TRANSKRIP 🔗
Ilyas Rahimi 0:00
As mentioned, we have a baseline. So at the moment, we do have a
baseline of how much bytes per day do we have from the each route.
Actually, that that's an established at the moment. And then we
would compare that to how many updates each of those bind unbound
and not resolver will take each day. So we counted the total
traffic would be created by a local route serving locally compared
to the hierarchical main way, which actually there is. So it's
just kind of looking at the small query-based queries driven daily
compared to a local route. How much traffic it costs.
George Michaelson 0:47
You're listening to Ping, a podcast by APNIC discussing all things
related to measuring the Internet. I'm your host George
Michaelson. This time, I'm talking to Ilyas Rahimi, who has just
graduated from the University of Amsterdam and Willem Toorop from
NLNet Labs. Ilyas did a master's thesis project on the impact of
Local Root zone serving. This was supervised by Willem and Marios
Avgeris, who is an assistant professor at UvA, where Ilyas
studied. Local Root is an operational practice of downloading the
state of the root zone from a publication point or in-band inside
the DNS protocol. Once the zone has been downloaded and validated,
the resolver can know authoritatively from locally held data some
queries simply can't resolve because their final top-level domain
element, the rightmost one, like .net in the fully qualified
domain name www.apnic.net, they can know this isn't in the root
zone. It could be a typo, or a misconfiguration, or a deliberate
attempt to cause a denial of service attack with reflection
traffic. Aside from minimizing the time and the traffic to refuse
to serve these unresolvable domain names, local route also permits
query minimization. Leakage of what name is being looked up can be
kept to a minimum. There is one step in the chain of query
resolution that doesn't have to be performed outside the local
resolver for a commonly used top-level domain like .com. This
won't be much different to having the cached copy locally, but for
less frequently used top-level domains, and there are now 1000s of
them, it's a small additional benefit. Ilyas's work looked at the
impact of different ways of fetching the top-level root zone and
the impact that this had on traffic from the resolver under
different models of how to fetch, how often to fetch, and the cost
benefit of fetching the entire root zone compared to the atomic
query through the cache as normal. Willem. Ilyas, welcome to Ping.
Willem Toorop 3:02
Thank you, George. Nice to be here again.
Ilyas Rahimi 3:05
Yes, indeed, it's an honor.
George Michaelson 3:06
People know Willem because he's presented on Ping a couple of
times before. But Ilyas, could you introduce yourself to the
audience?
Ilyas Rahimi 3:14
I'm Ilyas Rahimi. I recently graduated from OS3 of UvA
University of Amsterdam, and Willem used to present DNS in the
class, and I really loved that. And at some point, I just went to
him and I wanted to do a research based on DNS to him, and he
offered me this project, which I'm actually really really happy
about. And I learned a lot during this project, so that's the kind
of the start of this project.
George Michaelson 3:40
Willem, your office is on the campus of
Willem Toorop 3:43
yes, that's right.
George Michaelson 3:44
The
university, so although NLNet isn't formally in the teaching
framework, you have quite a high degree of engagement with masters
and PhD students.
Willem Toorop 3:52
Yes, absolutely. The security and network engineering master of
the University of Amsterdam is fantastic. They do two short
research projects a year, or they used to do that. I'm not sure if
they do that anymore. And they are ideal for if I want to
investigate something to involve a student, because then there's
the obligation that it has to happen, right? This research.
George Michaelson 4:19
Yeah, it's part of their marks!.
Willem Toorop 4:21
If I need to do it myself, then there's always something else with
priority.
George Michaelson 4:25
Yeah,
Willem Toorop 4:25
but I do do the guest lecture into DNS and DNSSEC. Yeah, on the
master.
George Michaelson 4:32
So Ilyas, we're getting a context here that this is a master's
thesis where you were looking at aspects of the DNS service. Yes.
Ilyas Rahimi 4:40
Yeah.
George Michaelson 4:41
You gave a presentation on this to the University of Amsterdam as
like a thesis defense, and it's on this change of behavior in the
DNS, where instead of all queries going to the root, the 13
canonical root servers, people are encouraged to download or
access a copy of the root zone and have it locally in the local
cache state of the resolver, and you were looking at behaviors of
this system.
Ilyas Rahimi 5:07
True. Yeah, as you mentioned, the local root serving by default is
actually a very hot topic, and we try to quantify the traffic
trade off between the queries and the zone distribution.
George Michaelson 5:17
How exactly did you go about doing this?
Ilyas Rahimi 5:20
Actually, two things to do that we took the RRSIG data, where we
have all the data of all the root, the identical IP addresses, and
the total amount of the data served. So you can actually calculate
the amount of bytes, say per root per day, by taking the identical
IPv4 addresses and aggregation of IPv6, combining them with the
total bytes which you have, and dividing them to each other, then
you will get per resolver a baseline of how much the data is per
resolver per day. Kind of
George Michaelson 5:50
It's
kind of a way of taking the aggregate
Load full transcript
traffic that's visible at
the root and say, if I have this many discretely unique addresses
coming to me with queries, I can make a model based on the
distribution of those queries across those addresses.
Ilyas Rahimi 6:06
Yeah, you have to search in the data. You have the UDP traffic,
TCP traffic, and the data coming, data going, and you have to
select their the their data which you are actually looking for.
And based on that, and we actually calculated the total amount of
MBs per day per route, as mentioned at the beginning, it was
around two MB per resolver per day.
George Michaelson 6:27
Two megabytes per resolver per day.
Ilyas Rahimi 6:31
Yeah,
George Michaelson 6:31
two men as a volume or as a query count. That's a megabytes
measure, not a count of query.
Willem Toorop 6:37
Yeah, the amount of traffic. So this is the baseline measurement,
right? [George: Yeah]. Question that LEAs asked himself is: How
would traffic on the Internet or per resolver be impacted if it
would follow local root, so serving the root zone locally? So
instead of sending queries to the route, it would do a transfer of
the root zone, and so how does that balance out? How many traffic
is there currently per letter per resolver to the root compared to
what the amount of traffic would then be?
George Michaelson 7:14
So, in in this model, Ilyas, how often would a resolver be
expected to fetch and update its belief in what the root zone
Ilyas Rahimi 7:22
is. Actually, two times a day, but this is the hierarchical way
where you will just do the query and the root will answer that. So
at this level, updating doesn't matter that much.
George Michaelson 7:34
Two fetches a day would represent the equivalent of doing a
complete fetch of the state of the root zone at that time, there
are what 4000 labels in the root at present. Plus, you have to
fetch all the associated signatures, all of the NS's, all of the
check sums, all of the key information, all the associated
records. It's not nothing, is it? I mean, doing this fetch is of
itself quite a large blob of data.
Ilyas Rahimi 8:01
That's true. That's the second baseline where we actually compare
the data which we actually gathered. I think the count will
actually be there on the VMs which we actually run to gather the
data of the local root. While the query driven is a small amount
of the query which you will actually easily get from the root.
George Michaelson 8:17
Did you, in effect, do a replay against the synthetic construct,
so you have the real root, an instance, and you are aware of all
the queries made to it and all of the addresses of things querying
it. Did you create like a test bed and rerun the equivalent of
that traffic from a set of addresses to represent that query load
under local root. How did you actually perform the measurement of
change between the two states?
Ilyas Rahimi 8:45
Yeah, for the local root serving is actually we took the four VMs.
What they actually do is they are having a copy of of the root
instead of querying the root each time. They will have their own
local copy of the root. But what is the difference in this setup
is that they don't query each time to the root, but they will do
often update their local copy. So the difference is actually how
often do they copy their local copy, so to say. And by copying
that, they had different approaches. We had four PMs, two doing
the HTTPS based transfer, which was a full transfer without
looking at the SOA record. While the other two, they did an
incremental update each time when the route actually changed and
the SOA changed. They did an update.
George Michaelson 9:31
You were testing real systems, real systems code like the bind
code or the NLNet product or knot. This wasn' a synthetic DNS
resolver. You were using production code.
Ilyas Rahimi 9:31
There's research going on where they actually show you how do even
just do the whole configuration, and we followed that lead to
configure the VMs, so we don't made any mistakes from there.
Willem Toorop 9:56
Yeah, there's an Internet draft currently by amongst others Warren
Kumari that proposes to make local root a best current practice.
Earlier version had in an appendix different configuration how to
configure different software to do local boot, and it has a
configuration for bind, for unbound and for knot resolver. And
Ilyas installed bind and unbound and knot resolver. The software
that could also fetch the root over HTTPS, which is unbound and
knot, had two versions: one doing DNS and the other one during
HTTPS. I have to correct myself. knot -knot resolver does not
support incremental transfer of the root, so it can do only HTTPS.
George Michaelson 10:44
This experiment is kind of dividing into two parts in my ideation,
my thinking. There is a part which is about a measurement, and I
believe from some stuff you've told me, Willem, that there's
actually Root Server Security and Stability Committee and the
Route Operations Group (RRSAC) have an idea of how they believe
measurements should be done for traffic at the route, and so a
component of this ILAS, I'm assuming you were looking at providing
data that would align with that methodology for counting. Yes,
Ilyas Rahimi 11:18
as mentioned, we have a baseline. So at the moment, we do have a
baseline of how much bytes per day do we have from the each root.
Actually, that that's an established at the moment, and then we
would compare that to how many updates each of those bind unbound
and knot resolver will take each day. So we counted the total
traffic would be created by a local root serving locally compared
to the
hierarchical main way, which actually there is. So it's
just kind of looking at the small query-based queries driven daily
compared to a local root. How much traffic it costs?
George Michaelson 11:18
This is using real-world traffic, using real-world server code,
using a substantive volume of queries that's representative of the
volume across the day, and two mechanisms for fetching local root:
one in band using DNS and one out of band using web.
Willem Toorop 11:30
Indeed, so the so it's comparing. Indeed, assuming that currently
no resolvers do local root, which is obviously not true because
there's a small number that would do it already. But the root
servers provide metrics about their own letters. This is in a
report. The metrics they provide it's in report RSSAC 02, and one
of the things they provide is yeah, not not so much the traffic
volume of each letter per day, but it's the number of queries of
certain sizes. So from that you can deduce the number of traffic
volume per day, and they also provide the unique IP addresses they
see at that letter per day. So if you divide the total number of
volume traffic by the number of unique IP addresses seen, then you
have an estimate of the number of of the traffic volume per
resolver per day for that specific letter.
George Michaelson 13:17
It's likely that that would actually be something like a Gaussian
distribution, and there would be some number of resolvers that
would do phenomenally few queries, and some number of resolvers
that would saturate. But in strict sense,
Ilyas Rahimi 13:30
yeah,
George Michaelson 13:30
that's probably equivalent to an average across all of them.
Willem Toorop 13:33
On the other side, the real world measurement that we do is so the
four different configurations with actual resolvers configured to
do local routes,
George Michaelson 13:44
and that was the other half of my bifurcating model. Because one
part is strictly measurement, but the other is like operational
experience and proof of functionality, isn't it? [Willem: Yes].
You were actually testing these instructions and these systems did
what they say on the label,
Willem Toorop 14:01
yes, and we then assume or can also observe that no queries will
be sent to the root anymore because they can all be deduced from
local data.
George Michaelson 14:13
Well, this actually is an interesting question because, as a
theory, I think that's wonderful. In principle, no queries, but
I'm starting to wonder: Is there no leakage? Is there no assertion
of queries that bypass the cache state in the resolver? Did it
actually zero traffic?
Willem Toorop 14:32
There is for some resolvers like bind and unbound. There will be
no traffic other than to look up the actual root servers and
occasional SOA query to see if there's a fresh version of the root
zone, but knot knot resolver has a different approach because they
have a different approach which is also currently proposed as an
Internet draft, by Paul Hoffman to fetch the root zone over HTTPS
and then fill the resolves with cache, right? So if certain names
would be evicted from the cache because there is no place anymore,
or then it's not used as often, such so, then those names will be
queried to the root.
George Michaelson 15:20
Yeah,
Willem Toorop 15:20
this this is actually not what Ilyas was looking into.
George Michaelson 15:24
Ilyas, you have a test bed, you have these resolver instances, you
establish a baseline based on traffic, and you therefore have a
model with rough edges around some number of local roots in the
existing state and the nature of a distribution of traffic load,
and you ran your experiment. Can you tell us a little bit about
the kind of data you saw?
Ilyas Rahimi 15:47
Yeah. So actually, we measure the total amount on occurring on
DNS, and what you actually see is quite accurate. Let's say the
bind results, the update per day was each time accurate about one
point 45 megabytes, and depending on the update, differed from day
to day. For example, on June 20 we had two updates, while on 21
and 22 and 23 we had three and four updates. So actually have
three updates depending on the time that you start, and that
counts actually.
George Michaelson 16:17
So it's not a fixed clock cycle zone update. It's a variant
depending on need.
Willem Toorop 16:23
Well, what we could see is that the root is actually always
updated, refined, so to say, twice a day. That's just a plain
refined. And if there are updates to the root, like some
delegation has new name servers or has new DS record from TLD.
[George: Yeah]. Then beside beside those two clients that happens.
George Michaelson 16:45
So they don't stage all queued updates until a checkpoint. They do
on demand re-sign if there's a need.
Willem Toorop 16:53
Yes. So therefore, there will always be two re-signs a day, and if
there's a change that day, then there will be a third one, or if
there are accidently two changes, what happened during the
measurement period that Ilyas did.
Ilyas Rahimi 17:08
Yeah, and it did actually differs if you are looking at DNS based
update or you are looking at HTTPS based update. On DNS based
update, it purely looked at the SOA changes. On the SOA changes,
it did an update, while on HTTPS based it doesn't do that. That
was also by unbound, but that is whether it was fixed already,
right?
Willem Toorop 17:28
Indeed. So maybe you should explain first what's wrong with
inbound or how naively inbound was implemented, so to say.
Ilyas Rahimi 17:37
Yeah, the polling full zone was actually based on 30 minutes
standard on HTTPS, and it didn't look if there was a change. So
what actually happened at the background each time with each
update of two point 19 megabytes, it updated after each 30
minutes. So at the end of the day, you would have 48 updates each
day on Unbound HTTPS base.
George Michaelson 17:58
Right, the HTTP protocol has a if changed functional behavior that
you can in effect get the equivalent of a time serial to know
whether or not the thing you last fetched has undergone change.
And it sounds like you're saying it just didn't bother looking at
that and it just did a prescriptive fetch.
Willem Toorop 18:17
That's right. And the root zone has a SOA records telling
resolvers or telling secondaries, secondary authoritative servers
to refresh or to have the zone every 30 minutes. And for
secondary, that means do a query for the SOA record. Has the
serial number changed? No. Then you can leave it at that. But if
it has changed, then you have to fetch the whole zone. But there
is no way to do a SOA query over HTTPS.
George Michaelson 18:49
Yeah.
Willem Toorop 18:49
Right. If the well, yeah, DNS over HTTPS, but that's not how the
root zone is served. I'd say.
George Michaelson 18:57
say. Ilyas, you ran this experiment. Is there broadly comparable
behavior from all of the resolvers in this model of testing?
Noting that unbound had this difference, is there in general
broadly the same behavior?
Ilyas Rahimi 19:12
Depends again. What we actually, as generally, can establish is
that the bytes per day per update that is fixed for each resolver.
So that depends on the amount of updates, how much each day
happens. That will make the difference in the MBs, so to say.
George Michaelson 19:27
I am a true believer in local root. At this point, as a naive
party, my assumption is you have been able to demonstrate that
there is a significant saving in traffic by deploying local root.
Ilyas Rahimi 19:40
No, that's actually the the other way around.
George Michaelson 19:44
Seriously,
Ilyas Rahimi 19:45
yeah. While local root actually provides some security in
dependencies, like if there's a small outage, local root is really
good at that. So hiding that if there's a small outage between the
root, local root would take care of that. Or if there's a DDoS
attack lets say.
George Michaelson 20:00
Yeah. So, although this is something that people want to do and
that has qualities of preserving privacy, visibility of queries to
the root, and potentially has an impact in terms of root server
load serving queries then are removed from the system, you appear
to be saying that on the client side, people have to do more work.
Ilyas Rahimi 20:21
Yeah. Yeah. It is almost two times the amount. But if we put it in
perspective, it's much more than that. If we just assume that a
10% of the total resolvers will use local root, at the best
part,
which is not actually, it would add a 3.8 terabytes per day. In
worst case, just adopting 10% of local root by unbound, while the
bug is solved, which we previously discussed, it would be around
27 terabytes a day additionally on the local root side.
George Michaelson 20:56
Wow!
Ilyas Rahimi 20:56
So the amount which it will be adding is really, really huge.
George Michaelson 21:01
But within the capacity of the root service system as a whole to
serve. I mean, this isn't actually a barrier that means cannot
proceed. It's just you can't proceed on a basis it reduces traffic
quite the way you might think.
Willem Toorop 21:15
Yeah, that's right. This was only looking at the amount, the the
traffic increase, so to say, or the change in traffic, and whether
it would increase or decrease, as well, with how the root is
currently updated, there will be a substantial traffic increase to
the root, which they probably can handle fine. But yeah,
George Michaelson 21:37
but this actually is a shift in logic because there was a
component of belief, certainly on my part, that this would reduce
traffic and therefore have beneficial impact in terms of future
growth, and that could still potentially be true, but it's not
immediately apparent. There has to be a second reason that would
justify making a deployment like this. So, what do you think is
the outcome that could make it rational to continue with doing a
local root activity. Is it in some other aspect of behavior of the
system as a whole?
Ilyas Rahimi 22:09
I think there comes the magic of Willem. He was looking at signing
the zone incrementally, so then instead of one huge update each
time, you will have an incremental part of that, which would
reduce significantly. So that would be the magic part of it. Maybe
Willem can expand on that.
Willem Toorop 22:27
That is true. But besides, even if there would be substantial
traffic increase, there are still many advantages of having the
route local to the resolver. As you mentioned, there's a privacy
or increased privacy that the root does not see as many queries as
it does see without local root. [George: Yeah] there's reduced
reliance on the root.
George Michaelson 22:53
So there's a resiliency story in this. There's a potential that
this could reduce weak points in the structural behaviors of the
system as a whole.
Willem Toorop 23:01
Yeah, and it's also if the resolver is verifying the zone MD and
validating the signature of zone MD, which is a hash of the whole
root zone. There is a security benefit as well because currently
referrals in the root zone are not signed. Only the delegation
signer resource record is signed, making the DNSSEC chain of
trust. But glue and name server data is not fine. So
George Michaelson 23:34
right, and Glue is not independently checked
of necessity. It's
used because it has to be when it's an in Bailewick reference to
the subsidiary record being referred to, but it doesn't invoke DNS
checks. Where having the zone fetched and the zone MD fetched and
all of that computation done means you can have inherently much
higher trust in that
Willem Toorop 23:55
data. Absolutely,
George Michaelson 23:57
there is also a quality that if you use the out of band fetch
mechanism. You do still have to fetch data, but you take that data
out of the DNS UDP plane and you move it into the HTTP TLS plane.
It could come from completely unassociated sources.
Willem Toorop 24:16
Yes, I think that's indeed also an idea that I've heard about to
let the content distribution network deal with the problem of
distributing the root zone instead of the the root serve
operators.
George Michaelson 24:31
So, Ilyas, did you see further work that could be carried out in
this field? Do you think there are other projects that master
students could look at in this?
Ilyas Rahimi 24:39
Yeah, my research is actually very time bound. I only had one
month to carry this whole experiment.
George Michaelson 24:45
Yeah,
Ilyas Rahimi 24:45
and as Willem said, the research are being changed at our
university. They will be getting three months, so that will be
much more time. And then you can actually look on the wire. We
actually what we did, we looked at the whole data on the wire. We
didn't look in depth how much is the pure DNS cost actually,
[George: right] Excluding the handshake, excluding all the
overhead, maybe that will give us a whole other insight, which I'm
sure it would. Removing all the handshakes and all the over wires,
it might give us a whole other aspect of what we are seeing at the
moment.
George Michaelson 25:19
Yeah, I think it's great when you do work like this that has the
potential not only to answer immediate questions but also points
at further work to be done. I think that's really fantastic.
Willem, you're going to follow up with this? You see potential for
more?
Willem Toorop 25:32
Yes, absolutely. Well, I actually have done some follow-up work
myself in what Ilyas already mentioned in incrementally signing
the root, yeah, looking at how that would influence the amount of
traffic.
George Michaelson 25:47
Perhaps that is something we could discuss another time on Ping,
because I think that would make quite a nice little story in and
of itself.
Willem Toorop 25:55
Absolutely, yes, definitely, we can do that.
George Michaelson 25:58
So, Willem, Ilyas, thank you very much for coming on Ping and
telling us about this measurement activity. That's fantastic. Is
there somewhere people can read about this?
Ilyas Rahimi 26:07
I think Willem already updated on his page, and he exchanged it
actually also in the Google Doc.
George Michaelson 26:14
Oh, I'll make sure that that is available on the blog that we post
alongside this recording.
Willem Toorop 26:19
Yes, Ilyas's of his work is also available on the NLNetLab's
website in the research and then student projects. He is right on
top of that list there, so you can download the paper there.
George Michaelson 26:34
Great, thank you very much.
Willem Toorop 26:35
You're welcome.
Ilyas Rahimi 26:36
Thank you for having us here.
George Michaelson 26:41
If you've got a story or research to share here on Ping, why not
get in contact by email to ping@apnic.net or via the APNIC social
media channels? Also, remember the measurement@apnic.net mailing
list on orbit is there to discuss and share relevant collaborative
opportunities, grants and funding opportunities, jobs and graduate
placings, or to seek feedback from the community on your own
measurement projects. Be sure to check out the APNIC website for
all your resource and community needs. Until next time.