B owsEm: Model-based Web Si e Loading Emula ion
Kilian Holzinge
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
Flo ian Klein
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
Daniel Pe i
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
S e an Lachni
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
Sebas ian Gallenmülle
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
Geo g Ca le
[email p o ec ed]
Technical Uni e si y o Munich
Ge many
Abs ac
Con inuous esea ch and enginee ing e o s aim o imp o e
he pe o mance o web ne wo king p o ocols, such as TCP,
QUIC, TLS, and HTTP. Pe o mance measu emen s a e con-
duc ed o assess he impac o changes o hese p o ocols,
hei implemen a ions, o he unde lying ne wo k in as uc-
u e. Howe e , his is a challenging ask due o he inc easing
complexi y o web si e deploymen s and b owse and se e
so wa e. In his wo k, we in oduce B owsEm, a web si e
loading emula o capable o ep oducing page loading wo k-
loads. I s unde lying model is based on da a sc aped om
eal web si es and akes ne wo k pa h cha ac e is ics, p o-
ocols, dependencies be ween HTTP ansac ions, and hei
indi idual iming aspec s in o accoun . Modula ly s uc u ed
and using wide-sp ead lib a ies like libcu l, i allows o es -
ing a wide ange o emula ion pa ame e s such as ne wo k
condi ions o p o ocol implemen a ions and ob aining e-
p oducible esul s. Ou e alua ion shows ha he emula ion
in oduces a ela i e e o smalle han
±
0.25 o 80% o
obse ed page load imes. In a measu emen campaign, we
ind ha he model shows sui able obus ness o a i icially
changed pa ame e s.
CCS Concep s
•Ne wo ks
→
Ne wo k pe o mance modeling;Ne wo k
expe imen a ion;Ne wo k pe o mance analysis;Ne -
wo k measu emen ;Applica ion laye p o ocols;T ans-
po p o ocols.
Keywo ds
web, b owse s, emula ion, measu emen , p o ocols
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 In e na-
ional License.
ANRW ’25, Mad id, Spain
©2025 Copy igh held by he owne /au ho (s).
ACM ISBN 979-8-4007-2009-3/2025/07
h ps://doi.o g/10.1145/3744200.3744759
ACM Re e ence Fo ma :
Kilian Holzinge , Flo ian Klein, Daniel Pe i, S e an Lachni , Sebas-
ian Gallenmülle , and Geo g Ca le. 2025. B owsEm: Model-based
Web Si e Loading Emula ion. In Applied Ne wo king Resea ch Wo k-
shop (ANRW ’25), July 22, 2025, Mad id, Spain. ACM, New Yo k, NY,
USA, 7 pages. h ps://doi.o g/10.1145/3744200.3744759
1 In oduc ion
Web pe o mance is impo an o he use expe ience and
he success o web applica ions. The demand o imp o e-
men s has led o he de elopmen o new p o ocols such
as QUIC and HTTP/3 o ex ensions o he web s ack such
as ansac ion mul iplexing, mul ipa h anspo p o ocols,
p io i iza ion schemes, o pos -quan um c yp og aphy. Engi-
nee ing and esea ch e o s con inuously assess he impac
o new de elopmen s, iden i y limi a ions, and p o ide p oo -
o -concep implemen a ions. Howe e , accu a ely e alua -
ing he in luence o changes o he ne wo k s ack on web
applica ions emains challenging. Mode n b owse s a e com-
plex so wa e sys ems, making i ha d o add unc ionali y
as p oo o concep s o ne wo k subsys ems such as he
anspo o HTTP p o ocol implemen a ion. Addi ionally,
ep oducibili y o in-b owse measu emen s is a challeng-
ing endea ou . In con as o ha , ully syn he ic e alua ion
app oaches ail o e lec ealis ic wo kload pa e ns.
The e is a me hodological gap o enable expe imen a ion
wi h he web ne wo k s ack while s ill being able o im-
pose ealis ic and ep oducible applica ion wo kloads. An
app oach is needed o emula e he b owse beha io while
gi ing esea che s he lexibili y o expe imen wi h unde -
lying anspo and applica ion p o ocol implemen a ions.
In his wo k, we in oduce B owsEm, an emula o o
web applica ion wo kloads ha gi es ne wo k esea che s a
lexible and adjus able ool o assess he impac o changes
unde ealis ic condi ions while a oiding in ac able com-
plexi y. Fo he emula ion, we de i e a model o he page
loading p ocess, cap u ing me a-da a o HTTP ansac ions.
The app oach educes he complexi y o clien and se e
web applica ions in ha he ac ual con en is no in e p e ed
ANRW ’25, July 22, 2025, Mad id, Spain Kilian Holzinge , Flo ian Klein, Daniel Pe i, S e an Lachni , Sebas ian Gallenmülle , and Geo g Ca le
o ende ed, bu eplaced wi h iming and dependency in-
o ma ion. We iden i y model pa ame e s by ex ac ing and
s o ing in o ma ion om he websi e loading beha io o
web b owse s and he ne wo k s ack. This sc aping p ocess
is ully au oma ed. The esul ing ile is used by he emula o
which eplica es he loading beha io on he anspo and
HTTP laye wi hou he need o se up complex in as uc-
u e o o modi y li e websi es.
The esul ing emula ion helps in unde s anding he impac
o changed ne wo k pa h cha ac e is ics and modi ica ions
o he anspo o applica ion laye on he loading ime o
websi es. Use s can ei he modi y model pa ame e s, such as
changing he HTTP e sion o eques s, o expe imen wi h
clien o se e side p o ocol ea u es, e. g., added new HTTP
p io i y signals o mul ipa h connec ions. The emula o is
designed in a modula way such ha di e en implemen a-
ions on clien and se e side a e possible. The e alua ion
shows ha B owsEm is capable o eplica ing ealis ic web
wo kloads wi h ema kable accu acy, achie ing a ela i e
page load ime (PLT) e o o
±0.25
o 80% o sampled si es.
The emainde o he pape is s uc u ed as ollows: The
nex sec ion p o ides an o e iew o ela ed wo k. Sec ion 3
de ines he scope o he sys em model o B owsEm. The
implemen a ion is explained in Sec ion 4. In Sec ion 5, we
e alua e and discuss esul s o pe o med measu emen s.
Finally, a e explaining he limi a ions o he app oach in
Sec ion 6, a conclusion summa izes he main indings.
The code o he ool is a ailable online1.
2 Rela ed Wo k
The demand o he ealis ic and ep oducible applica ion
o web a ic models da es back o he ea ly e sions o
HTTP. Ba o d and C o ella [
2
] s a e ha ei he analy i-
cal o ace-based app oaches can be used o gene a e syn-
he ic web wo kloads. Thei Scalable URL Re e ence Gene a o
(
SURGE
) ool was employed o in es iga e web pe o mance
in In e ne -wide measu emen s, whose a ic adhe ed o
ma hema ical models mimicking HTTP eques p ope ies
o use s in e ac ing wi h a web page [1].
SURGE
’s analy ical model has simila i ies wi h Mah’s [
10
]
and Choi and Limb’s [
7
]. I accoun s o s a is ically ep-
esen a i e ile and eques sizes, hei ela i e popula i y
o o he se e esou ces, e- eques likelihood, how hey
a e e e enced in a page, e e ence coun , and a ic bu s i-
ness. These alues we e de i ed om da a se s [
4
] which
empi ically cap u ed web use beha io .
SURGE
’s syn he ic
a ic ma ched he a iabili y o eal a ic and pu a mo e
ealis ic load on se e s han o he benchma king ools a
he ime. Pe Ba o d and C o ella’s de ini ion, B owsEm is
ace-based as i emula es ealis ic wo kloads h ough he
1h ps://gi hub.com/holzingk/B owsEm
playback o p e- eco ded samples. Acco ding o hem, such
an app oach has he endency o become a black box ha
obscu es sys em beha io and is ha d o adap o new e-
qui emen s. We belie e o ha e add essed such sho comings
h ough he use o a lexible and con igu able model o ma .
No ing signi ican web de elopmen s since he la e 1990s,
such as he ise o ideo s eaming applica ions, Lee and
Gup a [
9
] p oposed a a ic gene a o ecognizing ha use s
may change pages be o e embedded objec s a e ully down-
loaded o be using mul iple b owse s a once. P ies e al. [
15
]
compa ed Lee’s, Chois’, Ba o d’s, and Ihm’s [
8
] measu e-
men s wi h a a ic model based on he op million accessed
web si es and iden i ied he e ie al o la ge mul imedia
con en sp ead ac oss he globe as a end. Thei HTTP
a ic model was expe imen ally in eg a ed in o he
ns-3
QUIC module [
14
]. Web page load imes o e HTTP/1.1,
SPDY, and QUIC we e e alua ed wi h he Mahimahi ame-
wo k [
12
], which can accu a ely eco d and eplay HTTP
a ic locally using ully isola ed, composable UNIX shells
ha emula e mul i-se e applica ions and ne wo k condi-
ions. The Mahimahi-suppo ed inding ha he se ializa ion
o eques s unde u ilizes links led o imp o ing HTTP appli-
ca ion pe o mance o e long-delay links.
3 Scope o Emula ion Model
The scope o he emula ion model is shaped by he needs
o p o ocol esea che s and web si e ope a o s in es iga ing
he con inuously e ol ing web s ack.
Mode n web b owse s a e sophis ica ed so wa e sys ems.
Thei complexi y is d i en by web s anda ds and he need o
in e p e , execu e, ende , and display mul imedia con en .
When loading a websi e, he eques ed documen o en in-
cludes o he asse s, such as images, Ja aSc ip , o s yleshee s.
I a e e enced asse is hos ed on a di e en web se e , his
can lead o in o ma ion a eling on a di e en ne wo k pa h
wi h di e en p ope ies and addi ional o e head due o he
es ablishmen o a new connec ion.
In he ne wo king- ocused B owsEm model, all in e p e-
a ion o con en is abs ac ed a he clien and eplaced wi h
in e - ansac ion dependencies and iming in o ma ion. The
iming o indi idual HTTP ansac ions a e u he di ided
in o se e al phases, namely blocked,dns,connec , ls,send,
wai , and ecei e. The pa h be ween clien and se e is mod-
eled wi h capaci y, delay, and loss p ope ies. Se e -side
web applica ions also in oduce applica ion limi ed phases.
We model hose by delaying HTTP esponses by a speci ic
ime. Figu e 1 depic s a simpli ied d awing o he model.
B owsEm: Model-based Web Si e Loading Emula ion ANRW ’25, July 22, 2025, Mad id, Spain
Clien
◀▶
◀▶
◀▶
Se e 1
◀▶
Se e 2
◀▶
. . .
Se e n
◀▶
Figu e 1: Simpli ied emula ion model
4 App oach
The B owsEm wo k low consis s o ou main pa s. To build
ealis ic emula ion models, we i s de i e model pa ame-
e s by obse ing he page load p ocess o a web b owse ,
e e ed o as da a sc aping. Then, he collec ed da a a e
pos p ocessed and con e ed in o he wo kload da a model
ile. Since B owsEm use s ypically wan o assess he impac
o modi ica ions in he web s ack, manual edi s can be made
he e. Mo eo e , al e na e e sions o componen s such as
web se e s o QUIC lib a ies can be deployed inside he em-
ula o . The wo kload da a model is subsequen ly p o ided as
inpu o he page load emula ion, which eplica es he HTTP
and anspo laye wo kload. Finally, use s can inspec he
esul ing pe o mance me ics.
4.1 Da a Sc aping
In o de o emula e page load beha io , sample da a a e
needed. To be as ealis ic as possible, we chose o obse e
and eco d he beha io o a web b owse while loading
web si es and use he sc aped da a as inpu o he emula o .
Fo now, ou ool is limi ed o he Ch omium b owse bu
since he used in e ace is also o e ed by o he b owse s, we
an icipa e ha hey can be suppo ed as well. To au oma e
he eco ding o he page loading p ocess, B owsEm p o-
ides a Py hon ool. I ex ac s in o ma ion om he b owse
and om a ic cap u es. I uses he Ch ome De Tools P o-
ocol (CDP)
2
and PySha k
3
, a Py hon in e ace o in e ac
wi h he Wi esha k ne wo k packe analyze . HTTP A chi e
(HAR) iles, ansac ion dependencies, and a ic sni ing a e
a ailable da a sou ces. qlog, a s uc u ed logging o ma o
QUIC, would be an addi ional op ion [
11
]. Howe e , i is no
suppo ed by Ch omium.
HAR. All majo b owse s ha e buil -in de elope ools
ha eco d ne wo k ac i i y and page load imes. A de- ac o
s anda d o hose da a a e HAR iles which include all e-
ques s he b owse made while loading a page [
13
]. I is a
s uc u ed JSON log, con aining de ailed sub imings o each
esou ce: The blocked ime o a eques is he delay caused
by he b owse , e. g., due o a eached connec ion limi . Also,
2h ps://ch omede ools.gi hub.io/de ools-p o ocol/
3h p://kiminew .gi hub.io/pysha k/
he du a ion o a po en ial DNS lookup is encoded (dns),
along wi h he ime o connec ion es ablishmen (connec ),
du a ion o he T anspo Laye Secu i y (TLS) handshake
( ls) and ime o send he eques o he se e (send). In ad-
di ion, he wai ime indica es he ime om he comple ed
eques o he i s by e o he esponse, i. e., he ime o
i s by e (TTFB). Finally, he ecei e ime measu es he ime
o comple e he eques . We use he CDP o ini ia e page
loading wi h HAR ile eco ding and w i e a TLS key log ile,
allowing o dec yp eco ded packe aces.
T ansac ion dependencies. HAR iles include ini ia o ields.
They e e ence an o he ansac ion ha igge ed his pa -
icula one, which we use o in e a ansac ion dependency
ee. Al hough his in o ma ion some imes is missing, an
ini ia o can clea ly be iden i ied h ough manual inspec ion.
To add ess his p oblem, we c ea ed a dependency ee in-
e ence heu is ic. We assume ha child eques s ypically
s a sho ly a e he pa en began ecei ing i s esponse.
We iden i y he pa en ansac ion by inding he eques
wi h he smalles posi i e del a be ween i s esponse s a
ime and he child’s eques s a ime bounded by a speci ic
h eshold. Al hough by na u e his app oach is no ee o
e o s, e alua ion yielded su icien accu acy.
T a ic sni ing. Because no all ele an da a a e included
in he HAR ile, some da a a e ex ac ed om eco ded a ic
aces. PySha k dec yp s TLS and QUIC packe s wi h he
key log ile, enabling us o ma ch ele an lows o HTTP
eques s and o p o ide addi ional insigh s in o HTTP usage.
We assess c ucial pa h cha ac e is ics pe se e IP ad ess. We
use se e al mechanisms o assess he pa h ound- ip ime
(RTT). In he case o TCP, we use he median RTT ob ained
by ma ching he imes amp alue and imes amp echo eply,
i a ailable. As allbacks, in he ollowing o de , handshake
imes amps o ICMP echo eques and eply as es ima e a e
used espec i ely. The QUIC spin bi is an addi ional RTT
es ima ion sou ce ha is no ye used by he Ch omium
b owse . Ano he op ion would be o ack QUIC packe and
TCP segmen numbe s and co ela ing acknowledgemen s.
Fo he pa h capaci y, we in eg a e he passi e es ima ion
app oach PP a e [
5
] o which an open-sou ce implemen a-
ion has been p o ided [
3
]. Some connec ions only exchange
a limi ed numbe o packe s, likely ne e lea ing he slow
s a phase o conges ion con ol. As a consequence, some
esul s a e e oneous.
4.2 Wo kload Da a Model
The da a model desc ibes he emula ed wo kload and is se-
ialized as JSON. I combines inpu om he HAR ile and
analyzed sni ed a ic eco ded du ing web sc aping. The
ANRW ’25, July 22, 2025, Mad id, Spain Kilian Holzinge , Flo ian Klein, Daniel Pe i, S e an Lachni , Sebas ian Gallenmülle , and Geo g Ca le
main con en is a lis o HTTP ansac ions. Connec ion iden-
i ie s dis inguish connec ions in case o mul iplexed p o-
ocols such as HTTP/2 and HTTP/3. Absolu e imes amps
indica e he s a and end imes o he eques and esponse.
Rela i e eques imings aken om he HAR ile p o ide
addi ional in o ma ion. We append he eques and esponse
body leng h and he depending and dependan ansac ion
iden i ie s ob ained om he ini ia o ield o dependency
heu is ic. The da a model con ains ne wo k pa h in o ma ion
in he o m o loss pe cen age, RTT and es ima ed capaci y.
The s uc u ed ile allows o easy modi ica ion o he
wo kload model. Use s, o ins ance, can change he HTTP
e sion o indi idual eques s, adjus RTT o bandwid h con-
s ain s, a y he numbe o pa allel connec ions, o add
addi ional eques s. This makes i possible o simula e al e -
na i e e sions o a web si e’s beha io unde di e en web
s ack con igu a ions, while s ill p ese ing he ealism o he
o iginal wo kload.
4.3 Page Loading Emula ion
The page loading emula ion uses he Tokio
4
asynch onous
Rus un ime. B owsEm includes ne wo k pa h beha io in
i s da a model. I uses Linux ne wo k namespaces in con-
junc ion wi h he Ne Em ne wo k emula o o imi a e he
beha io o dis inc pa hs be ween he clien and he se e s
on a single node [6].
Fo each sepa a e se e in he model, a dedica ed Tokio
ask is spawned. The se e is de ined as an abs ac be-
ha iou using a Rus ai , so new o modi ied se e s can
be added by use s ha p o ide implemen a ions o ha in-
e ace. Cu en ly, he e a e h ee example implemen a ions.
We implemen ed he ai , o one, using Ac ix Web
5
which
is a Rus web amewo k ha suppo s HTTP/1.0, HTTP/1.1,
and HTTP/2. Fo ano he , using quinn as an HTTP/3 se e .
Las , we suppo using nginx as an example o a gene al
pu pose web se e . The ai de ines wo asynch onous
unc ions. The i s is a eques callback which akes he
wo kload model as inpu o ini ialize he se e , he second
is a callback ha handles he ac ual emula ed eques s.
The clien uses libcu l which is a widely used and lexible
HTTP clien lib a y. I was chosen because i p o ides a
single API o se e al HTTP e sions and also suppo s i e
QUIC implemen a ions: Cloud la e quiche, ng cp2, OpenSSL
3.2 QUIC, msquic and he expe imen al Linux Ke nel QUIC.
Fo ou e alua ion we use ng cp2. As an ini ializing s ep,
be o e he ac ual wo kload emula ion is pe o med, he clien
sends he wo kload model o he se e s. A he clien , each
connec ion is handled by a Tokio ask which i sel spawns
asks o he indi idual ansac ions which a e s a ed in
4h ps:// okio. s/
5h ps://ac ix. s/
0 20 40 60 80
100
101
102
se e coun
eques coun
(a) Reques and se e
coun pe loaded web si e
102104106
0
0.2
0.4
0.6
0.8
1
esponse size [B]
CDF
(b) HTTP esponse body
sizes o all web si es
Figu e 2: S a is ics o sc aped si es
acco dance wi h he iming in o ma ion ound in he load
model da a. The wai ime in he model da a includes he
ne wo k pa h RTT and addi ional delays caused by he se e
applica ion. In he emula ion, we sub ac he RTT om he
wai ime, since we apply ne wo k pa h emula ion, allowing
o change ne wo k pa h cha ac e is ics. Be o e a eques is
made in he emula o , he dependency ee is espec ed by
conside ing he delay om he inaliza ion o he pa en
esponse o he eques o he child. Depending on he model
da a, ei he a new connec ion o he se e is ini ia ed, o an
exis ing connec ion is eused. Simila o mos b owse s, he
TCP nodelay op ion is se .
5 E alua ion
We s a his sec ion wi h a da a d i en mo i a ion, ad o-
ca ing o he need o a ool such as B owsEm. The p ima y
goal o ou e alua ion is o quan i y he e o i in oduces
while ep oducing he loading beha io o web si es. We
compa e emula ion esul s o measu emen s om eco ded
b owse sessions. In addi ion, we assess he model s abili y
by modi ying emula ion pa ame e s such as he RTT and
in es iga e whe he esul s a e consis en wi h eal-wo ld
page loads unde simila condi ions.
5.1 Mo i a ion
To pe o m he e alua ion, we use he op 1000 web si es
om he DomCop Lis o Top 10 Million Domains
6
o sc ap-
ing. Some web si es caused issues, e. g., due o he use o
WebSocke s which we do no ye suppo o because head-
less use agen s a e blocked. The ollowing da a a e based on
922 sampled si es. We only eco d he loading p ocess o he
s a page and do no pe o m any in e ac ions wi h he si e.
We use a as connec ion wi hin a na ional science ne wo k.
In Figu e 2a, we p o ide an o e iew o he numbe o
se e s in ol ed and eques s coun o each page load. I can
be seen ha loading mos web si es es ablishes connec ions
o se e al di e en se e s and make ens o hund eds o
6h ps://www.domcop.com/ op-10-million-domains
B owsEm: Model-based Web Si e Loading Emula ion ANRW ’25, July 22, 2025, Mad id, Spain
0 200 400 600
0
0.2
0.4
0.6
0.8
1
ime [ms]
CDF
blocked
HTTP/1.0 (#18) HTTP/1.1 (#2259) HTTP/2 (#37077) HTTP/3 (#10156)
0 50 100 150 200
ime [ms]
dns
0 200 400
ime [ms]
connec
0 50 100 150 200
ime [ms]
ls
0 20 40 60
ime [ms]
send
0 100 200
ime [ms]
wai
0 20 40 60 80 100
ime [ms]
ecei e
(a) Dis ibu ion o ansac ion phases measu ed in b owse om sc aped da a
−0.1−0.05 0 0.05 0.1
0
0.2
0.4
0.6
0.8
1
ϵblocked
CDF
−0.1−0.05 0 0.05 0.1
ϵdns
−10 −5 0
ϵconnec
0 0.2 0.4 0.6 0.8 1
ϵ ls
−4−2 0
ϵsend
−40 −20 0
ϵwai
−4−2 0
ϵ ecei e
(b) Rela i e di e ences o ansac ion phases be ween measu emen s o eal web si es and emula ion
Figu e 3: Du a ion o ansac ion phases as CDFs o all eques s sepa a ed by HTTP e sion; di e en x-axis scales
eques s. Figu e 2b shows he dis ibu ion o esponse body
sizes ac oss all 49510 ansac ions in he sc aped da a se .
The as majo i y a e ans e s o smalle han 100 kB.
Figu e 3a depic s cumula i e dis ibu ion unc ions (CDFs)
o he du a ion o all phases o each HTTP e sion. They in-
clude all eques s ex ac ed om HAR iles gene a ed while
loading he 922 web si es. The da ase includes 18 HTTP/1.0,
2259 HTTP/1.1, 37077 HTTP/2, and 10156 HTTP/3 ansac-
ions. I shows ha majo con ibu o s o he ansac ion
du a ion a e he blocked,wai , and ecei e imes.
F om hese s a is ics, we conclude ha web si es ha e a
e y mixed loading beha io , leading o a huge a ie y o
wo kloads imposed on he HTTP and anspo laye . To
ep oduce hose ou side o a web b owse , while accoun ing
o he iming o ansac ion phases, a ool such as B owsEm
is needed — a pu ely syn he ic app oach is no su icien .
5.2 Me ics
To quan i y he emula ion quali y o B owsEm we in oduce
wo me ics. Fo he i s , we de ine he ela i e di e ence
be ween he sc aped page load ime
𝑠𝑃𝐿𝑇
and emula ed page
load ime
𝑒𝑃𝐿𝑇
as he PLT e o
𝜖𝑃𝐿𝑇 =
𝑠𝑃𝐿𝑇 −𝑒𝑃 𝐿𝑇
𝑠𝑃𝐿𝑇
. I cap-
u es how well he o e all loading beha io and eques
dependency s uc u e a e ep oduced. The second me ic,
𝜖𝑃=
𝑠𝑃−𝑒𝑃
𝑠𝑃
, measu es he ela i e di e ence o indi idual
phases o HTTP ansac ions. I can help o iden i y he oo
cause o emula ion de ia ions.
5.3 Emula ion E o
Figu e 3b shows ela i e di e ences o ansac ion phase
du a ions
𝜖𝑃
o e all obse ed ansac ions. Fo e e ence,
see also Figu e 3a, which shows he absolu e dis ibu ion o
sc aped ansac ion phase imes. The sample coun s can be
ound in he legend o Figu e 3. Due o a limi ed numbe o
HTTP/1.0 ansac ions, ela ed esul s need o be in e p e ed
wi h ca e. The blocked ime is always me by he ool, since
as he name sugges s, no ope a ions a e pe o med. Simila ly,
B owsEm eplaces DNS lookups wi h wai ime. The shown
connec imes la gely ma ches he b owse beha io . The e
is a small ac ion o HTTP/3 ansac ions wi h la ge ela i e
di e ences. Since HTTP/3 ansac ions ha e a e y sho
absolu e connec ime, we deem hose ou lie s as insigni i-
can . We model TLS and enc yp ion o ou se e s as well.
HTTP e sions g ea e han HTTP/1.0 euse connec ions,
esul ing in a la ge numbe o ansac ion wi h ls ime o
0
ms
. The emula ion ends o comple e he TLS handshake
as e . Possible easons o he de ia ions could be he use o
di e en enc yp ion lib a ies a he clien and se e com-
pa ed o he eal-wo ld measu emen , o in luences o an
e oneous RTT es ima e. The absolu e du a ions o he send
ime a e ypically e y sho . In he emula ion i ends o
be e en sho e han in he b owse . The e, i could be a -
ec ed by addi ional applica ion o conges ion limi a ions no
co e ed in ou model, hin ing owa ds w ong pa h capaci y
es ima ions. The wai ime is mos ly me by B owsEm. I is
ealized in he emula ed se e by wai ing o he measu ed
ime, whe eby he pa h RTT is deduced (see Sec ion 4.3) — a
possible cause o di e ences. The ecei e ime, albei being
gene ally longe due o la ge ans e sizes, shows simila
beha io as he send ime and he same explana ions apply. In
conclusion, while de ia ions om he obse ed b owse im-
ings exis , he emula ion ep esen s indi idual ansac ion
phases easonably well.
Nex , we assess he ela i e di e ence o he PLT, shown
in Figu e 4a. The PLT me ic
𝜖𝑃𝐿𝑇
indica es whe he he
ANRW ’25, July 22, 2025, Mad id, Spain Kilian Holzinge , Flo ian Klein, Daniel Pe i, S e an Lachni , Sebas ian Gallenmülle , and Geo g Ca le
−1−0.5 0 0.5 1
0
0.2
0.4
0.6
0.8
1
ϵP LT
CDF
(a) Dis ibu ion o ela-
i e di e ences be ween
sc aped and emula ed o al
page load imes
0 50 100 150 200
0
2
4
6
addi ional RTT delay [ms]
page load ime [s]
measu ed
emula ed
(b) Absolu e di e ences
o PLT be ween measu e-
men s o eal web si e and
emula ed scena io
Figu e 4: Page load ime and s udy o model obus ness
loading beha io o a web page, pa icula ly he dependency
ee o eques s and hei imings is accu a ely emula ed. We
ind ha abou 80% page loads ha e less han ±0.25𝜖𝑃 𝐿𝑇 .
5.4 S abili y
Using he model ile, use s can modi y emula ed eques pa-
ame e s such as pa h p ope ies o he HTTP e sion. To
e alua e he eliabili y o esul s, we assess he model s abil-
i y by compa ing hose emula ions agains measu emen s
aken di ec ly in a b owse unde he same condi ions.
Fo ha pu pose, we ake he model de i ed om sc ap-
ing a web si e unde baseline condi ions, i. e., wi hou in o-
ducing any addi ional delay. To s udy he e ec o ne wo k
la ency, we un he emula ion epea edly, wi h manually
inc eased pa h RTT p ope ies, keeping all o he aspec s
cons an . Fo each con igu a ion, we compa e he emula ed
esul s agains g ound- u h measu emen s ob ained om
ac ual b owse sessions unde he same added delay. This
allows us o e alua e he obus ness o ou emula ion model.
The esul s o his expe imen a e shown in Figu e 4b.
We can obse e ha he o e all PLT is s eadily inc easing
wi h he added RTT, as expec ed. The emula ed PLT closely
ollows he end o he eal b owse measu emen s, indi-
ca ing ha ou model is capable o accu a ely cap u ing he
impac o ne wo k la ency on web page loading imes. In
e ms o he ela i e e o
𝜖𝑃𝐿𝑇
, we can obse e an inc ease
wi h added RTT.
6 Limi a ions
Ou model is inhe en ly limi ed by which da a is measu eable
in he b owse and he ne wo k s ack. Thus, we only implic-
i ly cap u e b owse -speci ic beha io s such as clien -side
ende ing, Ja aSc ip execu ion, o CSS ende ing h ough
he iming o eques s and esponses. The e o e, some web
Quali y o Expe ience (QoE) me ics like Time o Fi s Pain
a e no cap u ed, bu could be explo ed in u u e wo k. The
use o s a ic ne wo k condi ions limi s ealism, as eal-wo ld
condi ions a y signi ican ly. Also, we do no di ec ly co e
applica ion limi ed phases in nei he se e no clien . The
ne wo k pa h es ima ion showed o be a sou ce o e o s.
While eques dependency modeling pe o ms well in many
cases, i may no ully cap u e he beha io o all web applica-
ions. Addi ionally, pa s o he model ely on da a epo ed
by Ch omium and may no gene alize o o he b owse s
such as Sa a i o Fi e ox, which use a di e en ne wo king
s ack. Fu u e wo k could enhance he model by adding mo e
complexi y, such as including p io i y signals and e alua ing
he impac on he accu acy o he emula ion.
7 Conclusion
To imp o e he web ne wo k s ack, p o ocol esea che s,
ne wo k enginee s, and si e ope a o s a e in e es ed in mea-
su ing pe o mance o web applica ions while ha ing ull
lexibili y o unde lying implemen a ion de ails, ne wo k pa-
ame e s, and con igu a ions. To achie e ha , eplica ing
deploymen s in ep oducible es bed se ups is edious be-
cause he deploymen , modi ica ion, and o ches a ion o
in ol ed b owse and se e so wa e is complex. B owsEm
add esses his gap by p o iding a lexible and ex ensible
measu emen amewo k, capable o eplica ing ne wo k
wo kloads o web applica ions wi h a ocus on ne wo k pa h
p ope ies, anspo and HTTP applica ion laye .
The ool includes an au oma ed da a sc ape o ex ac he
wo kload model om b owse beha io . We use HAR iles
and a ic aces as da a sou ces. Ga he ed in o ma ion is
collec ed in o a model ile. I con ains de ailed in o ma ion
on he o de and dependency o HTTP ansac ions and he
iming o hei indi idual phases as well as in e ed ne wo k
pa h p ope ies such as delay and pa h capaci y.
The emula o eplica es he page loading p ocess by eplac-
ing p ocessing ime spen wi h gene a ing o in e p e ing
web payload con en wi h wai ime. I s i es o mee he du-
a ion o all ele an ansac ion phases. Based on he model
in o ma ion, ne wo k pa hs a e emula ed on a single node
using Linux ne wo k namespaces. The clien uses libcu l as
a lexible and ex ensible HTTP lib a y, which has mul iple
backends o HTTP/3. We p o ide h ee implemen a ions
o he se e . All s anda d HTTP e sions a e suppo ed.
We e alua e B owsEm on a sample o 922 popula web
si es. We compa e he ime di e ence o indi idual ansac-
ion phases be ween b owse measu emen s and he emula-
ion. Resul s show ha , despi e some de ia ions, he emula-
ion app oxima es eal beha io su icien ly well. The o al
PLT has less han
±0.25
ela i e e o in 80% o emula ed
page loads. In an addi ional measu emen campaign, he em-
ula o was assessed o obus ness and yielded good s abili y
owa ds changed ne wo k pa h pa ame e s.
The code o he ool is a ailable online1.
B owsEm: Model-based Web Si e Loading Emula ion ANRW ’25, July 22, 2025, Mad id, Spain
Acknowledgmen s
This wo k was suppo ed by he EU Ho izon Eu ope p o-
g amme, p ojec s SLICES-PP (10107977) and G eenDIGIT
(101131207), by he Ge man Fede al Minis y o Educa ion
and Resea ch (BMBF), p ojec s 6G-li e (16KISK002) and 6G-
ANNA (16KISK107), by he Ge man Resea ch Founda ion,
p ojec Hype NIC (CA595/13-1), and by he Ba a ian Min-
is y o Economic A ai s, Regional De elopmen and Ene gy,
p ojec 6G Fu u e Lab Ba a ia.
Re e ences
[1]
Paul Ba o d and Ma k C o ella. 1999. Measu ing Web pe o mance
in he wide a ea. 27, 2 (1999), 37–48. doi:10.1145/332944.332953
[2]
Paul Ba o d and Ma k C o ella. 1998. Gene a ing ep esen a i e
Web wo kloads o ne wo k and se e pe o mance e alua ion. In
P oceedings o he 1998 ACM SIGMETRICS Join In e na ional Con-
e ence on Measu emen and Modeling o Compu e Sys ems (Madi-
son, Wisconsin, USA) (SIGMETRICS ’98/PERFORMANCE ’98). Asso-
cia ion o Compu ing Machine y, New Yo k, NY, USA, 151–160.
doi:10.1145/277851.277897
[3]
Simon Baue , Janluka Janelidze, Benedik Jaege , Pa ick Sa le , Pa ick
B zoza, and Geo g Ca le. 2023. On he Accu acy o Ac i e Capaci y
Es ima ion in he In e ne . In 2023 IEEE/IFIP Ne wo k Ope a ions and
Managemen Symposium (NOMS 2023). Miami, USA.
[4]
Ca los R Cunha, Aze Bes a os, and Ma k E C o ella. 1995. Cha ac e -
is ics o WWW clien -based aces. Technical Repo . Bos on Uni e si y
Compu e Science Depa men .
[5]
Taou ik En-Najja y and Guillaume U oy-Kelle . 2006. Pp a e: A pas-
si e capaci y es ima ion ool. In 2006 4 h IEEE/IFIP Wo kshop on End-
o-End Moni o ing Techniques and Se ices. IEEE, 82–89.
[6]
S ephen Hemminge e al
.
2005. Ne wo k emula ion wi h Ne Em. In
Linux Con AU, Vol. 5. 2005.
[7]
Hyoung-Kee Choi and J.O. Limb. 1999. A beha io al model o Web
a ic. In P oceedings. Se en h In e na ional Con e ence on Ne wo k
P o ocols (To on o, On ., Canada, 1999). IEEE Compu . Soc, 327–334.
doi:10.1109/ICNP.1999.801961
[8]
Sunghwan Ihm and Vi ek S Pai. 2006. Towa ds unde s anding mode n
web a ic. (2006).
[9]
Jeongeun Lee, Ma u i Gup a, and In ellon Co p. 2007. A NEW TRAFFIC
MODEL FOR CURRENT USER WEB BROWSING BEHAVIOR. h ps:
//api.seman icschola .o g/Co pusID:14057598
[10]
B uce A. Mah. 1997. An empi ical model o HTTP ne wo k a ic.
In P oceedings o INFOCOM ’97, Vol. 2. 592–600 ol.2. doi:10.1109/
INFCOM.1997.644510
[11]
Robin Ma x, Luca Niccolini, Ma en Seemann, and Lucas Pa due. 2025.
QUIC e en de ini ions o qlog. In e ne -D a d a -ie -quic-qlog-quic-
e en s-10. In e ne Enginee ing Task Fo ce. h ps://da a acke .ie .
o g/doc/d a -ie -quic-qlog-quic-e en s/10/ Wo k in P og ess.
[12]
Ra i Ne a ali, Ani udh Si a aman, Somak Das, Ameesh Goyal, Kei h
Wins ein, and Ha i Balak ishnan. 2015. Mahimahi: Accu a e Reco d-
and-Replay o HTTP. In 2015 USENIX Annual Technical Con e ence
(USENIX ATC 15).
[13]
Jan Od a ko, A ind Jain, and Andy Da ies. 2012. HTTP A chi e
(HAR) Fo ma . h ps://w3c.gi hub.io/web-pe o mance/specs/HAR/
O e iew.h ml
[14]
Umbe o Pa o, Fede ico Chia io i, Anay Aji Deshpande, Michele
Polese, And ea Zanella, and Michele Zo zi. 2020. Ex ending he ns-3
QUIC Module. In P oceedings o he 23 d In e na ional ACM Con e -
ence on Modeling, Analysis and Simula ion o Wi eless and Mobile Sys-
ems (Alican e Spain, 2020-11-16). ACM, 19–26. doi:10.1145/3416010.
3423224
[15]
Ras in P ies, Zsol Magya i, and Phuoc T an-Gia. 2012. An HTTP
web a ic model based on he op one million isi ed web pages. In
P oceedings o he 8 h Eu o-NF Con e ence on Nex Gene a ion In e ne
NGI 2012 (Ka lsk ona, Sweden, 2012-06). IEEE, 133–139. doi:10.1109/
NGI.2012.6252145
