Research Paper HTTP 3.0
Research Paper HTTP 3.0
Measuring HTTP/3:
Adoption and Performance
Martino Trevisan† , Danilo Giordano† , Idilio Drago ‡ , Ali Safari Khatouni?
†
Politecnico di Torino, ‡ University of Turin, ? Shopify
first.last@{polito.it, unito.it, shopify.com}
Abstract—The third version of the Hypertext Transfer Protocol and Facebook2 ). However, currently neither the real state of
(HTTP) is in its final standardization phase by the IETF. Besides its deployment nor the performance benefits of HTTP/3 have
arXiv:2102.12358v2 [cs.NI] 10 Nov 2021
better security and increased flexibility, it promises benefits in been measured yet.
terms of performance. HTTP/3 adopts a more efficient header
compression schema and replaces TCP with QUIC, a transport In this paper we fill this gap by running the first large-scale
protocol carried over UDP, originally proposed by Google and measurement study on HTTP/3 adoption and performance. We
currently under standardization too. Although HTTP/3 early first rely on the HTTPArchive Dataset to study to what extent
implementations already exist and some websites announce its the web ecosystem has already adopted HTTP/3. Then, we
support, it has been subject to few studies. We provide a first run additional campaigns to measure the benefits introduced
measurement study on HTTP/3 adoption and performance. We
testify how it has been adopted by some of the leading Internet by HTTP/3. Considering websites that adopt different versions
companies such as Google, Facebook and Cloudflare in 2020. of the HTTP protocol, we measure several metrics known
We run a large-scale measurement campaign towards thousands to indicate users’ Quality of Experience (QoE). Finally, we
of websites adopting HTTP/3, aiming at understanding to what emulate different network conditions on the paths connecting
extent it achieves better performance than HTTP/2. We find that our measurement platform to assess whether and how HTTP/3
adopting websites often host most web page objects on third-
party servers, which support only HTTP/2 or even HTTP/1.1. improves performance in different scenarios.
As excepted, websites loading objects from a limited set of third- Using the open-source HTTPArchive Dataset,3 we find thou-
party domains (avoiding legacy protocols) are those experiencing sands of websites supporting HTTP/3, most of them hosted by
larger performance gains. Our experiments however show that a handful of Internet hyper-giants, i.e., Facebook, Google, and
HTTP/3 provides sizable benefits only in scenarios with high
Cloudflare. We then automatically revisit websites supporting
latency or poor bandwidth.
HTTP/3 under diverse network conditions to measure the
Index Terms—HTTP/3; Performance; Measurements. performance benefits in terms of QoE-related metrics. We
visit 14 707 websites in total while emulating artificial latency,
I. I NTRODUCTION packet loss, and limiting the bandwidth. We run 2 647 260
visits over a period of one month. We find that HTTP/3
The Hypertext Transfer Protocol (HTTP) is the king of benefits emerge only on particular network conditions and
web protocols and is used to access the vast majority of strongly differ across websites. Our key findings are:
services on the Internet, from websites to social networks
• Google, Facebook and Cloudflare are the early adopters
and collaborative platforms. HTTP was born in the early 90s,
of HTTP/3, hosting almost the totality of currently web-
and its first version 1.1 was standardized in 1997 [1]. Only
sites supporting HTTP/3.
in 2014, HTTP/2 [2] was proposed, with substantial changes
• The majority of web page objects in websites support-
in the framing mechanisms. HTTP/3 is the third version of
ing HTTP/3 are still hosted on non-HTTP/3 third-party
HTTP and is currently in the final standardization phase at
servers.
the IETF [3]. It promises performance benefits and security
• We observe sizable performance benefits only in scenar-
improvements compared to HTTP/2. As a major change,
ios with high latency or low network bandwidth.
HTTP/3 replaces TCP as transport layer in favor of QUIC,
• The performance gains largely depends on the infrastruc-
a UDP-based protocol originally proposed by Google and
ture hosting the website, possibly due to optimizations on
currently being standardized by the IETF [4]. Furthermore,
server-side infrastructure.
it introduces a more effective header compression mechanism
• As expected, the websites relying on fewer connections
and exploits TLS 1.3 [5] (or higher) to improve the level of
to load objects are those benefiting the most.
security.
HTTP/3 is expected to take over the place of HTTP/2 in the The remainder of the paper is organized as follows: Sec-
next years, and some of the leading Internet companies already tion II describes HTTP/3 and illustrates related work. Sec-
announced its support during 2020 (e.g., the CloudFlare CDN1 tion III presents our datasets and how we have collected them.
Section IV illustrates our results on HTTP/3 adoption and
This work has been supported by the EU H2020 research and innovation
programme under grant agreement No. 644399 (MONROE) and by the 2 https://engineering.fb.com/2020/10/21/networking-traffic/
SmartData@PoliTO center on Big Data and Data Science. how-facebook-is-bringing-quic-to-billions/
1 https://blog.cloudflare.com/http3-the-past-present-and-future/ 3 https://httparchive.org/
2
Websites [%]
3
Fundamental for our analyses, the details of HTTP re-
sponses indicate the eventual Alt-Svc header, which is used
2
by servers to announce support to HTTP/3. By setting the
Alt-Svc header, the server has the possibility to inform the 1
client to make subsequent connections using HTTP/3 and may
specify the support to specific draft versions (e.g., 27 or 29). 0
We download the HTTPArchive dataset starting from 0 1 2 1 2 3 4 5 6 7 8 9 0 1 2
9/1 9/1 9/1 0/0 0/0 0/0 0/0 0/0 0/0 0/0 0/0 0/0 0/1 0/1 0/1
November 2019, when we observe the first websites offering 201 201 201 202 202 202 202 202 202 202 202 202 202 202 202
support to HTTP/3. We use the data to study the trend of
HTTP/3 adoption. The data sum up to 6.6 TB. Since we are Fig. 2: Percentage of websites in HTTPArchive that announce
interested in studying the adoption of HTTP/3 on websites, we support to HTTP/3, separately by IETF draft.
discard all visits to internal pages (less than half of the total)
and keep only visits to home pages. We refer to this dataset
as HTTPArchive. protocol. All visits to the same website are run consecutively,
cleaning all state between repetitions, i.e., browser cache, TCP
connections etc. Visits are repeated 5 times to get more reliable
B. HTTP/3 Performance – BrowserTime
results. Hence, we visit each website 4×3×3×5 = 180 times.
We use the most recent snapshot at the time of writ- BrowserTime collects several statistics for each visit, includ-
ing (December 2020) to build the list of websites currently ing details on all HTTP transactions as well as performance
supporting HTTP/3. We find 14 707 websites announcing metrics. We track two metrics that have been shown to be
support to HTTP/3. Next, we visit these websites with three correlated with users’ QoE [17] and can be estimated also at
HTTP versions (HTTP/1.1, HTTP/2, and HTTP/3) to quantify the ISPs [18]:
possible performance improvements. To this end, we rely on • onLoad: The time at which the browser fires the onLoad
BrowserTime, a dockerized tool to run automatic visits to web event – i.e., when all elements of the page, including
pages with a large set of configurable parameters.8 We use images, style sheets and scripts have been downloaded
BrowserTime to instrument Google Chrome to visit websites and parsed;
using a specific HTTP version. Important for our goal, Google 9
• SpeedIndex: Proposed by Google, it represents the time
Chrome allows specifying a set of domains to be contacted at which visible portions of the page are displayed.
with HTTP/3 on the first visit, i.e., without prior indication It is computed by capturing the video of the browser
via Alt-Svc header. We limit ourselves to Chrome, since screen and tracking the visual progress of the page during
we are not aware of similar functionalities in other browsers rendering.
(e.g., Firefox).
In total, we run 2 647 260 visits over a period of one month.
We are interested in studying the impact of HTTP/3 under
The visit metadata account for 189 GB, and we call this dataset
different network conditions. As such, we run our measure-
BrowserTime.
ments enforcing different network configurations. We run our
experiments using two high-end servers connected to the IV. HTTP/3 ADOPTION AND PERFORMANCE
Internet via 1 Gbit/s Ethernet and located in our university
campus. We call this baseline scenario Native, as reported in In this section we first provide an overview of the HTTP/3
Table II. adoption (Section IV-A). Since announcing HTTP/3 support is
Then, we enforce other configurations during the visits not the same as serving content using the protocol, we quantify
relying on the well-known Linux tc tool. Each network the amount of content served over HTTP/3 (Section IV-B).
configuration is defined by changing one of three parameters: Then, we study how HTTP/3 affects QoE-related performance
(i) extra latency, (ii) extra packet loss, or (iii) bandwidth limit. metrics (Section IV-C) and whether identified improvements
For each parameter, we use 4 different configurations, reported can be related to the provider hosting content (Section IV-D)
in Table II. In case of latency, we impose it on the uplink, or website characteristics (Section IV-E).
while packet loss and bandwidth limit are enforced on both up
and down links. For each network configuration, we visit each A. HTTP/3 adoption
website (i) enabling only HTTP/1.1, (ii) enabling HTTP/1.1 We first study to what extent HTTP/3 has been adopted
and HTTP/2, and (iii) enabling all three versions of the since its first proposal. To this end, we profit from the
8 https://www.sitespeed.io/documentation/browsertime/ 9 https://web.dev/speed-index/
4
100 12
HTTP/1.1
10
75
8 HTTP/3
50 6
4
25 Objects
Volume 2
0 0
Cloudflare Facebook Google Other
(12 455) (445) (1 094) (731)
6
Fig. 5: Share of objects/volume served on HTTP/3, separately 5
by provider.
Fig. 7: H3 Delta on different scenarios. onLoad (top) and SpeedIndex (bottom). Negative values indicate that HTTP/3 is faster.
Looking at the solid red lines, we notice that approximately in some other cases. In fact, in several tested cases, some
in 50 % of the cases websites load faster with HTTP/3 and websites can even perform worse when HTTP/3 is enabled.
in the remaining cases HTTP/3 is slower. When latency is
high, HTTP/3 gives sizable benefits compared to HTTP/2. If D. HTTP/3 performance by provider
we impose extra latency of 50 ms, 69 (74) % of websites
have lower onLoad time (SpeedIndex), meaning that they load Next we study whether HTTP/3 performance gains could be
faster. The number of websites loading faster increase to 76 related to the provider hosting the websites. As we observed
(81) % with 100 ms latency. With 200 ms, the number of sizable performance benefit for HTTP/3 only in cases of high
websites loading faster reach 81 (87) %, and the median H3 latency or poor bandwidth, we restrict our analyses to those
Delta is -0.08 (0.12). cases.
Figure 8 shows the distribution of H3 Delta for onLoad,
Focusing on experiments with bandwidth limitation (central separately by provider. We focus on scenarios with 200 ms
plots in Figure 7), different considerations hold. We observe extra latency and 1 Mbit/s bandwidth limit. We observe that
sizable benefits only for onLoad time with the bandwith the H3 Delta largely differs by provider. Focusing on latency
limited to 1 Mbit/s, where 69 % of websites load faster (Figure 8a), Facebook websites show the highest performance
with HTTP/3. Notice that this benefit cannot be introduced gain (H3 Delta −0.13 in median), represented in the figure
by indirect higher latency due to queuing delay (also called with the blue dashed line. Moreover, 95% of websites are
bufferbloat), as we limit the machine queues to 32 KB. In other loaded faster with HTTP/3 than with HTTP/2. Cloudflare (red
cases, no clear trend emerges, but we notice a larger variability solid line) shows the smallest benefits, with only 72% of
of the H3 Delta measure introduced by the constrained setup. websites loading faster. Google and the remaining websites
For example, in case of SpeedIndex, 56, 49, 58 % websites sit in the middle. Similar considerations hold for SpeedIndex,
load faster with HTTP/3 with 5, 2 and 1 Mbit/s bandwidth, not shown here for brevity.
respectively. Similar considerations hold for packet loss (right- With limited bandwidth (Figure 8b), we observe a very
most plots in Figure 7). Despite a larger variability, we cannot different situation. Here, Facebook has in general worst per-
identify any general trend, and the H3 Delta values are equally formance with HTTP/3 with 91% of its websites loading
distributed above and below 0. faster with HTTP/2. Conversely, Google (green dashed line)
In summary, we observe improvements on onLoad time with shows the best figures, with median H3 Delta −0.14 and 79%
poor bandwidth when using HTTP/3. HTTP/3 shows sizable of websites loading faster with HTTP/3. Cloudflare and the
benefits in case of high latency. We do not testify performance remaining websites exhibit no clear trend with roughly half of
benefits of HTTP/3 in scenarios with high packet loss and the websites loading faster with HTTP/3.
7
1.0 6
H3 Faster
5
Normalized Value
0.8 H3 ≈ H2
4 H2 Faster
0.6
CDF
Cloudflare 3
0.4 Facebook 2
Google
0.2 1
Other
0.0 0
n. main
s n. re ze
−0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0.2 0.3 0.4 0.5 f con t con 3 sha Page si
H3 Delta (onLoad) u m ber o ber of Do on larges HTTP/
N Num Objects
(a) Latency.
(a) Latency.
1.0
Cloudflare 6
0.8 Facebook H3 Faster
5
Normalized Value
H3 ≈ H2
0.6 Google
4 H2 Faster
CDF
Other
0.4 3
2
0.2
1
0.0
−0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0.2 0.3 0.4 0.5 0
n. s n. re e
H3 Delta (onLoad) o f con of Domain rgest con TP/3 sha Page siz
ber a HT
Num Number jects on l
(b) Bandwidth. Ob
(Figure 9b), rather than with high latency (Figure 9a). scale measurement campaign, we studied the performance of
Serving most objects with HTTP/3 (rather than with HTTP/3 under different network conditions, targeting thou-
HTTP/2) has a positive impact too, as we notice from the sands of websites. We found performance benefits emerging
fourth box group in Figure 9. Again, this is evident especially in scenarios with high latency or poor bandwidth. In the case
with bandwidth limit (Figure 9b), while with extra latency of high packet loss, HTTP/3 and HTTP/2 perform roughly
(Figure 9a) it is hard to find a clear trend. Finally, interesting the same. We found large performance diversity depending on
is the case of the web page size (last box group). In scenarios the infrastructure hosting the website. In general, we observed
with high latency, the websites benefiting from HTTP/3 are that websites taking benefits from HTTP/3 are those loading
small, while large ones typically perform better with HTTP/2. objects from a limited set of third-party domains, thus limiting
Conversely, when bandwidth is scarce, even if moderately, the number of issued connections and avoiding loading content
website loading faster with HTTP/3 are the large ones. using legacy protocols.
In summary, websites taking benefits from HTTP/3 are those
limiting the number of connections and third-party domains,
and fully adopting HTTP/3 on all web page objects. Page size R EFERENCES
has diverse implications depending on the network conditions.
These considerations hold in scenarios with high latency or [1] R. T. Fielding, H. Nielsen, J. Mogul, J. Gettys, and T. Berners-Lee,
limited bandwidth, while we do not observe any clear trend in “Hypertext Transfer Protocol – HTTP/1.1.” RFC 2068, Jan. 1997.
case of optimal network conditions or high packet loss, where [2] M. Belshe, R. Peon, and M. Thomson, “Hypertext Transfer Protocol
Version 2 (HTTP/2).” RFC 7540, May 2015.
metric distributions mostly overlap.
[3] M. Bishop, “Hypertext Transfer Protocol Version 3 (HTTP/3),” Internet-
Draft draft-ietf-quic-http-33, Internet Engineering Task Force, Feb. 2021.
V. D ISCUSSION AND FUTURE WORK Work in Progress.
[4] J. Iyengar and M. Thomson, “QUIC: A UDP-Based Multiplexed and
We dissected the performance of HTTP/3 under diverse net- Secure Transport,” Internet-Draft draft-ietf-quic-transport-33, Internet
work conditions, showing the impact of network latency and Engineering Task Force, Dec. 2020. Work in Progress.
bandwidth across websites. However, we run measurements [5] E. Rescorla, “The Transport Layer Security (TLS) Protocol Version 1.3.”
RFC 8446, Aug. 2018.
using only Google Chrome, as we are not aware of other [6] M. Varvello, K. Schomp, D. Naylor, J. Blackburn, A. Finamore, and
browsers that can be instrumented to use HTTP/3 since the first K. Papagiannaki, “Is the web http/2 yet?,” in International Conference
connection – i.e., without the need to observe the Alt-Svc on Passive and Active Network Measurement, pp. 218–232, Springer,
2016.
header previously. Moreover, we always visit websites with [7] M. Rajiullah, A. Lutu, A. S. Khatouni, M.-R. Fida, M. Mellia, A. Brun-
a fresh browser profile with empty cache and no pre-existing strom, O. Alay, S. Alfredsson, and V. Mancuso, “Web experience in
connections. Clearly, this setup limits the scope of our study mobile networks: Lessons from two million page visits,” in The World
Wide Web Conference, pp. 1532–1543, 2019.
as we cannot measure how HTTP/3 affects performance on
[8] D. Saif, C.-H. Lung, and A. Matrawy, “An early benchmark of quality of
subsequent visits or with a warm HTTP cache, as it will be experience between http/2 and http/3 using lighthouse,” arXiv preprint
the case for future users supporting HTTP/3. arXiv:2004.01978, 2020.
We limited ourselves to a subset of the websites adopting [9] R. Marx, J. Herbots, W. Lamotte, and P. Quax, “Same standards,
different decisions: A study of quic and http/3 implementation diversity,”
HTTP/3. Indeed, we included only a fraction of websites in Proceedings of the Workshop on the Evolution, Performance, and
hosted on the CloudFlare CDN, as its HTTP/3 support is Interoperability of QUIC, pp. 14–20, 2020.
partially disabled at the time of writing. Our measurements [10] S. Tellakula, “Comparing HTTP/3 vs. HTTP/2 Performance.” https://
blog.cloudflare.com/http-3-vs-http-2/, 2020.
will need to be run continuously to observe how the web
[11] L. Guillen, S. Izumi, T. Abe, and T. Suganuma, “Sand/3: Sdn-assisted
ecosystem will react to HTTP/3 in the near future, adopting novel qoe control method for dynamic adaptive streaming over http/3,”
the protocol and modifying the web page structure to optimize Electronics, vol. 8, no. 8, p. 864, 2019.
performance, in particular concerning domain sharding. [12] K. Wolsing, J. Rüth, K. Wehrle, and O. Hohlfeld, “A performance
perspective on web optimized protocol stacks: Tcp+ tls+ http/2 vs. quic,”
Finally, HTTP/3 and QUIC are not yet definitive IETF in Proceedings of the Applied Networking Research Workshop, pp. 1–7,
standards. Although recent modifications to the IETF draft 2019.
only concerned minor protocol features, it will be fundamental [13] J. Manzoor, L. Cerdà-Alabern, R. Sadre, and I. Drago, “On the perfor-
mance of quic over wireless mesh networks,” Journal of Network and
to provide a similar analysis once the final standards are Systems Management, vol. 28, no. 4, pp. 1872–1901, 2020.
approved. Similarly, we did not explore how different endpoint [14] G. Carlucci, L. De Cicco, and S. Mascolo, “Http over udp: An experi-
configurations affect HTTP/3 performance – e.g., the interac- mental investigation of quic,” in Proceedings of the 30th Annual ACM
Symposium on Applied Computing, SAC ’15, (New York, NY, USA),
tions of HTTP/3 and congestion control settings. Moreover, p. 609–614, Association for Computing Machinery, 2015.
whereas we covered different scenarios, it is widely known [15] A. M. Kakhki, S. Jero, D. Choffnes, C. Nita-Rotaru, and A. Mislove,
that network emulation is hard. Similar measurement studies “Taking a long look at quic: an approach for rigorous evaluation of
with actual end-users are still needed to confirm our findings. rapidly evolving transport protocols,” in Proceedings of the 2017 Internet
Measurement Conference, pp. 290–303, 2017.
2020, although most of the early adopters still host the ma- [16] M. Kosek, T. Shreedhar, and V. Bajpai, “Beyond quic v1–a first look
jority of objects on HTTP/2 third-party servers. With a large- at recent transport layer ietf standardization efforts,” arXiv preprint
arXiv:2102.07527, 2021.
VI. C ONCLUSIONS [17] D. N. da Hora, A. S. Asrese, V. Christophides, R. Teixeira, and D. Rossi,
We provided the first study on HTTP/3 adoption and per- “Narrowing the gap between qos metrics and web qoe using above-
the-fold metrics,” in International Conference on Passive and Active
formance, quantifying the performance benefits of HTTP/3 Network Measurement, pp. 31–43, Springer, 2018.
in several network scenarios. We testified how some of the [18] M. Trevisan, I. Drago, and M. Mellia, “Pain: A passive web performance
Internet leading companies started deploying HTTP/3 during indicator for isps,” Computer Networks, vol. 149, pp. 115–126, 2019.