consistency between the root operators, and some inconsistency within a given root letter, depending on where you query goes, or what protocol you use.
George,
I'm getting the sense here that you might think such inconsistency is bad, or unintentional.
Should this be a topic of study, perhaps one component should be whether it is beneficial or harmful to have such differences, both inter- and intra-operator.
DW
I propose RSSAC caucus explore an understanding of the MTU, MSS and fragmentation behaviour of all instances of the roots, in IPv6 and IPv4.
Just for your information, bind-9.11.0 (I don't know when, but at least from -P1) set 1220 to MAX TCP Segment size, regardless of the IP version being used (see at around line 3701 of lib/isc/unix/socket.c). This may reduce the efficiency of TCP transaction a little bit, but will cause no harm. We may need some more discussion to define a correct way in general (not limited to DNS case) how we can handle TCP segment size. We may able to control it with IPV6_USE_MIN_MTU and TCP_MAXSEG, however, each application must set appropriately. But this may not a business of RSSAC caucus. -- Akira Kato
participants (3)
-
Akira Kato -
George Michaelson -
Wessels, Duane