How Global Is Anycast?: Evaluating Performance Gaps

arXiv:2610.40833v1 [cs.NI] 29 Oct 2026

How Global Is Anycast?: Evaluating Performance Gaps Mohammad Shaikh Sadid Khan

Karishma Setya

ASM Rizvi

Department of Computer Science California State University, Fullerton sadidkhan@csu.fullerton.edu

Department of Computer Science California State University, Fullerton ksetya1@csu.fullerton.edu

Department of Computer Science California State University, Fullerton asmrizvi@fullerton.edu

Abstract—Anycast networks aim to reduce latency by directing users to nearby service sites. However, users can still experience high latency even when accessing services via anycast. Some anycast deployments are localized and intended to serve a specific geographic region, while others are globally distributed to support users worldwide, with the expectation of consistently low latency across locations. In practice, latency can remain poor for certain users due to factors such as routing inefficiencies and peering relationships. These inefficiencies are often more pronounced in regions such as Africa, Asia or South America, where Internet connectivity is generally less mature than in North America or Europe. Despite this, there has been no in-depth analysis of anycast performance in these continents. In this paper, we examine the latency experienced by users in these regions when accessing anycast networks. We analyze more than 10,000 anycast prefixes, quantify the severity of latency problems, and show that clients in Africa, Asia, and South America experience substantially higher latency than those in North America. We also investigate the underlying causes of this performance gap. Index Terms—anycast, performance, measurements

I. I NTRODUCTION In an anycast network, multiple sites provide the same service by announcing the same service prefix from all sites [1]. Since multiple sites announce the same service prefix, a user usually has multiple routing paths to different anycast sites. Based on the Border Gateway Protocol (BGP) path selection algorithm, a single path is selected from multiple paths to a single anycast site [2]. The BGP path selection process is generally expected to route users to the closest site, since—when other attributes like weights or local preferences are equal—BGP prefers the shortest autonomous system path (AS-path). Domain Name System (DNS) and Content Delivery Networks (CDNs) widely use anycast to serve clients from nearby locations by announcing the same prefix from multiple geographic sites. However, because BGP is not performanceaware and instead selects routes based on fixed attributes, anycast routing does not always deliver optimal latency [3]. Network operators deploy both local and global anycast networks [3], [4]. Local anycast is intended to serve nearby users, while global anycast is designed to serve users worldwide. However, in practice, some anycast networks do not achieve this goal for various reasons. First, Internet connectivity is generally more mature in regions such as North America and Europe than in many other continents, making it difficult for

operators to deliver comparable latency everywhere. Second, network administrators often prioritize deployments in locations with larger user bases and greater revenue potential. Third, undesirable routing behavior or peering relationships can make it difficult for network operators to provide comparable latency across all continents. Anycast polarization occurs when users in the same geolocation experience very different performance—some receive low latency while others see high latency [5]. Prior work shows that this phenomenon is not uncommon in regions such as North America and Europe. We hypothesize that polarization is even more severe in other continents where Internet infrastructure is less mature. Also, network operators deploy local anycast sites intended to serve specific autonomous systems (ASes) [1], [3]. This can further increase the likelihood of polarization. In this paper, we make three contributions. First, we show how ASes announce prefixes from different anycast sites (section IV). We demonstrate that a single AS can operate multiple anycast prefixes with markedly different performance. These cross-prefix differences highlight which prefixes are most important for clients in particular geographic locations (subsection IV-D). Second, to our knowledge, we are the first to quantify the degree of anycast polarization in regions such as Africa, Asia and South America and show that polarization is especially pronounced in these continents (section V). In Africa, even under the best-case scenario, more than 10% of probes experience latencies above 50 ms. Third, we examine anycast network configurations to identify the causes of polarization (section VI). We find that large transit providers sometimes lack connectivity to nearby anycast sites, causing clients to be routed to distant sites and incur higher latency. We also characterize how large ASes and local peering relationships influence anycast performance along with a comparative analysis among large ASes. II. R ELATED W ORK Prior work has evaluated anycast networks extensively. These studies examine when anycast outperforms unicast and when it leads to suboptimal site selection [6], identify routingpolicy and route-selection effects that contribute to anycast inefficiencies [1], and quantify how the number of anycast sites influences performance [7]. Other work highlights the

III. M ETHODOLOGY We at first get the list of anycast addresses, and then make ping measurements from RIPE probes to compare latencies across continents. We also make traceroutes to selective target prefixes to understand the underlying causes of polarization. We use DNS logs to verify that real clients exist behind the anycast prefixes. Getting anycast targets: We obtain the anycast addresses from a prior study [12] that used active measurements to identify anycast IPs. To ensure high confidence in these addresses, we follow their recommendations and retain only those prefixes satisfying (AB > 3)||(GCD > 1) value, where AB means anycast based approach, and Great-Circle Distance (GCD) is the latency based approach. RIPE measurements to the anycast targets: After compiling the anycast address list, we use RIPE Atlas [13] probes to measure latency to these anycast targets. We select at least 150 randomly selected probes per continent and run latency measurements to all anycast addresses, and then compare results across continents. To avoid bias in probe selection, we choose probes from a diverse set of autonomous systems and countries. Since probes within the same AS and country often exhibit similar routing behavior, we include at most two when multiple such probes are available. Otherwise, we prioritize country diversity within each continent to capture performance across different geolocations.

LLC lou

C

2500

dfla

nc.

I re, A

z ma

co on.

9 50

Inc

.

. Inc e tly, flar Fas Cloud

do Lon

n, L

Afil

LC

ia

.

nc s, I

.o. c .z.o , In . .pl Sp c LLC tuate n I wa e , l c o g i A Fly. Naz Goo Net

AS 54

5

AS 16

AS 13

69 AS 39

33

0

m,

11 AS 3 20 92 42 AS 12 04 1 AS 40 50 9 AS 15 96 7 AS 15 16 9 AS 63 91 1

gle

82

Destination Count

Goo

5000

ASN

(a) Top 10 ASNs with the number of destination prefixes 1.0 CDF of ASes

existence of regional anycast deployments that primarily serve clients within the same region [4]. In addition, prior studies demonstrate that performance can be improved through trafficengineering techniques such as AS-path prepending to steer clients toward better-performing sites [8], as well as through operational configuration changes [9]. In contrast, our goal is to characterize anycast performance and configuration practices in regions such as Africa, Asia, and South America, where Internet connectivity and interconnection ecosystems are comparatively less mature. The impact of large ASes and peering relationships on latency has been widely studied. Prior work shows that anycast deployments with many upstream providers tend to exhibit more single-hop paths than deployments with only a few providers [9], and that Internet Exchange Points (IXPs) can play a critical role in improving reachability and interconnection outcomes [10]. While earlier work characterized interdomain connectivity in Africa in 2015 [11], the region’s interconnection landscape has evolved substantially since then, motivating a renewed analysis focused specifically on anycast deployments. The work most closely related to ours identifies “polarization” effects arising from incomplete connectivity of large providers and local anycast sites [5]; however, it limits its analysis largely to North America and Europe, noting that polarization is even more severe in other continents. In contrast, we address this gap by analyzing the under-studied regions and by quantifying both the severity and underlying causes of polarization.

0.5

0.0

0

2000 4000 6000 8000 10000 12000 Prefix Count

(b) CDF of ASes

Fig. 1: Impact of large ASns on number of prefixes

We perform both ping and traceroute measurements to the anycast targets. We first use ping to assess latency, and when deeper insight into the network path is needed, we run traceroutes to better understand the underlying topology. Domain Name System (DNS) logs: To determine whether the anycast hostnames are queried by clients in specific geographic regions, we analyze DNS logs to identify where these hostnames are requested from. We use B-root server logs [14], which contain DNS query names and the corresponding source IP addresses. We then geolocate the source IPs using IP-togeolocation data from Geolite [15] and ipinfo [16]. IV. H OW D O AUTONOMOUS S YSTEMS (AS ES ) A NNOUNCE P REFIXES ? We first describe how different ASes announce anycast prefixes, quantify prefixes and deployment locations per AS, and distinguish high and low performing prefixes within large anycast providers. A. Anycast prefixes by Autonomous Systems (ASes) Autonomous Systems (ASes) announce prefixes for different services like Domain Name System (DNS) or Content Delivery Networks (CDNs). Some ASes announce a large number of anycast prefixes. The anycast census data indicates that 808 ASes announce approximately 13,000 anycast prefixes. The next question is: how many prefixes does each AS announce? Figure 1 illustrates how large ASes influence the distribution of anycast prefixes. Figure 1a lists the top 10 ASes; the top three (Google, Cloudflare, and Amazon) together account for over 8,000 anycast prefixes. Finally, Figure 1b shows that fewer than 5% of ASes announce more than 10,000 anycast prefixes. After the first 11,000 prefixes, the number of ASes rises sharply. These additional ASes each announce only a small number of anycast prefixes.

CDF over prefixes

1.0 Africa Asia Europe North America Oceania South America

0.5 0.0

0

5 10 15 20 # sites in continent (per prefix)

25

D. Big ASes with multiple prefixes

CDF of prefixes

Fig. 2: CDF of sites per continent.

1.0 0.5 0.0

75% of prefixes, the median latency is below 50 ms, indicating that when sites are present in Africa, most probes experience low latency. However, the 75th-percentile curve (green line) shows that for more than 50% of prefixes, at least 25% of probes observe latencies above 50 ms. This result shows that nearby sites matter, but even when a nearby site exists, clients may still experience poor latency.

p5_rtt_ms p50_rtt_ms p75_rtt_ms p95_rtt_ms

0 50 100 150 200 250 300 350 400 450 500 RTT (ms)

Fig. 3: Best case, median, and worst case latency across anycast prefixes (using probes from Africa)

Many prefixes within the same AS are deployed from the same set of anycast sites. When these prefixes share the same locations and have similar connectivity to upstream providers and peers, probes are likely to observe similar latencies across those prefixes. We will see how the performance varies across prefixes within an AS in subsection IV-D. B. How many sites in different continents? Anycast providers deploy different numbers of sites across continents depending on their requirements and constraints. The number of sites per continent is an important factor when comparing anycast performance across regions. We examine the number of sites on each continent to understand how anycast deployments vary across regions. Figure 2 shows the distribution of number of sites across continents using the anycast census data [12]. Oceania, Africa, and South America differ substantially from the other continents in terms of site counts. Specifically, 18% of prefixes have no sites in Oceania, while 25% and 40% of prefixes have no sites in Africa and South America, respectively. C. Impacts of number of sites on latency We observe that Africa, Oceania, and South America host fewer anycast sites than the other continents. Next, we examine how the number of sites affects latency. Figure 3 shows the latency from probes in Africa to all anycast prefixes. For each destination address, we compute the best-case (5th percentile), median, and worst-case latency (75th and 95th percentile). For about 75% of prefixes, at least 5% African probes observe very low latency because they have a nearby site (blue line in Figure 3). This aligns with our earlier finding that 25% of anycast prefixes have no sites in Africa (subsection IV-B). For

We investigate how large ASes that announce many anycast prefixes configure their deployments. Do they ensure that all prefixes achieve similar latency, or do they configure different prefixes differently? This question is important because clients can be served using the prefixes that perform better than others. If clients or DNS resolvers can choose from multiple prefixes, they can select the best-performing one. We evaluate this question for the top 20 ASes by number of prefixes. All these prefixes announce more than 50 prefixes (Figure 1a shows top 10). Figure 4 shows the latency distribution across prefixes for the top 20 ASes. For each prefix, we compute the fraction of probes from Africa (due to space constraints, we present this analysis only for Africa) that observe good latency (RTT ≤ 50 ms). We report this percentage for each anycast prefix announced by that AS (Figure 4a). We use a red-to-green color scale, where red indicates a lower fraction of probes experiencing good latency and green indicates a higher fraction (Figure 4a). We also take the median of this fraction across prefixes within each AS (Figure 4b). We report the interquartile range (IQR) across prefixes to capture how this fraction varies across an AS’s prefixes. Five of the top 20 ASes have no prefix that show good latency for more than 50% probes (no green data points in Figure 4a). Afilias Inc, Nazwa.pl, OVH SAS, Windstream Communications, and Tencent Building are examples of top anycast ASes for which only a small fraction of probes observe good latency. According to the anycast census data, these five ASes have no anycast sites in Africa, leading to higher overall latency [12]. Google, Cloudflare, Amazon, Microsoft, and WoodyNet each have multiple prefixes for which roughly 80% of African probes observe good latency. According to the anycast census data [12], Google has at least four anycast sites in Africa, Cloudflare has five, Amazon has three, and WoodyNet has seven, whereas Microsoft has no anycast sites on the Africa continent according to anycast census data. Despite this, some Microsoft prefixes still provide good latency to many African probes. To investigate, we run traceroutes from a subset of African probes to these prefixes and examine the penultimate hop within Microsoft’s network. We observe that a South Africa probe reaches Microsoft in Johannesburg (e.g., ae85-0.ier01.jnb02.ntwk.msn.net), indicating local ingress into Microsoft’s backbone. In contrast, a Tunisia probe is routed to Marseille (e.g., ae102-0.icr02.mrs21.ntwk.msn.net), suggesting that some Northern African traffic is served via nearby European site while still achieving low latency due to geographic proximity.

% probes with RTT ≤ 50 ms (0=red, 100=green)

og Clo le LL ud C Am flare azo , Inc n .c . om Fas , In t c. l y Clo , In ud c. fla re Lo Afi nd lia on s, I Fly nc. .io , In Na c. zw a Go .pl og Ne le LL tAc C tua Ve te, I n rca ra, c OV LLC HS A Ak S am Ide ai nti ty Dig ita l AC E Mic ros o Go ft Da d Wo dy od Wi yNe nd t, I str nc e . Ten am C ce n t o mm Bu ild ing

Go

100 75 50 25 0

AS 39 6 AS 982 13 AS 335 16 AS 509 5 AS 411 20 3 9 AS 242 12 AS 041 40 AS 509 15 AS 967 15 AS 169 63 AS 911 12 AS 008 16 AS 276 2 AS 134 20 2 AS 726 13 6 93 AS 41 AS 807 39 5 87 87 AS AS 42 AS 702 13 9 22 03

% probes with RTT ≤ 50 ms (per prefix)

100 75 50 25 0

ft Da dd y Wo od yN Wi nd str ea m Co Ten mm ce nt Bu ild ing

AC E

OV HS

tAc tua te, Ne

Na zw

a.p l

Inc

Go

Mic ros o

Ve rca ra

Ak am a Ide i nti ty D

og le Go Fly .io

Afi lia s

98

AS 2 13 3 AS 35 16 5 AS 09 54 AS 113 20 92 AS 42 12 0 AS 41 40 5 AS 09 15 9 AS 67 15 1 AS 69 63 9 AS 11 12 0 AS 08 16 2 AS 76 21 AS 342 20 7 AS 266 13 93 4 AS 1 80 AS 75 39 87 87 AS 4 AS 2 7 AS 029 13 22 03

Fas tly

igi tal

et

do n ud fla re Lo n Clo

og le Clo ud fla re Am azo n .c om

Go

100 75 50 25 0

AS 3

96

% probes with RTT≤50ms (median for prefixes±IQR)

ASN distribution across prefixes (a) Top 20 ASes and their latency

ASN

(b) Top 20 ASes and the median percentage of probes with good latency across anycast prefixes

Fig. 4: Top 20 ASes and their latency distribution based on the vantage points from Africa

Some ASes configure only a small subset of their prefixes to provide good latency to African probes. To quantify this, we compute, for each AS, the median fraction of its anycast prefixes that achieve good latency, along with the IQR across prefixes (Figure 4b). Eight of the top 20 ASes exhibit very low median values (points in Figure 4b), including Fastly, Nazwa.pl, NetActuate, OVH SAS, Akamai, ACE, Windstream Communications, and Tencent Building. For these ASes, most announced prefixes deliver poor latency to African probes. Most ASes configure their anycast prefixes heterogeneously: within the same AS, some prefixes provide good latency while others perform poorly. This produces a clear split between green and red points in Figure 4a. A smaller set of ASes exhibit more homogeneous configurations, where nearly all prefixes show similar latency behavior. For example, Fly.io, Vercara, and Identity Digital display broadly consistent latency distributions across their prefixes. This performance gap suggests that, even within the same provider, some anycast prefixes outperform others. Therefore, when multiple prefixes offer the same service, selecting the right prefix is important for achieving better performance. E. Real clients behind prefixes In subsection IV-D, we showed that large ASes often announce multiple prefixes and that performance can vary

substantially across those prefixes. In particular, some prefixes exhibit poor performance in certain geographic regions. This raises a natural follow-up question: are clients in those regions actually being served by (or querying for) the poorly performing prefixes? To determine whether clients from a given geographic region use these prefixes, we analyze DNS data to verify whether clients in that region issue DNS queries associated with them. We first perform reverse DNS lookups to obtain hostnames for the anycast IP addresses. We then use Broot DNS data to check whether these hostnames appear in observed DNS queries. When a hostname is found, we extract the corresponding source IP addresses from the DNS logs and geolocate them. This allows us to identify the geographic regions from which the anycast hostnames were queried, and thus assess whether the relevant prefixes are being used by clients in those regions. We selected 33 hostnames across six providers and analyzed five hours of DNS data. We chose hostnames from each AS according to their performance groups: up to three hostnames from the well-performing group and three from the poorly performing group. These providers were chosen from the top 10 ASes ranked by number of announced prefixes. We did not include all top-10 providers because, for some of them, we were unable to obtain hostnames for every anycast prefix. We

then searched the DNS logs for queries to these hostnames to identify the corresponding source IP addresses and infer their geographic locations. We used hostnames from six ASes—WoodyNet, Afilias, NetActuate, Amazon, Google, and Akamai—and selected multiple prefixes from each AS representing both best-case and worst-case latencies. In the B-root DNS logs, we were able to identify hostnames from both best and worst performing groups associated with WoodyNet, Afilias, NetActuate, and Amazon. We did not observe hostnames for Google or Akamai because we analyzed only five hours of data; additionally, client traffic may have been directed to other root servers, and some clients may have been served from cache. This result shows that African clients query prefixes from both the bestperforming and worst-performing latency groups. V. H OW IS THE POLARIZATION ? Next, we present a comparative analysis of polarization and performance of the anycast networks. Polarization occurs when probes from the same geographic region observe markedly different performance—some receive good latency while others experience poor latency. For a given continent, we define a prefix as polarized if more than 5% of probes observe good latency (RTT ≤ 50 ms) while the remaining probes observe poor latency. In Africa, 75.7% of prefixes (9,831 out of 12,993) have at least 5% of probes with RTT ≤ 50 ms. For the remaining 24.3% of prefixes, most probes experience poor latency, which we suspect is due to the absence of anycast sites in Africa or they are local anycast prefixes for another continent. Moreover, among the prefixes where at least 5% of probes observe good latency, we find that all of them exhibit polarization. Probes from different continents exhibit distinct latency distributions. Figure 6 shows the latency distribution for probes from four continents. As expected, North America has substantially lower polarization than Africa, Asia, and South America. For probes in Africa, for 20% of anycast prefixes, at least 30% of African probes observe RTTs above 50 ms. For 75% of prefixes, at least 45% of African probes observe RTTs above 50 ms. Asia exhibits a better best-case scenario than Africa; however, for over 50% of prefixes, more than 40% of probes experience latencies exceeding 50 ms. South American probes show lower polarization for roughly 45% of prefixes, while for the remaining prefixes Africa performs better. North American probes perform markedly better than both Africa, Asia and South America: 60% of prefixes exhibit less than 10% polarization, which is even better than the best-case scenario (12.3% minimum polarization) observed in Africa. Figure 6 shows how polarization differs across prefixes announced by different ASes. Across all continents, we observe a wide range of polarization levels rather than a single consistent pattern. Best performing anycasts: Table I lists the ASes with the lowest polarization observed for African and South American probes. The smallest minimum polarization we find is 12.3% for AS21556 (NASA) in Africa, indicating that even for its

best-performing prefix, 12.3% of African probes still see RTTs above 50 ms despite the presence of nearby anycast sites. Other ASes with similarly low minimum polarization include WoodyNet, the University of Maryland, Cloudflare, DNScast, Microsoft, Google, Quad9, and Amazon for African probes. For these ASes, the prefixes that perform the best exhibit minimum polarization values between 12% and 18%. The best-case scenario in Asia is slightly better than Africa—still showing 11.9% to 14.0% polarization. South American probes show a better best-case outcome, with minimum polarization as low as 4.4%, achieved by Fastly. Many South American probes are routed to anycast sites in North America that are geographically relatively close, leading to lower latency. Latency variation across countries: Countries in Africa and South America experience different latency levels to anycast prefixes. For each country, we first compute the median RTT across all probes in that country for each prefix; we then report the 75th-percentile RTT across prefixes. We cannot collect measurements for countries without RIPE Atlas probes, and we include only countries with at least three probes. Many African countries exhibit median latencies around 100 ms from their probes to at least 25% of the prefixes (e.g., Egypt, South Sudan, Uganda, and Kenya). In more extreme cases, some countries experience median latencies above 150 ms to 25% of the anycast prefixes (e.g., Chad, Cameroon, and Congo). Overall, southern Africa (e.g., South Africa and Zimbabwe) tends to perform better than central regions of the continent. Figure 5a shows the latency distribution across countries in Africa. In Asia, multiple countries—including India, South Korea, Japan, and parts of Southeast Asia—exhibit good median latency (∼25 ms) to at least 75% of the prefixes (Figure 5b). In contrast, Middle Eastern countries such as Iran and Saudi Arabia show median latencies of 50–75 ms, while China exceeds 150 ms. In South America, latency distributions are generally better than in Africa. Even so, probes in Brazil, Paraguay, Argentina, and Chile typically achieve lower latencies than those in Bolivia, Peru, Colombia, and Venezuela. We can see countrywise latency map in Figure 5c. VI. R EASONS BEHIND POLARIZATION Next, we make a root-cause investigation of the factors driving the observed performance issues and polarization. A. Why Do Even the Best-Performing Anycast Prefixes Exhibit Polarization? We examine the penultimate AS hops on paths to the anycast prefixes with the lowest polarization to better understand how these networks are connected. We expect to observe different local and regional ASes on the paths that carry traffic from sources to nearby anycast sites. To test this, we analyze traceroutes from vantage points of Africa, Asia, and South America to the low-polarization anycast prefixes and extract the corresponding penultimate AS hops. We then identify which of these penultimate hops belong to Tier-1

ASN AS name AS396982 Google AS15169 Google AS16509 Amazon AS13335 Cloudflare NA AS54113 Fastly AS398787 GoDaddy AS2635 Automattic AS55002 F5 AS19905 Vercara AS29085 Porsche

Min pol (%) ASN AS name 1.1 AS54113 Fastly 1.2 AS13335 Cloudflare 1.2 AS15169 Google 1.2 AS396982 Google 1.8 SA AS209242 Cloudflare 1.8 AS398787 GoDaddy 2.3 AS55532 Squiz 2.3 AS62659 Q2 Soft 2.3 AS53332 C2Hosting 2.4 AS2559 Visa

Min pol (%) ASN AS name 4.4 AS13335 Cloudflare 4.9 AS209242 Cloudflare 6.5 AS398787 GoDaddy 6.9 7.1 AS AS23467 New Relic 7.5 AS19254 Koch 7.6 AS55532 Squiz Pty 7.9 AS4749 WiseTech 8.2 AS50756 Mikrofinanso 8.3 AS399566 Bigcommerce

Min pol (%) 11.9 12.6 12.7

ASN AS name Min pol (%) AS21556 NASA 12.3 AS1335 Cloudflare 13.6 AS42 WoodyNet 13.7 AS15169 Google 15.2 13.3 AF AS10886 U-of-Maryland 15.5 13.4 AS187 DNScast 15.9 13.5 AS8075 Microsoft 17.3 13.6 AS19281 Quad9 17.6 13.9 AS396982 Google 17.7 14.0 AS8068 Microsoft 17.9

TABLE I: Top ASes with lowest polarization in North America (NA), South America (SA), Asia (AS), and Africa (AF).

Venezuela Colombia Guyana Ecuador 175

Kazakhstan

Egypt

Libya

200

Mauritania

Niger Mali Burkina Faso Chad Senegal Benin Nigeria Guinea Côte d'Ivoire Ghana Cameroon

Sudan

Eritrea

S. Sudan

Dem. Rep. Congo

Namibia

150

Somaliland Ethiopia

Central African Rep. Kenya Gabon Congo Uganda

Angola

175

Somalia

125 100 75 50

Tanzania 25

Zambia Mozambique Zimbabwe Madagascar Botswana

Mongolia

Georgia Uzbekistan Kyrgyzstan Azerbaijan Turkmenistan Tajikistan Turkey Syria Afghanistan Iraq Iran Jordan Pakistan United Arab Emirates Saudi Arabia Oman

China

South Korea Japan

India

125

Peru

75

Bangladesh MyanmarLaos Vietnam Thailand Cambodia Philippines Malaysia

Brazil

Paraguay 50

Sri Lanka

Argentina 25

Indonesia

0

South Africa

0

(b) Asia.

50 40 30 20 10

Chile

(a) Africa.

60

100

Bolivia

Nepal

Yemen

70

150

North Korea

75th percentile RTT across prefixes (median over probes per country) [ms]

Algeria

75th percentile RTT across prefixes (median over probes per country) [ms]

Morocco

75th percentile RTT across prefixes (median over probes per country) [ms]

200

Tunisia

0

(c) South America.

CDF of prefixes

Fig. 5: Country-level median latency variation in Africa, Asia, and South America (75th percentile latency across prefixes).

1.0 0.5 0.0

Africa (AF) Asia (AS) South America (SA) North America (NA)

0

20

40

60

80

% of probes per prefix with > 50ms latency

Fig. 6: CDF of polarization level across anycast prefixes in Africa, Asia, North America, and South America

providers to assess (i) how dependent these paths are on Tier1 transit and (ii) how Tier-1 involvement affects latency. We define Tier-1 providers as the top 15 ASes in the CAIDA AS rank list (https://asrank.caida.org) as these ASes have the largest customer cones. Figure 7 summarizes the diversity of penultimate AS hops and the latency differences between Tier-1 and non–Tier-1 paths (we only show the results for Africa due to space constraints). We first count the number of distinct penultimate ASes observed for each destination and then determine how many of them are Tier-1 ASes. Figure 7a shows the number of distinct ASes appearing as penultimate hops: some destinations are reached through more than 80 distinct penultimate ASes, while the least diverse case has about 20. Most destinations fall around 60 distinct penultimate ASes, indicating substantial heterogeneity in how African networks reach these

anycast sites. Although the paths include relatively few distinct Tier-1 ASes, these ASes have a substantial impact on latency. Latency also differs markedly depending on whether paths traverse Tier-1 providers. When traffic is routed through Tier1 ASes, median latency is consistently higher, reaching over 100 ms in the worst case (orange bars in Figure 7b). In contrast, paths that avoid Tier-1 provider typically exhibit much lower median latency—often below 5 ms—suggesting delivery to a nearby (local) anycast site (blue bars in Figure 7b). Which Tier-1 ASes are more active?: We examined the most active Tier-1 ASes observed in African routes, since their presence in end-to-end paths is often associated with increased latency. Our goal is to characterize the routing paths these ASes take and investigate why they exhibit poor latency performance. AS174 (Cogent Communications) appears most frequently on paths to our destinations. Other prominent Tier-1 networks include AS3491 (PCCW Global), AS3356 (Level 3/Lumen), and AS5511 (Orange). Latency varies substantially depending on which Tier-1 AS is present along the path. Figure 8 summarizes the latency distributions, with the x-axis ordered by how prevalent each Tier1 AS is in the observed paths. The most prevalent network, AS174 (Cogent), exhibits a median latency above 100 ms when it appears on the path. AS3356 (Lumen) also shows a broad latency range where there is significant gap between the median and 75th percentile latency. In contrast, AS3491 (PCCW Global) and AS5511 (Orange) have comparatively stable latency distributions: the worst-case latency remains

Global unique Tier-1 penultimate ASNs

80 60 40 20 0

Wo od AS4 yN 2 e AS t 1 Go 516 og 9 le DN AS1 Sc 87 ast A Clo S1 ud 333 Un fl 5 ive rsi A are ty S1 of 08 Ma 86 Am A r… azo S16 n.c 50 o 9 AS m 39 Go Go 698 og o le A gle 2 As S1 ia 39 Pa 07 ci 0 AS f… 3 Na Go 9408 tio og 9 na l A AS le ero 21 na 55 ut… 6 AS Qu1928 ad 1 9 A Mic S8 ros 075 oft A Mic S8 Int ros 068 ern of et Sy A t st S3 Mik rofi ems 557 na A … ns S5 ov 0 ay 75 a 6 Ko ch A … Ind S1 us 925 trie 4 s

Count

Global unique penultimate ASNs

(a) Number of unique penultimate ASNs with Tier-1 count

Median RTT (paths with Tier-1)

100 75 50 25 0

Am A azo S16 n.c 50 om 9 A Mic S8 ros 075 oft Wo od AS4 yN 2 et A M S icr 80 Un os 68 ive rsi A oft ty S1 of 08 Ma 86 r… A D Int S ern NSca 187 s et Sy A t ste S3 Na ms 557 tio na l A AS … ero 21 na 55 ut… 6 AS Qu1928 ad 1 AS 9 C l 1 o 3 Mik rofi udfla 335 na A re ns S5 ov 0 ay 75 a 6 Ko ch A … Ind S1 us 925 tr 4 Sq A ies uiz S5 Go Pty 553 og le A Ltd 2 As S1 ia 39 Pa 07 ci 0 AS f… 3 Go 9698 og 2 AS le 39 Go 408 og 9 l AS e Go 1516 og 9 le

Median RTT (ms)

Median RTT (paths without Tier-1)

(b) Median RTT comparison when AS paths include Tier-1 and non Tier-1 ASes

Fig. 7: Number of unique ASes in the routing paths

Final RTT (ms)

200

150

100

50

Are lio nS we de AS n) 67 62 (Te lec om Ita lia ) AS 32 57 (GT TC om m)

om m) TAT AC

AS 12 99 (

Or an ge )

AS 64 53 (

Glo ba l)

AS 55 11 (

PC

CW

Lu me n) AS 34 91 (

AS 33 56 (

AS 17 4(

Co ge nt)

0

Tier-1 ASN (Org) — sorted by prevalence in the path

Fig. 8: Latency distribution for Tier-1 ASes when the Tier-1 ASes are in the routing path

within 25 ms for AS3491 and within 50 ms for AS5511. These large ASes sometimes steer traffic to a distant anycast site because they lack connectivity to the nearest anycast location. For instance, when traffic traverses AS174 (Cogent), it may be routed from Africa to a European anycast site, since Cogent is connected to the European site and not to a local African site. This is one of the reasons behind polarization in

anycast networks. We find anycast networks sometimes use multiple big ASes to have complete coverage and redundancy as shown in Figure 7a. However, having multiple ASes in the paths may have adverse impacts on latency if all the big transits are not connected locally to an anycast site. While it is hard to predict how many big ASes should be there as the transit providers, it is essential to have complete connectivity with the transit providers. If the transit is only connected to a few anycast locations, then there is a high chance of having polarization. Characteristics in Asia and South America: We also assess latency variation in Asia and South America by comparing paths that include a Tier-1 AS with those that do not. We observe similar trends: when Tier-1 ASes appear in the path, packets often take a longer route. In the best-case scenario, paths traversing a Tier-1 AS mostly show a median latency of over 50 ms for Asia and 30 ms for South America, whereas paths without a Tier-1 AS have a median latency of under 10 ms in both Asia and South America. AS4637 (Telstra Global), AS9002 (RETN Limited), AS3491 (PCCW Global), AS3356 (Lumen), and AS1299 (Arelion) are the most common Tier-1 ASes in Asia. Among these, AS4637 (Telstra Global) shows a wide range of latency ranging from 5 ms to 170 ms. We suspect this AS is locally connected to some anycast prefixes but reaches other prefixes

Paths through Tier-1 (%)

X-2: ASes and names with >50% polarization

N

100 75 50 25 0

9 tw 117 N e AS1ryaka A c. In

s, ork

3 758 ger nal AS4ostinnatio H ter In

Lower polarization

6 155 ics AS2nautAd… o Aer ace nal Sp atio and

o Wo

2 AS4Inc. t, e N dy

1 ien 061 Med AS5nipp K

7 G 805 b i t A AS5ecure S

9 c 185 er In AS2enlay Z

1 ited 756 Lim AS1ARUS L

Median latency with Tier-1 in path (ms)

1 . 204 Inc AS1filias, A

6 088 o f ASe1rsityland v ry Uni Ma

C

87 AS1Scast DN

5 333 c. ASa1re, In dfl l ou

075 ft ASc8rosoion i M porat C or

9 516 LC AS1gle L o Go

1 928 d 9 AS1 Qua

82 969 LLC AS3ogle Go

068 ft ASc8rosoion i M porat C or

X-1: Lower polarization (specified AS order)

ks, 9 twor 11k7a Ne 1 S A rya A c. In

3 758 ger nal AS4ostinnatio H te r In

7 G 805 bit A AS5ecure S

1 ie n 061 Med AS5nipp K

9 c 185 er In AS2enlay Z

1 i te d 756 Lim AS1ARUS L

ns c pee ctri uro… E 2 Ele 515aux IPCoor 939cane 2 6 S S A ese ork A urri R e tw H LC N L

6 nc 525 pe I AS5etsko N

1 . 204 Inc AS1filias, A

Lower polarization >50% polarization

200 100

6 155 ics AS2nautAd… o Aer ace nal Sp atio and

6 nc 525 p e I AS5etsko N

>50% polarization

(a) Percentage of Tier-1 AS in the paths X-2: ASes and names with >50% polarization

N

ns c pee ctri uro… E 2 El e 5 9 P r I 1 3 e o 5 x o 9 n AS2eseauork C AS6urrica R etw H LC N L

2 AS4Inc. , t e N ody Wo

6 088 of ASe1rsityland v ry Uni Ma

C

5 333 c. ASa1re, In dfl lo u

87 AS1Scast N D

075 ft ASc8rosoion i M porat Co r

9 516 LC AS1gle L Go o

1 928 d9 AS1 Qua

82 969 LLC AS3ogle Go

068 ft ASc8rosoion i M porat Co r

X-1: Lower polarization (specified AS order)

(b) Median latency variation

Fig. 9: Comparison of Tier-1 existence and latency variation between lower and > 50% polarization

via more distant locations. We recommend anycast providers should establish local connectivity with this provider, as many Asian clients rely on this AS. We also find that AS3356 (Lumen), AS174 (Cogent), AS1299 (Arelion), and AS6762 (Telecom Italia) are the most frequently observed ASes in paths from South American probes. Other routing inefficiencies: To examine how routing inefficiencies can cause poor latency from local ASes, we analyze traceroute data for the best-performing prefixes. We observe that certain local ASes incur substantial delays before reaching the destination. For example, when AS2609 (Tunisia Backbone) appears as the penultimate AS hop, the median latency is 175 ms. To understand this behavior, we examined the routing paths traversing AS2609 and found routes to e.rootservers.net (AS21556 – NASA). Although this root DNS server has instances in Africa and the corresponding prefix is among the best-performing in the region, paths that traverse AS2609 exhibit significantly higher latency. In these traces, we see multiple routing hops within Tunisia, and the geolocation data suggests an inflated and inefficient routing path. We also found similar behavior for AS16637 (MTN South Africa), when they carry Cloudflare traffic. We also observe cases where local ASes carry traffic to distant regions and effectively behave like transit providers. For instance, AS12956 (Telxius Cable) routes DNSCast traffic from Chad to Europe, resulting in latencies exceeding 100 ms.

Network operators should investigate such routing inefficiencies to identify the cause of elevated delays. In many cases, providers can mitigate the problem by discontinuing prefix announcements through upstream networks that consistently route traffic inefficiently. In Asia, we observed inefficient paths when traffic traverses AS4134 (Chinanet Backbone) to reach a Cloudflare prefix. When analyzing the best-performing prefixes from South America, we observe similar behavior in a few local ASes. B. How Do Poorly Performing Anycast Prefixes Differ from the Best-Performing Prefixes? Next, we examine configuration differences under high polarization. Our goal is to understand why certain prefixes exhibit higher polarization and why elevated latency persists across a larger fraction of probes. We select 10 prefixes with polarization greater than 50%— that is, more than half of their probes from Africa observe latency above 50 ms. We then run traceroutes from the same set of probes to these anycast prefixes. For each traceroute, we determine whether the path to the destination includes a Tier-1 AS and identify the penultimate AS hop connected to the anycast site. Finally, we compare the latency distributions of paths that traverse Tier-1 ASes with those that use only non–Tier-1 ASes. We compare these results with the best-case scenario, which has the lowest polarization (Figure 9). Our objective is to

identify differences in anycast configuration that contribute to higher polarization. Figure 9a shows the percentage of paths that traverse a Tier-1 AS. In most instances where polarization exceeds 50%, we observe a higher fraction of Tier-1–transiting paths (the red line lies above the blue line). In some cases, the majority of paths use a Tier-1 AS when polarization is above 50%. One example is when the anycast deployment belongs to a Tier-1 network itself (AS6939). In other cases, such as AS55256 and AS12041, the anycast deployment appears to rely heavily on Tier-1 ASes. These anycast prefixes depend more on Tier-1 ASes instead of maintaining direct peering relationships with local ASes. Figure 9b compares latency in the lowest-polarization case versus the > 50% polarization case for paths that traverse Tier1 ASes. We observe higher polarization when performance over Tier-1 paths is significantly worse (the red line sits well above the blue line). In some cases, the median latency exceeds 200 ms when Tier-1 ASes appear on the path. This suggests these Tier-1 ASes are not well connected to local or nearby anycast sites, forcing traffic onto longer routes to reach the destination and leading to higher latency. We also compare the median latency in the lowestpolarization case and the high-polarization case for paths that do not traverse a Tier-1 AS. Even without Tier-1 ASes, the higher-polarization cases still exhibit higher median latency than the best-case scenario. However, when Tier-1 ASes are absent, median latency remains much lower (typically < 50 ms) than in the Tier-1–transiting case. Overall, polarization appears to be driven primarily by a greater reliance on Tier-1 transit and by Tier-1 connectivity that steers traffic to distant anycast sites. When comparing to the best case scenario, we observe many of the same Tier-1 ASes on paths when polarization exceeds 50%, but their performance is markedly worse in these cases. For instance, AS3491 (PCCW) exhibits a median latency of about 100 ms, whereas in the best-case scenario it was below 10 ms. This highlights the importance of strong local connectivity with large transit providers. In some poorly performing anycast prefixes, different large ASes exist in the paths; however, it is generally preferable to have broad connectivity to local anycast sites rather than relying on several Tier-1 ASes that are connected to only a limited subset of sites. VII. C ONCLUSION In this paper, we evaluate anycast networks to identify polarization in regions such as Africa, Asia, or South America, where Internet connectivity and interconnection ecosystems are comparatively less mature. We analyze more than 10,000 anycast prefixes by probing them from RIPE Atlas vantage points located in Africa, Asia, and South America. Using these measurements, we quantify the extent of polarization in each region. We illustrate how large providers deploy their networks and prefixes, highlighting cases with both low and high latency. We find that polarization is substantially more severe in Africa, Asia, and South America than in North America. To understand the underlying causes, we complement latency

measurements with traceroutes to infer anycast configurations and routing behavior, and we show that large ASes frequently steer traffic to distant anycast sites, contributing directly to polarization. We also highlight routing inefficiencies where traffic from a local AS takes a longer path to the destination. Overall, our results suggest that improving the connectivity between large providers and local anycast sites is the key to mitigating polarization. R EFERENCES [1] Z. Li, D. Levin, N. Spring, and B. Bhattacharjee, “Internet anycast: performance, problems, & potential,” in Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, 2018, pp. 59–73. [2] M. Caesar and J. Rexford, “Bgp routing policies in isp networks,” IEEE network, vol. 19, no. 6, pp. 5–11, 2005. [3] T. Koch, E. Katz-Bassett, J. Heidemann, M. Calder, C. Ardi, and K. Li, “Anycast in context: A tale of two systems,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference, 2021, pp. 398–417. [4] M. Zhou, X. Zhang, S. Hao, X. Yang, J. Zheng, G. Chen, and W. Dou, “Regional ip anycast: Deployments, performance, and potentials,” in Proceedings of the ACM SIGCOMM 2023 Conference, 2023, pp. 917– 931. [5] A. Rizvi, T. Huang, R. Esrefoglu, and J. Heidemann, “Anycast polarization in the wild,” in International Conference on Passive and Active Network Measurement. Springer, 2024, pp. 104–131. [6] M. Calder, A. Flavel, E. Katz-Bassett, R. Mahajan, and J. Padhye, “Analyzing the performance of an anycast cdn,” in Proceedings of the 2015 Internet Measurement Conference, 2015, pp. 531–537. [7] R. de Oliveira Schmidt, J. Heidemann, and J. H. Kuipers, “Anycast latency: How many sites are enough?” in International Conference on Passive and Active Network Measurement. Springer, 2017, pp. 188– 200. [8] M. Zhou, Y. Chen, J. Zheng, Y. Xu, P. Hu, Y. Tang, W. Yin, J. Lin, Q. Yu, Y. Su et al., “{AnyPro}:{Preference-Preserving} anycast optimization based on strategic {AS-Path} prepending,” in 23rd USENIX Symposium on Networked Systems Design and Implementation (NSDI 26), 2026, pp. 2587–2602. [9] S. McQuistin, S. P. Uppu, and M. Flores, “Taming anycast in the wild internet,” in Proceedings of the Internet Measurement Conference, 2019, pp. 165–178. [10] L. Prehn, F. Lichtblau, C. Dietzel, and A. Feldmann, “Peering only? analyzing the reachability benefits of joining large ixps today,” in International Conference on Passive and Active Network Measurement. Springer, 2022, pp. 338–366. [11] R. Fanou, P. Francois, and E. Aben, “On the diversity of interdomain routing in africa,” in International Conference on Passive and Active Network Measurement. Springer, 2015, pp. 41–54. [12] R. Hendriks, M. Luckie, M. Jonker, R. Sommese, and R. van RijswijkDeij, “Laces: An open, fast, responsible and efficient longitudinal anycast census system,” in Proceedings of the 2025 ACM Internet Measurement Conference, ser. IMC ’25. New York, NY, USA: Association for Computing Machinery, 2025, p. 445–461. [Online]. Available: https://doi.org/10.1145/3730567.3764484 [13] RIPE NCC, “RIPE Atlas: Active Internet Measurements,” https://www. ripe.net/analyse/internet-measurements/ripe-atlas/, accessed: 2026-0825. [14] USC/B-Root Operations and USC/LANDER Project, “Day in the life of the internet (ditl) april 2026 dataset: Broot message/question logs,” IMPACT Data Repository, Apr. 2026, iMPACT ID: USC-LANDER/DITL B Root message question20260407/rev15744. [Online]. Available: http://www.isi.edu/ant/lander [15] MaxMind, Inc., “GeoLite2 City Database,” https://dev.maxmind.com/ geoip/geolite2-free-geolocation-data, 2025, accessed: 2026-08-25. [16] IPinfo, “IPinfo: IP Address Data and Geolocation,” https://ipinfo.io, 2016, accessed: 2026-08-25.

Related documents

• Multi-network broadcasting within the Internet#208903
• Transport Layer Performance Combining Multipath and Network Coding in Mobile Satellite Networks#1215312
• The Heterogeneous Effects of Data Transfer Restrictions on Data Centre Geography#680813
• Unveiling TCP BBR Dominance in Starlink Internet: Experimental Insights and Analysis#349594
• Data driven approach for Outdoor Channel Prediction in 5G and Beyond#1109051