Understanding the deployment of public recursive resolvers

May 6th, 2024 by Matthew Luckie

This is the third in a series of essays (two earlier blog posts [1, 2]) about CAIDA’s new effort to reduce the barrier to performing a variety of Internet measurements.

We were recently asked about running Trufflehunter, which infers the usage properties of rare domain names on the Internet by cache snooping public recursive resolvers, on Ark. The basic idea of Trufflehunter is to provide a lower bound of the use of a domain name by sampling caches of large recursive resolvers.  One component of Trufflehunter is to identify the anycast instances that would answer a given vantage point’s DNS queries.   Trufflehunter uses a series of TXT queries to obtain the anycast instance of a given public recursive resolver, and considers four large public recursive resolvers: Cloudflare (, Google (, Quad9 (, and OpenDNS (  The original paper describes the queries in section 4.1, which we summarize below, highlighting the interesting parts of each response with blue color.

Cloudflare returns an airport code representing the anycast deployment location used by the VP:
$ host -c ch -t txt id.server
id.server descriptive text "AKL"

Google returns an IP address representing the anycast deployment used by the VP, which can be mapped to an anycast deployment location with a second query:
$ host -t txt o-o.myaddr.l.google.com
o-o.myaddr.l.google.com descriptive text ""
o-o.myaddr.l.google.com descriptive text "edns0-client-subnet <redacted>/24"

$ host -t txt locations.publicdns.goog
locations.publicdns.goog descriptive text " icn " ... " syd " " cbf " ...

Quad9 returns the hostname representing the resolver that provides the answer:
$ host -c ch -t txt id.server
id.server descriptive text "res100.akl.rrdns.pch.net"

and finally, OpenDNS returns a bunch of information in a debugging query, which includes the server that handles the query:
$ host -t txt debug.opendns.com
debug.opendns.com descriptive text "server r2004.syd"
debug.opendns.com descriptive text "flags 20040020 0 70 400180000000000000000007950800000000000000"
debug.opendns.com descriptive text "originid 0"
debug.opendns.com descriptive text "orgflags 2000000"
debug.opendns.com descriptive text "actype 0"
debug.opendns.com descriptive text "source <redacted>:42845"

Each of these queries requires a slightly different approach to extract the location of the anycast instance.  Our suggestion is for experimenters to use our newly created python library, which provides programmatic access to measurement capabilities of Ark VPs (described in two earlier blog posts [1, 2]).  The code for querying the recursive resolvers from all VPs is straight forward, and is shown below.  First, on lines 15-27, we get the Google mapping, so that we can translate the address returned to an anycast location.  This query requires TCP to complete, as the entry is larger than can fit in a UDP payload (the response is 9332 bytes at the time of writing this blog).  This query is synchronous (we wait for the answer before continuing, because the google queries that follow depend on the mapping) and is issued using a randomly selected Ark VP.  Then, on lines 29-38, we issue the four queries to each of the anycasted recursive resolvers from each VP.  These queries are asynchronous; we receive the answers as each Ark VP obtains a response.  On lines 40-82, we process the responses, storing the results in a multi-dimensional python dictionary, which associates each Ark VPs with their anycast recursive resolver location, as well as the RTT between asking the query and obtaining the response.  Finally, on lines 84-96, we dump the results out in a nicely formatted table that allows us to spot interesting patterns.

01 import argparse
02 import datetime
03 import ipaddress
04 import random
05 import re
06 from scamper import ScamperCtrl
08 def _main():
09   parser = argparse.ArgumentParser(description='get public recursive locs')
10   parser.add_argument('sockets')
11   args = parser.parse_args()
13   ctrl = ScamperCtrl(remote_dir=args.sockets)
15   # pick an ark VP at random to issue the query that gets the mapping
16   # of google recursive IP to location
17   goog_nets = {}
18   obj = ctrl.do_dns('locations.publicdns.goog',
19     inst=random.choice(ctrl.instances()),
20     qtype='txt', tcp=True, sync=True)
21   if obj is None or len(obj.ans_txts()) == 0:
22     print("could not get google mapping")
23     return
24   for rr in obj.ans_txts():
25     for txt in rr.txt:
26       net, loc = txt.split()
27       goog_nets[ipaddress.ip_network(net)] = loc
29   # issue the magic queries to get the instance that answers the query
30   for inst in ctrl.instances():
31     ctrl.do_dns('o-o.myaddr.l.google.com', server='', qtype='txt',
32       attempts=2, wait_timeout=2, inst=inst)
33     ctrl.do_dns('id.server', server='', qclass='ch', qtype='txt',
34       attempts=2, wait_timeout=2, inst=inst)
35     ctrl.do_dns('id.server', server='', qclass='ch', qtype='txt',
36       attempts=2, wait_timeout=2, inst=inst)
37     ctrl.do_dns('debug.opendns.com', server='', qtype='txt',
38       attempts=2, wait_timeout=2, inst=inst)
40   # collect the data
41   data = {}
42   for obj in ctrl.responses(timeout=datetime.timedelta(seconds=10)):
43     vp = obj.inst.name
44     if vp not in data:
45       data[vp] = {}
46     dst = str(obj.dst)
47     if dst in data[vp]:
48       continue
49     data[vp][dst] = {}
50     data[vp][dst]['rtt'] = obj.rtt
52     for rr in obj.ans_txts():
53       for txt in rr.txt:
54         if dst == '':
55           # google reports an IPv4 address that represents
56           # the site that answers the query. we then map
57           # that address to a location using the mapping
58           # returned by the locations TCP query.
59           try:
60             addr = ipaddress.ip_address(txt)
61           except ValueError:
62             continue
63           for net, loc in goog_nets.items():
64             if addr in net:
65               data[vp][dst]['loc'] = loc
66               break
67         elif dst == '':
68           # Cloudflare replies with a single TXT record
69           # containing an airport code
70           data[vp][dst]['loc'] = txt
71         elif dst == '':
72           # Quad9 reports a hostname with an embedded
73           # airport code.
74           match = re.search("\\.(.+?)\\.rrdns\\.pch\\.net", txt)
75           if match:
76             data[vp][dst]['loc'] = match.group(1)
77         elif dst == '':
78           # opendns reports multiple TXT records; we want the one
79           # that looks like "server r2005.syd"
80           match = re.search("^server .+\\.(.+?)$", txt)
81           if match:
82             data[vp][dst]['loc'] = match.group(1)
84   # format the output
85   print("{:15} {:>13} {:>13} {:>13} {:>13}".format(
86     "# vp", "google", "couldflare", "quad9", "opendns"))
87   for vp, recs in sorted(data.items()):
88     line = f"{vp:15}"
89     for rec in ('', '', '', ''):
90       cell = ""
91       if 'loc' in recs[rec]:
92         rtt = recs[rec]['rtt']
93         loc = recs[rec]['loc']
94         cell = f"{loc} {rtt.total_seconds()*1000:5.1f}"
95       line += f" {cell:>13}"
96     print(line)
98 if __name__ == "__main__":
99    _main()

The code runs quickly — no longer than 10 seconds (if one of the VPs is slow to report back), illustrating the capabilities of the Ark platform.

The output of running this program is shown below.  We have highlighted cells where the RTT was at least 50ms larger than the minimum RTT to any of the large recursive resolvers for the given VP.  These cells identify low-hanging fruit for operators, who could examine BGP routing policies with the goal of selecting better alternative paths.

# vp             google   couldflare         quad9      opendns
abz-uk.ark    lhr  24.4    LHR  16.5     lhr  16.9    lon  16.4
abz2-uk.ark   lhr  23.4    MAN   8.7     man   8.4   man1   8.5
acc-gh.ark    jnb 246.7    JNB 167.3     acc   0.5   cpt1 236.6
adl-au.ark    mel  28.1    ADL   5.5     syd  22.6   mel1  14.0
aep-ar.ark    scl  29.4    EZE   8.0   qaep2   4.6   sao1  34.0
aep2-ar.ark   scl  38.2    EZE   6.6   qaep2   5.7   sao1  42.5
akl-nz.ark    syd  26.9    AKL   2.6    akl2   4.0    syd  26.1
akl2-nz.ark   syd  29.2    AKL   4.9    akl2   4.0    syd  26.9
ams-gc.ark    grq   5.0    FRA   6.3     ams   0.8    ams   1.1
ams3-nl.ark   grq   6.0    AMS   2.1     ams   1.3    ams   1.9
ams5-nl.ark   grq   5.1    AMS   1.2     ams   1.9    ams   1.7
ams7-nl.ark   grq   6.0    AMS   2.4     ams   1.8    ams   2.3
ams8-nl.ark   grq  14.8    AMS  11.5     fra  17.7    ams  11.8
arn-se.ark    lpp   9.7    ARN   2.0     arn   0.8   cph1  10.5
asu-py.ark    scl  55.4    EZE  22.6     asu   0.8   sao1  57.0
atl2-us.ark   atl   9.8    DFW  23.6   qiad3  20.5   atl1   4.6
atl3-us.ark   atl  22.7    ATL  13.1     atl  17.0   atl1  19.7
aus-us.ark    dfw  80.7    DFW  56.6     dfw  46.6    dfw  17.6
avv-au.ark    mel 142.3    MEL   9.4     syd  22.6   mel1   8.5
bcn-es.ark    mad  38.0    BCN  13.3     bcn   0.5   mad1   9.8
bdl-us.ark    iad  13.3    EWR   5.0     lga   4.9    ash  10.6
bed-us.ark    iad  30.6    BOS  15.2     bos  13.5   bos1  19.8
beg-rs.ark    mil  62.5    BEG   1.2     beg   0.7   otp1  12.9
bfi-us.ark    dls  11.7    SEA   3.8   xsjc1  96.9    sea  10.5
bjl-gm.ark    bru  87.5    CDG  71.1     bjl   6.7   cdg1  73.4
bna-us.ark    atl  11.4    BNA   9.0     atl   9.5   atl1   9.3
bna2-us.ark   cbf  25.5    ORD  19.7     ord  12.8    chi  12.8
bos6-us.ark   iad  17.6    BOS   5.6   qiad3  15.5   bos1   5.7
bre-de.ark    grq  21.4    TXL  23.1     ber  17.2    ams  23.7
btr-us.ark    dfw  20.3    ORD  29.5     dfw  13.2    dfw  11.4
bwi2-us.ark   iad  12.7    IAD   4.4   qiad3   3.6   rst1   3.1
cdg-fr.ark    mil  24.9    MRS   1.5     mrs   1.0   mrs1  11.6
cdg3-fr.ark   bru   8.5    CDG   3.9     ams  13.7   cdg1   5.5
cgs-us.ark    iad   4.5    EWR   7.9     iad   2.2    ash   2.1
cjj-kr.ark    hkg  63.4                qhnd2  70.7    hkg  47.3
cld4-us.ark   lax  22.7    LAX  15.7     bur 104.1    lax  14.3
cld5-us.ark   lax  52.6    LAX  20.7     bur 103.1    lax  21.9
cld6-us.ark   lax  14.5    LAX   5.2   qlax1   7.0    lax   6.0
cos-us.ark    cbf  22.9    DEN   7.3     ord  36.0   den1  12.6
dar-tz.ark    jnb 224.6    NBO  14.0     dar   1.0    jnb  51.1
dar2-tz.ark   jnb 223.0    DAR   0.9     dar   0.9    jnb  48.0
dmk-th.ark    sin  32.2    SIN  28.0     bkk   2.4   nrt2  95.0
dtw2-us.ark   cbf  18.1    DTW   4.6     dtw   4.2    chi  10.7
dub-ie.ark    lhr  18.9    DUB   1.2     dub   0.7   dub1   0.7
dub2-ie.ark   lhr  18.7    DUB   1.8     dub   1.1   dub1   1.2
dub3-ie.ark   lhr  35.5    ZRH  45.6     dub  16.9   dub1  11.6
ens-nl.ark    grq   9.3    AMS   5.2     ams   4.0    ams   4.9
eug-us.ark    dls  15.7    SJC  18.3     sea   6.5    sea   6.5
fra-gc.ark    fra   9.1    FRA   1.1     fra   1.1    fra   1.1
gig-br.ark    gru 125.2    GIG   2.2   qrio1   1.4   rio1   1.1
gva-ch.ark                 ZRH   5.3     gva   1.0    fra  10.8
gye-ec.ark    chs 128.7    MIA  66.8   quio2  10.7    mia  77.1
ham-de.ark    grq  19.7    HAM   8.5     ber  11.0    fra  17.7
her2-gr.ark   mil  64.9                  ath   7.5   mil1  32.6
hkg4-cn.ark   tpe  13.6    HKG   1.0   qhkg3   4.9    hkg   9.5
hkg5-cn.ark   hkg  16.3    HKG   2.6   qhkg3   3.9    hkg   3.6
hlz2-nz.ark   syd  40.9    AKL  16.9     akl  13.8    syd  38.2
hnd-jp.ark    nrt   6.5    NRT   3.4   qhnd2  80.3    nrt   3.6
hnl-us.ark    dls  82.3    HNL   1.1     sea  75.2    sea  75.3
iev-ua.ark    waw  45.5    FRA  29.8   qwaw2  14.3   wrw1  14.9
igx2-us.ark   iad  16.1    EWR  16.9   qiad3  11.5   atl1  11.8
ind-us.ark    atl  25.1    IND   1.1     ord   8.5   atl1  11.0
ixc-in.ark    del  78.8    DEL   0.8   qsin1   0.7   mum2   0.7
jfk-us.ark    iad   9.6    EWR   1.5     lga   0.6    nyc   1.0
kgl-rw.ark    jnb 231.6    NBO  15.0     kgl   3.3    jnb  73.6
ktm-np.ark    del  98.6    DEL  30.4     ktm   7.6   mum1  94.4
las-us.ark    lax  15.5    LAX   8.4     pao  15.9    lax   8.2
lax3-us.ark   lax  10.5    LAX   3.0     bur   2.0    lax   2.0
lcy2-uk.ark   lhr  11.5    MAN   7.4     lhr  11.2    lon   2.2
lej-de.ark    fra  21.2    FRA   9.7     fra  12.5    fra   9.6
lex-us.ark    iad  17.9    IAD  15.7     iad  17.3   atl1  25.9
lgw-uk.ark    lhr  31.3    LHR  24.2     lhr  24.6    lon  23.6
lhe2-pk.ark   dia 195.0    KHI  35.0   qsin4 113.1    sin 130.6
lis-pt.ark    mad  33.0    LIS   4.5     lis   3.9   mad1  18.0
lke2-us.ark   dls  21.7    SEA  21.4     sea  15.1    sea  11.7
lun-zm.ark    jnb 189.9    JNB  32.2     jnb  23.3    jnb  23.3
lwc-us.ark    tul  18.0    MCI   1.4   xsjc1 106.3    dfw  10.4
lwc2-us.ark   dfw  24.4    DFW  15.7   qlax1  46.9    dfw  16.5
mdw-us.ark    cbf  25.7    ORD   7.0     ord   3.4    chi   4.4
med2-co.ark   mrn  96.4    MIA  50.7   qbog1  22.7    mia  48.9
mhg-de.ark    fra  30.7    FRA  19.1     fra  21.4    fra  18.4
mia-gc.ark    mrn  19.7    MIA   1.1     mia   1.0    mia   0.6
mnl-ph.ark    hkg  87.9    MNL   1.8   qsin1  50.8    hkg  75.8
mnz-us.ark    iad  13.4    IAD   5.8   qiad3   5.2   rst1   4.0
mru-mu.ark    jnb 206.1    JNB  43.1     mru   1.0    jnb  43.3
msy-us.ark    dfw  34.4    DFW  21.9   qlax1  55.2    dfw  23.0
mty-mx.ark    tul  44.1    MFE   4.5   qiad3  39.6    dfw  16.7
muc-de.ark    fra  30.3    FRA   9.1     fra   8.8    fra   8.9
muc3-de.ark   zrh  25.9    MUC   6.1     fra  11.0    fra  12.2
nap2-it.ark   mil  37.8    MXP  19.8     fra  38.0   mil1  15.2
nbo-ke.ark    jnb 220.5    NBO   2.4     nbo   2.2    jnb  60.7
nic-cy.ark    mil  61.3    LCA   1.0     mrs  35.2   mrs1  35.4
nrn-nl.ark    grq   8.6    AMS   5.5     ams   2.9    ams   3.1
nrt-jp.ark    nrt   4.8    NRT   1.2     hkg 101.9   nrt2   1.0
nrt3-jp.ark   nrt   6.5    NRT   3.5     pao 101.7   nrt2   3.5
oak5-us.ark   lax  19.3    SJC   5.9    sjc0   4.4    pao   4.4
okc-us.ark    dfw   9.6    MCI   9.6     dal   8.2    dfw   7.3
ord-us.ark    cbf  13.0    ORD   2.9     iad  19.0    chi   1.6
ory4-fr.ark   bru   6.5    CDG   2.0     cdg   1.4   cdg1   1.7
ory6-fr.ark   bru   6.4    LHR   8.8     lhr   8.4    lon   8.6
ory7-fr.ark   bru   7.7    CDG   4.0     cdg   5.0   cdg1   4.0
ory8-fr.ark   bru   6.3    CDG   2.0     lhr  25.7    lon   9.4
osl-no.ark    lpp  46.5    OSL   0.9     osl  25.8   sto1   9.0
pbh2-bt.ark   bom 105.2    MAA  92.0     pbh   1.8    sin  91.7
per-au.ark    syd  48.9    PER   2.0     per   1.2    syd  45.3
per2-au.ark   syd 140.0    PER   2.5     per   1.5    syd  44.9
phl-us.ark    iad  15.9                  iad  14.0   rst1  15.9
pna-es.ark    mad  33.9    BCN  13.5     bcn  13.3   mad1  19.3
prg-cz.ark    fra  16.9    PRG   1.1   qbts1   5.9   prg1   0.6
prg2-cz.ark   fra  29.8    PRG   1.7   qfra3  14.0   prg1   1.5
pry-za.ark    jnb 167.3    JNB   1.2     jnb   0.6   cpt1  18.4
puw-ru.ark    lpp  18.2    DME   1.5     beg  67.1    fra  34.3
pvu-us.ark    lax  25.2    SLC   3.8     slc   2.7    sea  18.4
rdu-us.ark    iad   9.5    IAD   8.1     iad   7.7    ash   8.6
rdu2-us.ark   iad  11.3    IAD   8.1     iad   8.5    ash   8.6
rdu3-us.ark   iad  27.5    IAD  18.6     iad  18.4   atl1  23.8
rkv-is.ark    lhr  49.0    KEF   2.0     kef   0.8   dub1  23.0
san-us.ark    lax  10.4    LAX   4.7     bur  87.2    lax   3.1
san2-us.ark   lax  39.0    LAX   6.5   qlax1   7.5    lax   5.6
san4-us.ark   lax  26.6    LAX  20.2     bur 106.8    lax  17.3
sao-br.ark    gru 117.3    GRU   2.5   qgru1   1.3   rio1   9.6
scq-es.ark    mad  31.7    MAD  11.0     mad  10.8   mad1  11.3
sea3-us.ark   dls  10.1    SEA   3.9     pao  24.6    sea   3.0
sin-gc.ark    sin   4.0    SIN   1.5   qsin1  93.5    sin   1.1
sin-sg.ark    sin   3.2    SIN   2.3   qsin1  12.2    sin   0.7
sjc2-us.ark   lax  15.2    SJC   1.4    sjc0   0.5    pao   2.1
sjj-ba.ark    waw  71.1    BUD  57.5     vie 213.3   mil1  89.4
sjo-cr.ark    chs  63.0    SJO   2.2     mia  51.8    mia  78.6
snn-ie.ark    lhr  20.4    DUB   4.4     dub   3.9   dub1   3.9
sql-us.ark    lax  19.6    SJC   2.2     pao   1.2    pao   0.9
stx-vi.ark    chs  38.3    MIA  26.4     mia  25.5    mia  25.1
svo2-ru.ark   lpp  22.4    DME   5.8     fra  41.1    fra  38.2
swu-kr.ark    hkg  77.6    ICN   6.5   qsin1  76.9   nrt2  38.5
syd3-au.ark   syd   4.8    SYD   0.9     syd   1.0    syd   0.5
tij-mx.ark    lax  15.8    LAX   6.6   qlax1   6.6    lax   5.6
tlv-il.ark    mil  80.2    MRS  43.6     tlv   2.0   mil1  50.5
tlv3-il.ark   mil  64.2    TLV   2.1     tlv   2.1   tlv1   3.0
tnr-mg.ark                 JNB 215.7                  jnb 199.1
tpe-tw.ark    tpe  10.3    TPE   5.2     tpe   3.1    sin  74.4
vdp-dk.ark    lpp  21.9    CPH   5.6     arn  13.4   cph1   5.4
vie-at.ark    fra  22.5    FRA  13.8     vie   2.1   prg1   6.5
waw-pl.ark    waw  20.1    HAM  31.3   qwaw2   1.3   wrw1   0.9
wbu-us.ark    cbf  13.4    DEN   2.7     den   1.8   den1   1.5
wlg2-nz.ark   syd  35.3    AKL  17.0    akl2  16.5    syd  32.4
ygk-ca.ark    yyz  29.5    YYZ  16.7     iad  35.5    yyz  16.1
yyc-ca.ark    dls  26.5    YYC   4.6     sea  21.8    yvr  16.7
zrh-ch.ark    zrh  14.5    ZRH   1.1    zrh2   0.6   mil1   8.1
zrh2-ch.ark   zrh  16.2    ZRH   1.1    zrh2   1.0    fra   6.8
zrh4-ch.ark   zrh  15.8    ZRH   1.5    zrh2   0.8   mil1   9.0


ITDK 2024-02

April 23rd, 2024 by Matthew Luckie

CAIDA has released the 2024-02 Internet Topology Data Kit (ITDK), the 24th ITDK in a series published over the past 14 years. In the year since the 2023-03 release, CAIDA has expanded its Ark platform with both hardware and software vantage points (VPs), and re-architected the ITDK probing software. We have been busy modernizing the software to enable us to collect ITDK snapshots more regularly, as well as annotate the router-level Internet topology graph with more features.

For IPv4, the ITDK probing software is based primarily around two reliable alias resolution techniques. The first, MIDAR, probes for IPID behavior that suggests that responses from different IP addresses had IPID values derived from a single counter, and thus the addresses are assigned to the same router. This inference is challenging because of the sheer number of router addresses observed in macroscopic Internet topologies, and the IPID value is held in a 16-bit field, requiring sophisticated probing techniques to identify distinct counters.  The second, iffinder, probes for common source IP addresses in responses to probes sent to different target IP addresses.

In the past few months, we have replaced the MIDAR and iffinder probing component on the Ark VPs to use alias resolution primitives present in scamper (specifically, the midarest, midardisc, and radargun primitives). We used the recently released scamper python module, and 902 lines of python, which executes on a single machine at CAIDA to coordinate the probing from many VPs.

The following table provides statistics illustrating the growth of the ITDK over the past year, driven by the expansion of Ark VPs. Overall, we increased the number of Ark VPs providing topology data from 93 to 142, the number of addresses probed from 2.6 to 3.6M, doubled the number of VPs that we use for alias resolution probing, and found aliases for 50% more addresses than a year ago.  Note that we use the term “node” to distinguish between our router inferences, and the actual routers themselves.  By definition all routers have at least two IP addresses; our “nodes with at least two IPs” are the subset of routers we were able to observe with that property.

2023-03 2024-02
Number of addresses probed: 2.64M 3.58M
Number of ark VPs: 93 142
Number of countries: 37 52
Alias resolution:
Number of ark VPs for MIDAR: 55 101
Number of ark VPs for iffinder: 46 101
Nodes with at least two IPs: 75,660 107,976
Addresses in nodes with at least two IPs:  284,479 425,964

For 2024-02, we also evaluated the gains provided by SNMPv3 probing, following a paper published in IMC 2021 that showed many routers return a unique SNMP Engine ID in response to a SNMPv3 request; the basic idea is that different IP addresses returning the same SNMPv3 Engine ID are likely aliases.  Of the 3.58M addresses we probed, 672K returned an SNMPv3 response.  We inferred that IP addresses belonged to the same router when they return the same SNMP Engine ID, the size of the engine ID was at least 4 bytes, the number of engine boots was the same, and the router uptime was the same; we did not use the other filters in section 4.4 of the IMC paper.  This inferred 47,770 nodes with at least two IPs, many of which were shared with existing nodes found with MIDAR + iffinder. In total, when we combined MIDAR, iffinder, and SNMP probing, we obtained a graph with 124,857 nodes with at least two IPs, covering 515,524 addresses. We are including both the MIDAR + iffinder and MIDAR + iffinder + SNMP graphs in ITDK 2024-02.

Our ITDK also includes an IPv6 graph derived from speedtrap, which infers that IPv6 addresses belong to the same router if the IPID values in fragmented IPv6 responses appear to be derived from a single counter, and a graph derived from speedtrap and SNMP. For IPv6, the gains provided by SNMP are more significant, as the effectiveness of the IPv6 IPID as an alias inference vector wanes.  Of the 929K IPv6 addresses we probed, 68K returned an SNMPv3 response.

2023-03 2024-02
Number of addresses probed: 592K 929K
Number of ark VPs: 36 54
Number of countries: 18 25
Speedtrap output:
Nodes with at least two IPs: 4,945 4,129
Addresses in nodes with at least two IPs: 12,638 10,886
Speedtrap + SNMP output:
Nodes with at least two IPs: 8,935
Addresses in nodes with at least two IPs:  – 35,164

Beyond the alias resolution, the nodes are also annotated with their bdrmapIT-inferred operator (expressed as an ASN) as well as an inferred geolocation. We look up the PTR records of all router IP addresses with zdns, following CNAMEs where they exist, and provide these names as part of the ITDK.  For router geolocation, we used a combination of DNS-based heuristics, IXP geolocation (routers connected to an IXP are likely located at that IXP), and Maxmind GeoLite2.

We inferred DNS-based geolocation heuristics using RTT measurements from 148 Ark VPs in 52 countries to constrain Hoiho, which automatically infers naming conventions in PTR records as regular expressions, and covered 819 different suffixes (e.g.,  ^.+\.([a-z]+)\d+\.level3\.net$ and ^.+\.([a-z]{3})\d+\.[a-z\d]+\.cogentco\.com$ extract geolocation hints in hostnames for Level3 and Cogent in the above figure). There is no dominant source of geohint observed in these naming conventions; 443 (54.1%) embedded IATA airport codes (e.g. IAD, WAS for the Washington D.C. area), 310 (37.9%) embedded place names (e.g. Ashburn for Ashburn, VA, US), 87 (10.6%) embedded the first six characters of a CLLI code (e.g. ASBNVA for Ashburn), and 12 (1.4%) embedded locodes (e.g. USQAS for Ashburn, VA, US). Interestingly, the operators that used CLLI and locodes had conventions that were more congruent with observed RTT values than operators that used IATA codes or place names.  For the nodes in the ITDK, hoiho provided a geolocation inference for 127K, IXP provided a geolocation inference for 14K, and maxmind covered the remainder.  The rules we inferred are usable via CAIDA’s Hoiho API.

ITDKs older than one year are publicly available, and ITDK 2024-02 is available to researchers and CAIDA members, after completing a simple form for access.

Acknowledgment: We are grateful to all of the Ark hosting sites, MaxMind’s freely available geolocation database, and academic research access to Iconectiv’s CLLI database to support this work.

A First Look at Suspicious IRR Records

February 15th, 2024 by Ben Du

The Internet Routing Registry (IRR) is a set of distributed databases used by networks to register routing policy information and to validate messages received in the Border Gateway Protocol (BGP). 


First deployed in the 1990s, the IRR remains the most widely used database for routing security purposes, despite the existence of more recent and more secure alternatives such as the Resource Public Key Infrastructure (RPKI). Yet, the IRR lacks a strict validation standard and the limited coordination across different database providers can lead to inaccuracies. Moreover, it has been reported that attackers have begun to register false records in the IRR to bypass operators’ defenses when launching attacks on the Internet routing system, such as BGP hijacks. 


In our paper, IRRegularities in the Internet Routing Registry, we at CAIDA/UC San Diego, in collaboration with Georgia Tech and Stanford, proposed a workflow to identify suspicious IRR records. In this post, we succinctly describe how we quantified the inconsistencies across all IRR databases, identified likely suspicious IRR records, and validated our results against relevant studies.


Reported false IRR records


Each IRR database is managed independently under different policies and registration processes. The five RIRs (RIPE, ARIN, APNIC, AFRINIC, and LACNIC) manage authoritative IRR databases. Routing information registered in those IRR databases undergoes a validation process against the address ownership information to ensure correctness. IRR databases operated by other institutions are non-authoritative IRR databases and are not strictly validated.


To increase the likelihood of launching a successful BGP hijack attack, malicious actors may inject false records into non-authoritative IRR databases. There have been reported cases of successful BGP hijacking attempts that also abused the IRR.


In one prominent case, an attacker successfully hijacked Amazon’s address space that was used to host Celer Network’s cryptocurrency exchange website. The attacker, AS209243 (QuickHost.uk), pretended to be an upstream provider of AS16509 (Amazon) by registering false objects in ALTDB. In a different case, AS207427 (GoHosted.eu) registered false IRR objects for 3 UCSD-announced prefixes and hijacked those prefixes in BGP for more than a month.


Workflow to identify suspicious IRR records


Following the diagram in Figure 1, we consider an IRR record in a non-authoritative database suspicious if it satisfies the following conditions:

  1. The IRR record conflicted with corresponding records (containing the same prefix but different origin) in the authoritative IRR database.
  2. The prefix in the IRR record from step 1 was originated in BGP by multiple ASes, one of which is the AS in the IRR record.


We validate our inferred suspicious IRR records with the Resource Public Key Infrastructure (RPKI), a more recent and secure alternative of the IRR. We also check if the origin ASes in the suspicious IRR records were classified as serial hijacker ASes by Testart et al. published at IMC 2019.

Figure 1. Workflow to identify suspicious IRR records (solid arrows) and methods to validate our results (dotted arrows).


Baseline: Inconsistency across IRR databases


We discuss the results of the first step in the workflow above. To understand the baseline characteristics of the IRR databases, we analyze the consistency between all pairs of IRR databases.


Figure 2 shows the percentage of records with the same prefix but different origin ASes between pairs of IRRs. We found that most IRR databases have mismatching records with one another, consistent with persistent neglect by IRR users and thus an increasing number of outdated entries. We also noticed instances where a company registered records in multiple IRR databases, but only updated the records in one IRR database, causing inter-IRR inconsistency. 


Most surprising were the mismatching records between pairs of authoritative IRR databases, since each RIR only allows registration of records containing address blocks managed by that RIR, which do not overlap with each other. We speculate that those mismatching records correspond to address space that was transferred across RIRs, and the address owner from the previous RIR did not remove the outdated object. As of January 2024, two months since our paper was published, the RIRs have removed all inconsistent IRR records in their authoritative databases. We provide the updated results on github https://github.com/CAIDA/IRR-IRRegularities-Analysis

Figure 2. Fraction of inconsistent records in the IRR on the Y-axis with respect to the IRR on the X-axis. The denominators shown in Figure 1b in the paper.


Suspicious records in RADB


Out of 1.5 million RADB records (1.2 million unique prefixes), we identified 34,199 potentially suspicious records. We further checked the RPKI consistency of those records. Out of those 34,199 records, 4,082 records had a mismatching ASN, 144 had prefixes that were too specific, and 9,450 had no matching ROA in RPKI. To further narrow down the list of suspicious IRR records, we removed the ones whose AS appear in other RPKI-consistent records (assuming those ASes were unlikely to be malicious), leaving 6,373 suspicious records. Network operators who use IRR-based filtering should carefully consider those suspicious records.


We also compared our list of 34,199 suspicious records with the list of serial hijackers from Testart et al.  and found 5,581 records registered by 168 serial hijacker ASes. We found one of those ASes to be a small US-based ISP with 10 customers according to CAIDA’s AS Rank. Another serial hijacker AS was a European hosting provider with more than 100 customers, which was also known to be exploited by attackers to abuse the DNS system. However, networks may have registered both suspicious and benign records, which can complicate the inference of suspicious IRR records.



We provided a first look at inconsistencies across IRR databases and proposed an approach to infer suspicious activities in the IRR without external sources of ground truth. We found IRR databases prone to staleness and errors, confirming the importance of operators transitioning to RPKI-based filtering. We hope this work inspires new directions in automating the detection of abuse of IRRs, ideally in time to prevent or thwart an attacker’s ultimate objective. We publicly provide our analysis code on Github https://github.com/CAIDA/IRR-IRRegularities-Analysis with more recent sample data and results.

Developing active Internet measurement software locally to run on Ark

January 24th, 2024 by Matthew Luckie

In the first part of our blog series, we introduced our brand-new python module for scamper, the packet-prober underpinning much of Ark’s ongoing measurements. One aspect that we highlighted was the ability for potential users of Ark to develop their code locally, before running it on the Ark platform. When I develop measurement applications, I use a couple of local Raspberry Pis and my own workstation to get the code correct, and then copy the code to the CAIDA system to run the experiment using available Ark vantage points. The goal of this blog article is to describe different ways to locally develop your measurement experiment code.

Example #1: Starting small with one scamper process.

The easiest way to begin is with one scamper process running on the same system where you develop your python code. With scamper installed (we recommend that you use a package listed on the scamper website), start a scamper process, and make it available for measurement commands on a Unix domain socket. For example, you might run scamper as follows:

$ scamper -U /tmp/scamper -p 100

This will create a Unix domain socket to drive scamper at /tmp/scamper, and tell scamper that it can probe at up to 100 packets/second. You can adjust these parameters to what is appropriate locally.

You can then develop and debug your measurement code in Python. To use this scamper process, your Python code might begin as follows:

01 from scamper import ScamperCtrl
03 # use the scamper process available at /tmp/scamper
04 ctrl = ScamperCtrl(unix="/tmp/scamper")
06 # do a simple ping to and print the outcome
07 o = ctrl.do_ping("")
08 if o.min_rtt is not None:
09   print(f"{o.min_rtt.total_seconds()*1000):.1f} ms")
10 else:
11   print("no reply")

Example #2: Coordinating measurements among VPs.

Once you are comfortable using the python module with a single local scamper instance, you might want to test your code with multiple scamper instances, each representing a distinct vantage point. The scamper software includes the sc_remoted interface to support that. sc_remoted has features for authenticating endpoints with TLS, but you might choose to initially operate endpoints without the complexity of TLS.

sc_remoted listens on a port for inbound scamper connections, and makes Unix domain sockets — one for each VP — available in a nominated directory. The best idea is to create an empty directory for these sockets. You might run sc_remoted as follows:

$ mkdir -p /path/to/remote-sockets
$ sc_remoted -U /path/to/remote-sockets -P 50265

The first command creates the directory, because sc_remoted will not create that directory for you. The second command starts sc_remoted listening on port 50265 for incoming scamper connections, and will place Unix domain sockets in /path/to/remote-sockets as they arrive.  Note, we use /path/to as a placeholder to the actual path in your local file system that is appropriate for your environment; you might put these sockets in a directory somewhere in your home directory, for example.

Then, on the systems that you want to act as vantage points, the following command:

$ scamper -p 100 -R -M foo.bar

will (1) start a scamper process, (2) tell it that it can probe at up to 100 packets-per-second, (3) connect it to the specified IP address and port to receive measurement commands from, and (4) tell it to identify itself as “foo.bar” to the remote controller. If you go into /path/to/remote-sockets, you might see the following:

$ cd /path/to/remote-sockets
$ ls -l
total 0
srwx------ 1 mjl mjl 0 Jan 22 16:57 foo.bar-

This socket represents the scamper process you just started. The filename begins with foo.bar, the parameter that you gave to scamper to identify itself. After the dash is the IP address and port number that the remote controller observed the remote system coming from. You can connect as many additional scamper instances as you like, and you will see them listed in the directory individually. You should name each differently with something meaningful to you (foo.bar, bar.baz, etc) so that you can identify them on the system on which you’re writing your python code.

The python code we wrote in Example #1 above might be modified as follows:

01 import sys
02 from scamper import ScamperCtrl
04 if len(sys.argv) != 2:
05   print("specify path to unix domain socket")
06   sys.exit(-1)
08 # use the remote scamper process available at the specified location
09 ctrl = ScamperCtrl(remote=sys.argv[1])
11 # do a simple ping to and print the outcome
12 o = ctrl.do_ping("")
13 if o.min_rtt is not None:
14   print(f"{o.min_rtt.total_seconds()*1000):.1f} ms")
15 else:
16   print("no reply")

And run as:

$ python ping.py /path/to/remote-sockets/foo.bar-\:12369

If you have multiple remote-sockets in the directory, you can add them individually, or use all sockets in the directory. For example:

01 import sys
02 from datetime import timedelta
03 from scamper import ScamperCtrl
05 if len(sys.argv) != 3:
06   print("usage: single-radius.py $dir $ip")
07   sys.exit(-1)
09 ctrl = ScamperCtrl(remote_dir=sys.argv[1])
10 for i in ctrl.instances():
11   ctrl.do_ping(sys.argv[2], inst=i)
13 min_rtt = None
14 min_vp = None
15 for o in ctrl.responses(timeout=timedelta(seconds=10)):
16   if o.min_rtt is not None and (min_rtt is None or min_rtt > o.min_rtt):
17     min_rtt = o.min_rtt
18     min_vp = o.inst
20 if min_rtt is not None:
21   print(f"{min_vp.name} {(min_rtt.total_seconds()*1000):.1f} ms")
22 else:
23   print(f"no responses for {sys.argv[2]}")

and run this command:

$ python single-radius.py /path/to/remote-sockets

We encourage you to reach out via email if you have questions about using the module. In the first instance, you can email ark-info at caida.org.

Towards a Domain Specific Language for Internet Active Measurement

January 16th, 2024 by Matthew Luckie

This is the first in a series of essays about CAIDA’s new effort to reduce the barrier to performing a variety of Internet measurements.

Network operators and researchers often require the ability to conduct active measurements of networks from a specific location in order to understand some property of the network. However, obtaining access to a vantage point at a given location is challenging, as there can be significant trust barriers that prevent access: a platform operator has to provide vantage point hosts with guarantees about the activity that their vantage points will exhibit, and a platform operator has to trust that users will stay within those guarantees. Current access control to active measurement infrastructure has two extremes: access that allows for trusted users to run arbitrary code, and API access that allows arbitrary users to schedule a (limited) set of measurements, and obtain their results. Prior research thrusts in active measurement design have focused on interfaces that allow a user to request a host to send arbitrary packets, leaving the implementation of the measurement to the user. However, this design pattern limits the guarantees that a platform operator can provide a vantage point host regarding what their vantage point will actually do. A domain-specific language for conducting active measurements can alleviate these concerns because it (1) allows a platform operator to specify the measurements that a user can run, and communicate to the host what their vantage point will do, (2) provides users reference implementations of measurement applications that act as building blocks to more complex measurements.

Over the past six months, in consultation with members of the active measurement community, CAIDA has been working towards a next-generation measurement infrastructure, built on the existing Archipelago (Ark) platform. One aspect of this platform is the notion of a researcher development environment that allows for complex, distributed, and reactive measurements built on a well-defined set of measurement primitives. In an effort to make the Ark platform easier to use for measurement researchers, while also providing important access control, CAIDA has developed the scamper python module that provides an interface to the measurement primitives available on each of the Ark nodes. Today, we are releasing the source code of that module, so that researchers can develop and test complex measurements locally, before running vetted experiments on Ark.

Scamper architecture chart

The architecture of scamper. Measurement tasks are supplied from one or more input sources, including from an input file, from the command line, or from a control socket.

The module provides user-friendly interfaces to existing scamper functionality. We illustrate the module with some examples.

Example #1: Implementation of RTT-based geolocation measurement

The following is an implementation of the well-known single-radius measurement, which conducts delay measurements to an IP address from a distributed set of vantage points, and reports the shortest of all the RTTs obtained with the name of the monitor, which on Ark, is derived from its location (e.g., lax-us, hlz2-nz, ams-nl, and so on). Researchers and operators might use this technique to understand approximately where a system is located, using the RTT constraint of the vantage point that reports the shortest delay.

01 import sys
02 from datetime import timedelta
03 from scamper import ScamperCtrl
05 if len(sys.argv) != 3:
06   print("usage: single-radius.py $dir $ip")
07   sys.exit(-1)
09 ctrl = ScamperCtrl(remote_dir=sys.argv[1])
10 for i in ctrl.instances():
11   ctrl.do_ping(sys.argv[2], inst=i)
13 min_rtt = None
14 min_vp = None
15 for o in ctrl.responses(timeout=timedelta(seconds=10)):
16   if o.min_rtt is not None and (min_rtt is None or min_rtt > o.min_rtt):
17     min_rtt = o.min_rtt
18     min_vp = o.inst
20 if min_rtt is not None:
21   print(f"{min_vp.name} {(min_rtt.total_seconds()*1000):.1f} ms")
22 else:
23   print(f"no responses for {sys.argv[2]}")

This implementation takes two parameters (lines 5-7) — a directory that contains a set of unix domain sockets, each of which represents an interface to a single vantage point, and an IP address to probe. On line 9, we open an interface (represented by a ScamperCtrl object) that contains all of the vantage points in that directory. We then send a ping measurement to each of the vantage point instances (lines 10-11). These ping measurements operate in parallel — the ping measurements on each of the nodes operate asynchronously. We then collect the results of the measurements (lines 13-18), noting the minimum observed RTT, and the vantage point where it came from. We pass a 10-second timeout on line 15, so that a vantage point that experiences an outage after we send the measurements does not hold up the whole experiment. Finally, on lines 20-23, we print the result of the measurement.

Example #2: RTTs to authoritative name servers of a specific domain

The next example shows a more complex measurement. Let’s say we want to know the RTTs to the authoritative name servers for a zone from a single vantage point.

01 import sys
02 from datetime import timedelta
03 from scamper import ScamperCtrl
05 if len(sys.argv) != 3:
06   print("usage: authns-delay.py $vp $zone")
07   sys.exit(-1)
09 ctrl = ScamperCtrl(remote=sys.argv[1])
11 # get the list of NS for the zone
12 o = ctrl.do_dns(sys.argv[2], qtype='NS', wait_timeout=1, sync=True)
14 # issue queries for the IP addresses of the authoritative servers
15 ns = {}
16 for rr in o.ans():
17   if rr.ns is not None and rr.ns not in ns:
18     ns[rr.ns] = 1
19     ctrl.do_dns(rr.ns, qtype='A', wait_timeout=1)
20     ctrl.do_dns(rr.ns, qtype='AAAA', wait_timeout=1)
22 # collect the unique addresses out of the address lookups
23 addr = {}
24 for o in ctrl.responses(timeout=timedelta(seconds=3)):
25   for a in o.ans_addrs():
26     addr[a] = o.qname
28 # collect RTTs for the unique IP addresses
29 for a in addr:
30   ctrl.do_ping(a)
31 for o in ctrl.responses(timeout=timedelta(seconds=10)):
32   print(f"{addr[o.dst]} {o.dst} " +
33         (f"{(o.min_rtt.total_seconds() * 1000):.1f}"
34          if o.min_rtt is not None else "???"))

This implementation takes two parameters (lines 5-7) — a single vantage point, and a zone name to study. As before, we open an interface to that VP on line 9, and then issue a DNS query for the authoritative name servers for the zone (line 12).

There are a couple of interesting things to note about line 12. First, we do not pass a handle representing the VP instance to the do_dns method, as the ScamperCtrl interface only has a single instance associated with it — it is smart enough to use that single instance for the measurement. Second, we pass sync=True to make the measurement synchronous — the method does not return until it has the results of that single measurement. This is shorter (in lines of code) and more readable than issuing an asynchronous measurement and then collecting the single result. Then, we issue asynchronous queries for the IPv4 and IPv6 addresses for the name servers returned (lines 14-20) and send ping measurements for each of the addresses (lines 29-30). Finally, we print the names of the nameservers, their IP addresses, and the minimum RTT observed to each.

The scamper python module supports most of the measurement primitives currently available in scamper: ping, traceroute, DNS query, alias resolution, HTTP, and simple UDP probes. We’ve used the module internally to: (1) reproduce the data collection methodology of a recent router fingerprinting method; (2) study the deployment of Netflix fast.com speed test endpoints, (3) implement MIDAR; (4) study anycast open resolvers, and (5) monitor serial number changes of zones amongst a set of nameservers authoritative for a zone. One important feature of our module is that it provides a python interface, which means that you can use our measurement module alongside existing python modules that parse JSON, etc. The documentation for the module is publicly available and we will write additional blog entries in the coming days that show more of its features, in a digestible form.

In the short term, measurement researchers can request access to the infrastructure to run vetted experiments by emailing ark-info at caida.org. Note: the access does not provide a login on any of the Ark nodes. Rather, it provides access to a system that can access the measurement primitives of the vantage points, as illustrated above. Again, we are releasing the python module to allow researchers to develop and debug experiments locally, before running them on Ark. We hope that this approach provides a convenient development lifecycle for researchers, as the CAIDA system you will have access to when running the experiments will not necessarily have the local development environment that you are accustomed to.

The module itself is written in cython, providing a wrapper around two C libraries in scamper that do much of the heavy lifting. You build and install the module using the instructions on the scamper website, install the module using the Ubuntu PPA (preferred if you are using Ubuntu), or install the module using one of the packages available on other operating systems as these become available.

CAIDA contributions to ACM’s Internet Measurement Conference (IMC) 2023

November 14th, 2023 by CAIDA Webmaster

ACM’s Internet Measurement Conference (IMC) is an annual highly selective venue for the presentation of Internet measurement and analysis research. The average acceptance rate for papers is around 25%. CAIDA researchers co-authored three papers and one poster that was be presented at the IMC conference in Montreal, Quebec on October 26 – 28, 2023. We link to these publications below.

On the Importance of Being an AS: An Approach to Country-Level AS Rankings
Bradley Huffaker, Alexander Marder, Romain Fontugne, kc claffy. ACM Internet Measurement Conference (IMC), 2023.
Recent geopolitical events demonstrate that control of Internet infrastructure in a region is critical to economic activity and defense against armed conflict. This geopolitical importance necessitates novel empirical techniques to assess which countries remain susceptible to degraded or severed Internet connectivity because they rely heavily on networks based in other nation states. Currently, two preeminent BGP-based methods exist to identify influential or market-dominant networks on a global scale-network-level customer cone size and path hegemony–but these metrics fail to capture regional or national differences.

We adapt the two global metrics to capture country-specific differences by restricting the input data for a country-specific metric to destination prefixes in that country. Although conceptually simple, our study required tackling methodological challenges common to most Internet measurement research today, such as geolocation, incomplete data, vantage point access, and lack of ground truth. Restricting public routing data to individual countries requires substantial downsampling compared to global analysis, and we analyze the impact of downsampling on the robustness and stability of our country-specific metrics. As a measure of validation, we apply our country-specific metrics to case studies of Australia, Japan, Russia, Taiwan, and the United States, illuminating aspects of concentration and interdependence in telecommunications markets. To support reproducibility, we will share our code, inferences, and data sets with other researchers.

IRRegularities in the Internet Routing Registry
Ben Du, Gautam Akiwate, Cecilia Testart, Alex C. Snoeren, kc claffy, Katherine Izhikevich, Sumanth Rao. ACM Internet Measurement Conference (IMC), 2023.
The Internet Routing Registry (IRR) is a set of distributed databases used by networks to register routing policy information and to validate messages received in the Border Gateway Protocol (BGP). First deployed in the 1990s, the IRR remains the most widely used database for routing security purposes, despite the existence of more recent and more secure alternatives. Yet, the IRR lacks a strict validation standard and the limited coordination across diferent database providers can lead to inaccuracies. Moreover, it has been reported that attackers have begun to register false records in the IRR to bypass operators’ defenses when launching attacks on the Internet routing system, such as BGP hijacks. In this paper, we provide a longitudinal analysis of the IRR over the span of 1.5 years. We develop a workflow to identify irregular IRR records that contain conflicting information compared to different routing data sources. We identify 34,199 irregular route objects out of 1,542,724 route objects from November 2021 to May 2023 in the largest IRR database and find 6,373 to be potentially suspicious.

Coarse-grained Inference of BGP Community Intent
Thomas Krenc, Alexander Marder, Matthew Luckie, kc claffy. ACM Internet Measurement Conference (IMC), 2023.
BGP communities allow operators to influence routing decisions made by other networks (action communities) and to annotate their network’s routing information with metadata such as where each route was learned or the relationship the network has with their neighbor (information communities). BGP communities also help researchers understand complex Internet routing behaviors. However, there is no standard convention for how operators assign community values, and significant efforts to scalably infer community meanings have ignored this high-level classification. We discovered that doing so comes at a significant cost in accuracy, of both inference and validation. To advance this narrow but powerful direction in Internet infrastructure research, we design and validate an algorithm to execute this first fundamental step: inferring whether a BGP community is action or information. We applied our method to 78,480 community values observed in public BGP data for May 2023. Validating our inferences (24,376 action and 54,104 informational communities) against available ground truth (6,259 communities) we find that our method classified 96.5% correctly. We found that the precision of a state-of-the-art location community inference method increased from 68.2% to 94.8% with our classifications. We publicly share our code, dictionaries, inferences, and datasets to enable the community to benefit from them.

CAIDA also contributed to one extended abstract:

Empirically Testing the PacketLab Model
Tzu-Bin Yan, Zesen Zhang, Bradley Huffaker, Ricky K. P. Mok, kc claffy, Kirill Levchenko. ACM Internet Measurement Conference (IMC) Poster, 2023.
PacketLab is a recently proposed model for accessing remote vantage points. The core design is for the vantage points to export low-level network operations that measurement researchers could rely on to construct more complex measurements. Motivating the model is the assumption that such an approach can overcome persistent challenges such as the operational cost and security concerns of vantage point sharing that researchers face in launching distributed active Internet measurement experiments. However, the limitations imposed by the core design merit a deeper analysis of the applicability of such model to real-world measurements of interest. We undertook this analysis based on a survey of recent Internet measurement studies, followed by an empirical comparison of PacketLab-based versus native implementations of common measurement methods. We showed that for several canonical measurement types common in past studies, PacketLab yielded similar results to native versions of the same measurements. Our results suggest that PacketLab could help reproduce or extend around 16.4% (28 out of 171) of all surveyed studies and accommodate a variety of measurements from latency, throughput, network path, to non-timing data.

A summary of the EU’s Digital Services Act and the role of academic research

August 29th, 2023 by David Clark and kc claffy

A new working paper from the CAIDA GMI3S project summarizes key aspects of the European Union’s Digital Services Act (DSA), providing insight into how the EU plans to regulate large online platforms and services.

The DSA is intended to create a safer and more transparent online environment for consumers by imposing new obligations and accountability on companies like Meta, Google, Apple, and Amazon. The regulations target issues like illegal content, disinformation, political manipulation, and harmful algorithms. Systemic societal risks, like disinformation and political influence campaigns, are a major focus of the regulations.

Some key takeaways from the DSA overview:

  • Very large online platforms and search engines face the most stringent requirements, like risk assessment audits, transparency reporting, and independent oversight.
  • The rules apply to any company offering services within the EU market, regardless of where they are based. This prevents tech giants from circumventing the law.
  • The Act establishes clear legal definitions and penalties around illegal content like hate speech, terrorist propaganda, and child sexual abuse material.
  • Covered platforms must assess and mitigate these and other systemic threats. They are expected to assess how the design and implementation of their service, and manipulation and use of it, including violations of terms of service, contribute to such risks.
  • Providers serving advertising must clearly identify its sponsor (who paid for it on and on whose behalf), and the main parameters used to its target audience, and where applicable, how to change them.
  • Providers of recommendation systems must provide the main parameters including the most important criteria determining what is recommended, and the reasons for the importance of these criteria.
  • A critical component is requiring platforms to provide access to data for vetted independent researchers studying systemic risks and harms. This enables evidence-based analysis of problems and solutions.

The DSA signals that the EU is taking a hard line on enforcing accountability and responsible practices for dominant online platforms that have largely operated without oversight. The success of the regulations will depend on effective coordination and enforcement across the EU’s member states. However the Act provides a potential model for balancing innovation and consumer protection in the digital marketplace. Consumers may benefit from transparency and accountability of manipulative algorithms, more control over data, and protections against online harms.

Importantly, the data access mandate signals the EU’s commitment to leveraging academic expertise in shaping a healthier digital ecosystem. Researchers will be able to investigate systemic issues like algorithmic bias, political polarization, misinformation dynamics, and the mental health impacts of social media.

This DSA working paper provides valuable context on this ambitious attempt to regulate the digital economy.

To read further:

CAIDA’s 2022 Annual Report

July 10th, 2023 by kc

The CAIDA annual report summarizes CAIDA’s activities for 2022 in the areas of research, infrastructure, data collection and analysis. The executive summary is excerpted below:
Read the rest of this entry »

Hoiho API (Holistic Orthography of Internet Hostname Observations)

February 13th, 2023 by Bradley Huffaker

In December 2021, CAIDA published a method and system to automatically learn rules that extract geographic annotations from router hostnames. This is a challenging problem, because operators use different conventions and different dictionaries when they annotate router hostnames. For example, in the following figure, operators have used IATA codes (“iad”, “was”), a CLLI prefix (“asbnva”), a UN/LOCODE (“usqas”), and even city names (“ashburn”, “washington”) to refer to routers in approximately the same location — Ashburn, VA, US. Note that “ash” (router #4) is an IATA code for Nashua, NH, US, that the operators of he.net and seabone.net used to label routers in Ashburn, VA, US. Some operators also encoded the country (“us”) and state (“va”).

Our system, Hoiho, released as open-source as part of scamper, uses CAIDA’s Macroscopic Internet Topology Data Kit (ITDK) and observed round trip times to infer regular expressions that extract these apparent geolocation hints from hostnames. The ITDK contains a large dataset of routers with annotated hostnames, which we used as input to Hoiho for it infer rules (encoded as regular expressions) that extract these annotations. CAIDA has released these inferred rulesets in recent ITDKs.

Today, CAIDA is launching an API (api.hoiho.caida.org) and web front end (hoiho.caida.org) which returns extracted geographic locations from a user-provided list of DNS names. The API uses the rules that CAIDA infers with each ITDK. For embedded IATA, UN/LOCODE, and city names, the API returns the city name and a lat/long representing the location. For embedded CLLI codes, the API returns the CLLI code; please contact iconectiv for a dictionary that maps CLLI codes to locations.

Try the API out, and let us know if you find it useful!

[HOIHO] Luckie, M., Huffaker, B., Marder, A., Bischof, Z., Fletcher, M., and claffy, k., 2021. “Learning to Extract Geographic Information from Internet Router Hostnames.” ACM SIGCOMM Conference on emerging Networking EXperiments and Technologies (CoNEXT),

Studying Conformance of MANRS Members

January 21st, 2023 by Ben Du

In November 2022, 85% MANRS members were conformant to Action #1 and Action #4.


The Mutually Agreed Norms on Routing Security (MANRS) initiative is an industry-led effort to improve Internet routing security. MANRS encourages participating networks to implement a series of routing security practices.  In our paper, Mind Your MANRS: Measuring the MANRS Routing Ecosystem, we at CAIDA (UC San Diego), in collaboration with Georgia Tech, and IIJ Research Lab, provided the first independent look into the MANRS ecosystem by using publicly available data to analyze the routing behavior of participant networks. MANRS membership has increased significantly in recent years, but our research goal was to get more clarity on the impact of the MANRS initiative on the state of overall Internet routing security.   In this post, we summarize how we characterized the growth of MANRS members, explain our process of analyzing ISP conformance with the MANRS practices we studied, compare RPKI ROA registration status between MANRS and non-MANRS members, and reflect on implications of our analysis for the future of MANRS. 


We first analyzed what types of networks have joined MANRS over time, and whether MANRS members are properly implementing the routing security practices (MANRS conformance).  The two practices (which MANRS calls actions) we focused on in our study are: 

  1. Participating ISPs will register their IP prefixes in a trusted routing database (either Resource Public Key Infrastructure (RPKI) or one of the databases of the Internet Routing Registry (IRR).   This practice is “MANRS Action #4”.
  2. Participating ISPs will use such information to prevent propagation of invalid routing information. This practice is “MANRS Action #1”.


Our paper analyzed the MANRS ecosystem in May 2022. Since MANRS is a growing community, for this post we have updated our analysis using data collected in November 2022 to capture a more recent view of the MANRS ecosystem. We have also published our analysis code here for interested readers to reproduce the analysis using the latest available data.


MANRS growth

We first downloaded a list of MANRS members. The Internet Society kindly provided us the dates when each MANRS participant joined the programs. We found that between 2015 and November 2022, 863 ASes joined MANRS. Over this 7-year period, an additional 12.1% of routed IPv4 address space was originated by MANRS ASes. Plotting growth by ASes and by address space (Figure 1) shows that most of these new ASes were based in the LACNIC region, but that those ASes originated little or no address space into BGP.   



Figure 1 – MANRS participation grew between 2015 and 2022, but the picture looks quite different if measured by number of ASes vs. % of routed address space. 

MANRS Conformance 

We examined whether MANRS (ISP and CDN) members properly implemented MANRS Action #4 and #1 according to the MANRS requirements:

  • To conform to Action #4, members must register at least 90% (100% for CDNs) IP prefixes in IRR or RPKI.
  • To conform to Action #1, members must filter out customer BGP announcements that do not match IRR or RPKI records.


We downloaded BGP prefixes and their IRR/RPKI status from the Internet Health Report (IHR) maintained by IIJ Research Labs. We found that in November 2022, 893 (95.9%) of all 931 MANRS ASes conformed to MANRS Action #4 (prefix registration). Figure 2 shows that in November 2022, 3.7% of the address space originated by MANRS ASes was contained in prefixes that either were not registered or were incorrectly registered in IRR or RPKI. We also conducted case studies of non-conformant MANRS CDN members  and found that one large CDN was not conformant because one of their 7000+ prefixes was RPKI-invalid. Please refer to section 8.4 of the paper for more details. 





Figure 2 – Most ASes participating in MANRS conformed with Action #4, and correspondingly, most of the address space those ASes originated into BGP was IRR or RPKI valid, i.e., had records that matched observed BGP announcements. 


To evaluate whether MANRS members filtered out customer BGP announcements that do not match IRR or RPKI records (Action #1), we downloaded BGP prefixes, their IRR and RPKI statuses, and their upstream ASes from the Internet Health Report. We then calculated the prevalence of IRR/RPKI Invalid prefixes propagated through each MANRS network. 


Figure 3 shows that in November 2022, 790 (84.9%) of 931 MANRS ASes conformed to the MANRS Action #1 . Figure 3 also shows that 141 (15.1%) MANRS ASes did not conform to Action #1. However, not all of the address space propagated by these ASes was incorrectly registered in RPKI or IRR.  In fact, those 141 ASes propagated 96.7% of the address space propagated by MANRS ASes, but only 1.5% of that total was incorrectly registered. In addition, we found that 25 out of 27 MANRS members that are large transit providers (i.e., had > 180 customer ASes) did not fully conform with MANRS Action #1, suggesting that conformance was hard to achieve for networks with complex routing relationships.





Figure 3 – MANRS ASes that did not conform to MANRS Action #1 only propagated a small fraction of address space announced by MANRS ASes that was not IRR or RPKI Valid. (b) shows 95.2% of MANRS-propagated address space was IRR/RPKI Valid despite being propagated by Action #1 non-conformant members.


Are MANRS members more likely to register in RPKI? 

Our study found that, except for a few cases, MANRS organizations tended to conform with the two actions we studied. However, to estimate the impact of the MANRS initiative on the state of routing security, we compared the behavior of MANRS and non-MANRS ASes. 


We first compared these two subsets of ASes in terms of registration of RPKI ROAs of prefixes announced in BGP.  In November 2022, 60.1% of routed IPv4 address space originated by MANRS ASes was covered by RPKI ROAs, compared with only 38.8% of all routed IPv4 addresses covered by ROAs. Figure 5 shows that in November 2022, IPv4 address space originated by MANRS ASes was more likely to be registered in RPKI in all RIR regions except APNIC. In the APNIC region, we found significant RPKI registration by non-MANRS networks from JPNIC and TWNIC, possibly due to local RPKI outreach efforts.  Overall, this difference suggests a positive influence of MANRS members on the adoption of RPKI. 


Similarly, changing the view from routed address space to the originating ASes, we found that in November 2022, MANRS members were more likely to originate at least 80% RPKI Valid prefixes in BGP compared to their non-MANRS counterparts in all RIR regions (Figure 6).


Figure 5 – In November 2022, IPv4 address space originated by MANRS ASes was more likely to be registered in RPKI in all RIR regions except APNIC.


Figure 6 – In November 2022, MANRS ASes were more likely to originate RPKI Valid prefixes than non-MANRS ASes.


Future for MANRS

In November 2022, we found 71 MANRS ASes that registered their prefixes only in IRR but not RPKI. Registering only in an IRR database is less optimal than registering in RPKI, since some IRR databases may contain inaccurate records due to looser validation standards (See our paper IRR Hygiene in the RPKI Era). We recommend that in the future, MANRS  members register in RPKI in addition to IRR databases.  We also recommend that MANRS add a conformance checker to its existing observatory to further motivate its members to maintain good routing security practices. We have published our analysis code to facilitate such conformance checking.