Burst Computing: Quick, Sudden, Massively Parallel Processing on Serverless Resources

This pape is included in he P oceedings o he
2025 USENIX Annual Technical Con e ence.
July 7–9, 2025 • Bos on, MA, USA
ISBN 978-1-939133-48-9
Open access o he P oceedings o he
2025 USENIX Annual Technical Con e ence
is sponso ed by
Bu s Compu ing: Quick, Sudden, Massi ely Pa allel
P ocessing on Se e less Resou ces
Daniel Ba celona-Pons, Uni e si a Ro i a i Vi gili and Ba celona Supe compu ing
Cen e ; Ai o A jona, Ped o Ga cía-López, En ique Molina-Giménez, and
S epan Klymonchuk, Uni e si a Ro i a i Vi gili
h ps://www.usenix.o g/con e ence/a c25/p esen a ion/ba celona-pons
Bu s Compu ing: Quick, Sudden, Massi ely Pa allel P ocessing
on Se e less Resou ces
Daniel Ba celona-Pons†‡ Ai o A jona†Ped o Ga cía-López†
En ique Molina-Giménez†S epan Klymonchuk†
†Uni e si a Ro i a i Vi gili ‡Ba celona Supe compu ing Cen e
Abs ac
We p esen bu s compu ing, a no el se e less solu ion
ailo ed o bu s -pa allel jobs. Unlike Func ion-as-a-Se ice
(FaaS), bu s compu ing es ablishes job-le el isola ion using
a no el g oup in oca ion p imi i e o launch la ge g oups
o wo ke s wi h gua an eed simul anei y. Resou ce alloca-
ion is op imized by packing wo ke s in o ewe con aine s,
which accele a es hei ini ializa ion and enables locali y. Lo-
cali y signi ican ly educes emo e communica ion compa ed
o FaaS and, combined wi h simul anei y, i allows wo ke s
o communica e synch onously wi h message passing and
g oup collec i es. Consequen ly, applica ions un easible in
FaaS a e now possible. We implemen bu s compu ing a op
OpenWhisk and p o ide a communica ion middlewa e ha
seamlessly le e ages locali y wi h ze o-copy messaging. E al-
ua ion shows educed job in oca ion and communica ion la-
ency o a 2
×
speed-up in Te aSo and a 98.5% educ ion
in emo e communica ion in PageRank (13
×
speed-up) com-
pa ed o s anda d FaaS.
1 In oduc ion
The cloud o e s compu e in as uc u e on demand, bu p o-
isioning, adjus ing, and managing hese esou ces o la ge-
scale da a p ocessing applica ions is an a duous ask, espe-
cially o non-expe s. Fu he mo e, when he load is unp e-
dic able, dynamic, wi h a ying olumes o da a, use -d i en,
and some imes in e ac i e, inding he igh scale o a oid
misp o isioning [25,39,54] becomes e y complex.
Func ion-as-a-Se ice (FaaS) has gained ac ion as a so-
lu ion o he esou ce p o isioning p oblem as i o e s apid,
on-demand, no-ops scaling and a pay-as-you-go billing model
a e y ine g anula i y (MB pe ms). Use s do no need o
se up a clus e , bu he se ice simply accep s unc ion in-
oca ions and ully manages he es . Mo eo e , i s esou ce
bu s abili y has se FaaS aside om adi ional engines like
Spa k o Dask, allowing o s a housands o sho -li ed unc-
ions in seconds ins ead o minu es (see Table 1). Se e al
esea ch wo ks [
15
,
16
,
23
,
44
] ha e used FaaS o a my iad
o da a- and compu e-in ensi e asks.
Table 1: Time o p o ision cloud compu e esou ces on
di e en se ices and echnologies.
Technology To al CPUs NodesaS a -up ime
EMR Spa k 96 6 296 s
24 431 s
Da ap oc 96 6 95 s
24 113 s
Dask 128 8 184 s
64 253 s
Ray 128 8 187 s
64 229 s
Kna i e (Kube ne es) 960 960 54 s
OpenWhisk 960 960 21 s
AWS Lambda (2 GiB) 960 960 6 s
Bu s Compu ing 960 960 1.7 s
a
AWS EMR Spa k and GCP Da ap oc use m5 and E2-s anda d VM
amilies, espec i ely. Dask and Ray a e deployed on use -managed
m6i amily EC2 VMs. Kna i e, OpenWhisk, and Bu s deployed on
20 c7i.12xlal ge VMs bu s a 960 unc ions/wo ke s.
This has b ough a new concep in cloud compu ing ha
e e s o he abili y o quickly espond o sudden, pa allel
wo kloads wi hou p o isioning a clus e in ad ance. Fouladi
e al
. [16]
alk abou a “bu s able supe compu e -on-demand”
and a “bu s -pa allel swa m o housands o cloud unc ions,
all wo king on he same job.” Howe e , li e a u e admi s ha
he cu en FaaS model is oo na ow and p ecluding o mas-
si ely pa allel da a p ocessing p og ams (MPP) [24].
In his wo k, we highligh ha he key issue o FaaS hinde -
ing bu s -pa allel jobs is i s lack o g oup awa eness. Indeed,
FaaS use s need mul iple independen se ice calls o spawn
a lee o wo ke s, which become s ongly isola ed om each
o he . We no e ha such ine-g ained isola ion is damaging
and unnecessa y o collabo a i e jobs, and hus p opose o
aise he mul i- enan bounda ies o he job le el.
We p esen bu s compu ing, a new cloud compu ing model
o deal wi h quick, sudden, massi ely pa allel wo kloads,
which we call bu s s. To his end, we o e a g oup in o-
ca ion p imi i e o handle he whole job as a uni . To he bes
o ou knowledge, we a e he i s o implemen his ea u e
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 39
in a FaaS pla o m, clea ly di e ing om all o he esea ch
e o s ha su e he bu den o handling and o ches a ing
indi idual unc ion in oca ions. A g oup in oca ion allows o
op imize esou ce alloca ion, ensu e wo ke pa allelism, and
pe o m packing: unning mul iple wo ke s co-loca ed in he
same en i onmen . In addi ion o speed up wo ke s a -up
la ency, his enables wo ke locali y and simul anei y, which
can be exploi ed o imp o e code and da a loading, and o
aid powe ul wo ke - o-wo ke communica ion pa e ns (e.g.,
b oadcas , all- o-all) ha seamlessly le e age sha ed memo y
channels wi h ze o-copy mechanisms.
F om he use side, a bu s spawns a lee o wo ke s ha
communica e wi h message-passing, a simple bu e y powe -
ul abs ac ion ha c ea es a no el se e less subs a e e sa-
ile o many applica ions beyond wha is easible in FaaS. This
gi es ex ensi e con ol o he job o ad anced use s and allows
o design compu e engines o amewo ks on op (e.g., DAG-
based) o simpli y de elopmen , manage he execu ion, and
handle ailu es. Massi ely pa allel compu a ions a e p ime
bu s applica ions, especially when sudden, unp edic able, and
use -d i en in na u e. Ba ch-like jobs a e also good candida es
when un in e ac i ely. Some examples a e da a p ocessing,
analy ics, and machine lea ning wo kloads like explo a o y
model uning, SQL,
k
-means, and la ge-scale so ing. Bu s s
may be s a eless (e.g., g id sea ch o Mon e Ca lo simula ions)
o s a e ul (e.g., able joins and agg ega ions).
We make he ollowing con ibu ions:
•
We p esen bu s compu ing, a no el cloud se ice model
o sho , sudden, massi ely pa allel jobs (bu s s). We be-
lie e ha no cloud endo o esea ch e o has c ea ed
he necessa y subs a e o suppo hem.
•
Bu s compu ing e ol es FaaS wi h a key no el g oup
in oca ion p imi i e (a la e) ha aises mul i- enan iso-
la ion om a single unc ion in oca ion o he whole job.
In consequence, he sys em launches massi e p ocess
g oups as e , wi h gua an eed pa allelism, and packs
wo ke s oge he o exploi locali y.
•
We implemen a bu s compu ing pla o m by ex ending
OpenWhisk, a s a e-o - he-a FaaS sys em. Ou imple-
men a ion includes a specialized Rus wo ke un ime
and a bu s communica ion middlewa e ha seamlessly
le e age wo ke locali y wi h collec i e code/da a load-
ing and ze o-copy messaging.
•
Unde e alua ion on se e al bu s -pa allel wo kloads
agains FaaS, bu s compu ing imp o es job in oca ion
la ency (up o 11.5
×
as e ), wo ke simul anei y (up
o 26.5
×
lowe median absolu e de ia ion), and g oup
communica ion (up o 98% in a b oadcas ), o a speed-
up o 13×in PageRank and 2×in Te aSo .
2 Mo i a ion: in sea ch o bu s abili y
Many wo ks a e le e aging se e less se ices o massi e
da a p ocessing [
3
,
7
,
9
,
15
,
16
,
23
,
40
,
60
] despi e cu en
(FaaS) hind ances [
5
,
18
,
24
] due o he esou ce bu s abili y
0 2 4 6 8 10
Time (s)
0.0
0.5
1.0
CDF
100 x 256 MiB
100 x 10 GiB
1000 x 256 MiB
1000 x 10 GiB
1
Figu e 1: S a -up ime (cold s a ) o 100 and 1000 FaaS
unc ions in AWS Lambda o wo memo y sizes.
o his model [
17
,
38
]. Applica ions bene i om quick, on-
demand, no-ops esou ces a e y ine g anula i y, and pay
p ecisely o wha hey need, when hey need i .
This has b ough wha we call bu s s, massi ely pa allel p o-
cessing (MPP) wo kloads ha appea suddenly and p ocess
la ge, a iable olumes o da a in a e y sho ime (unde 1 o
2 minu es). Such applica ions ha e dynamic esou ce needs
ha canno be p edic ed easily, hus se e less bu s abili y be-
comes essen ial [
49
,
50
,
52
,
56
]. Conside , e.g., an in e ac i e,
scien is -d i en wo k low in a Jupy e No ebook, whe e he
use dynamically explo es la ge da ase s and modi ies pa ame-
e s ha signi ican ly impac he wo kload size. Al hough hey
may esemble ba ch jobs, hey a e cha ac e ized by hei sud-
den, spo adic occu ence, highly dynamic and unp edic able
da a olumes, and he expec a ion o low-la ency execu ion.
Rep esen a i e use cases include in e ac i e model uning
ia g id sea ch, explo a o y da a analysis wi h SQL que ies
and algo i hms such as logis ic eg ession, and da a p epa a-
ion ope a ions like il e ing and so ing. The MilliSo and
MilliQue y benchma ks [
31
] exempli y such sho - unning
wo kloads. Addi ional scena ios include eal- ime da a s eam
o ideo eed p ocessing, whe e bo h da a olume and analy -
ical complexi y may luc ua e d ama ically o e ime.
Cu en da a p ocessing solu ions such as Spa k, Dask,
Flink, o Ray ail o suppo bu s s. A long-li ed deploymen
o hese engines is imp ac ical, as i would easily become mis-
p o isioned. They a e no o e ed as a se ice by any cloud
ei he , which could pallia e he issue by mul iplexing jobs
om mul iple enan s, and hus o ces pe -use deploymen s
ha a e oo slow o se up [
38
], e en on cloud-managed o -
e ings (e.g., Amazon EMR). Table 1shows ha s a ing one
o hese echnologies is in ole able o c i ical spo adic o
dynamically sized applica ions.
In con as , FaaS se ices p o ide a la ge-scale compu e
subs a e much as e . Fig. 1shows ha AWS Lambda may
spawn a lee o 1000 unc ions in
6 s
; a much mo e app op i-
a e ime ange o bu s s.
1
FaaS is also mo e a ac i e han
Con aine -as-a-Se ice (CaaS) o managed Kube ne es se -
1
No e ha small unc ions (
256 MiB
) incu highe in oca ion la ency
han la ge ones (
10 GiB
). Also ound on o he p o ide s (e.g., GCP), i is
likely due o he o e head o scheduling ine -g ained esou ces.
40 2025 USENIX Annual Technical Con e ence USENIX Associa ion
Clien
Con olle
Ins ance 0
Ins ance 3
Ins ance 1
Ins ance 4
Wo ke 0 Wo ke 1
Wo ke 3 Wo ke 4
In oke (x6)
A1 A2
FaaSPla o m
Clien
Con olle
Pack 0
Pack 1
Fla e
B1
B2
Bu s Pla o m
Ins ance 2
Wo ke 2
Ins ance 5
Wo ke 5
Wo ke 0 Wo ke 1
Wo ke 3 Wo ke 4
Wo ke 2
Wo ke 5
B3
g=3
Remo e indi ec
Sha ed memo y
Wo ke communica ion:
A3
Ex e nal Comms. Se e
Ex e nal Comms. Se e
Figu e 2: Running a da a p ocessing job o 6 wo ke s in FaaS
and bu s compu ing wi h g anula i y (g) 3.
ices due o simple abs ac ions [
26
] and quicke esou ce al-
loca ion. Fo example, Kna i e, a Kube ne es-based FaaS-like
implemen a ion, is no iceably slowe in spawning wo ke s
han dedica ed FaaS pla o ms (Table 1).
2.1 FaaS is holding us back
A e iew o he li e a u e will show us ha unning bu s s a op
FaaS b ings many challenges [
18
,
24
,
38
]. We highligh h ee
ic ion poin s: (F1) wo ke isola ion, (F2) job agmen a ion
wi h complex o ches a ion, and (F3) huge da a mo emen .
To illus a e hem, Fig. 2 ollows he execu ion o a pa allel
job on a FaaS pla o m. I shows a pa allel job wi h 6 wo ke s.
The job could be emba assingly pa allel (s a eless) such as
a da a il e ing, o equi e he wo ke s o coo dina e a some
poin (s a e ul) such as a able join o , mo e in ensi ely due
o i s i e a i e na u e, PageRank.
F1 appea s because mul i- enan isola ion is a he le el o a
unc ion in oca ion. FaaS spawns unc ion ins ances indepen-
den ly, one a a ime, equi ing mul iple HTTP eques s
A1
o ob ain he 6 wo ke s. Besides he added la ency o se e al
eques s, his is an issue o pa allel jobs because he pla -
o m is no awa e o hese wo ke s being collabo a o s, and
hus canno gua an ee hei pa allelism. This c ea es delays o
skews be ween wo ke s ha po en ially ha m job execu ion.
Take, o ins ance, Fig. 1, whe e he las unc ion s a s up
o
6 s
a e he i s one.
2
E en mo e, he pla o m popula es
iden ical en i onmen s (ins ances) o each in oca ion
A2
,
2
Fu he e alua ion (no shown in he plo ) e eals ha his dispe si y
may inc ease o 44 s in GCP, o 20 s in an OpenWhisk deploymen .
which s esses he sys em wi h code, dependency, and da a
3
loading ha c ea es memo y duplica ion [41,53].
F2 occu s when wo ke s need o coo dina e. Fo ins ance,
Te aSo à la MapReduce includes a da a shu le amids he
job, and PageRank i e a i ely globally agg ega es a ec o .
Wo ke s canno communica e e ec i ely because hey may
no exis a he same ime (F1). Ins ead, wo ke s ead and w i e
in e media e da a asynch onously h ough an ex e nal s o age
solu ion. This pa e n (depic ed in Fig. 3) c ea es job ag-
men a ion ( unc ion s ages) and complica es i s o ches a ion,
especially in i e a i e algo i hms like PageRank (un easible
wi h his app oach [
6
]). Fi s , i inc eases da a mo emen and
equi es wo ke ec ea ion a each s age (adding code and da a
loading o e head). Second, i needs an ac i e wo k low o -
ches a ion p ocess o moni o he s a e o wo ke s and o e see
he o e all job p og ess.
4
Cen alized solu ions add a mos ly-
idle d i e componen while decen alized ask scheduling
adds a laye o complexi y o applica ions [
9
,
29
,
32
]. Nei he
can sol e he unde lying p oblem o wo ke isola ion.
These issues a e emphasized by F3. Wi h so many iny iso-
la ed wo ke s, mos communica ion pa e ns (e.g., a shu le)
equi e nume ous emo e connec ions
A3
. In da a p ocessing
wo kloads, his may esul in e y la ge da a ans e s, p e-
cluded by he (FaaS) lack o di ec communica ion [10,35].
3 Bu s Compu ing
Bu s compu ing is a no el pa adigm o unning bu s s in
he cloud. I o e comes he abo e ic ions wi h wo key
p inciples ha e ol e FaaS: g oup awa eness and locali y
exploi a ion. Fig. 2shows how his changes job in oca ion.
FaaS hinde s wo ke collabo a ion because mul i- enan
isola ion is a he le el o a single unc ion (F1). Because a job
belongs o a single enan , i makes sense o aise isola ion o
he job le el and handle all i s wo ke s as a g oup. To his end,
bu s compu ing p o ides a g oup in oca ion p imi i e, which
we call la e
B1
, o ins an ly launch massi e p ocess g oups
wi h gua an eed pa allelism. To he bes o ou knowledge, we
a e he i s o implemen his kind o p imi i e in a se e less
sys em. Fla es b ing g oup awa eness o he se ice, which
is key o pe o m wo ke packing
B2
, i.e., unning mul iple
wo ke s o a job in he same isola ed en i onmen . Packing
es ablishes wo ke locali y and enables se e al op imiza ions
discussed below.
In a la e, all wo ke s ha e gua an eed pa allelism and
access o he job con ex (e.g., he bu s size, IDs, o locali y),
which allows hem o communica e synch onously in pa e ns
un easible in FaaS, such as wo ke - o-wo ke message passing
and collec i es, ha simpli y job o ches a ion and a oid F2.
This di e ence is depic ed in Fig. 3.F3 is add essed because
communica ion
B3
can seamlessly exploi locali y and use
sha ed memo y mechanisms be ween wo ke s in he same
pack, which educes emo e ans e s.
3Fo ins ance, hype pa ame e uning uses he same da a in all wo ke s.
4
This can be pain ul since FaaS does no p o ide moni o ing mechanisms.
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 41
Clien /O ch.
𝜆1
𝜆2
Code/da a ge /send Compu e
Clien /T igge
FaaS
Bu s
Communica ion
···𝜆3
𝜆4
𝜆1
𝜆2
Wai
Se ice delay
Fla e
Figu e 3: Timeline compa ison o a pa allel job wi h FaaS
and bu s compu ing.
3.1 Wo ke packing and communica ion
Wo ke packing To un a la e, he bu s pla o m alloca es
n
wo ke s in o
m
packs; we say ha
n
is he bu s size. The
numbe o wo ke s pe pack is he bu s ’s g anula i y (
g=
n/m
). Thus, Fig. 2shows a bu s o size
n=6
whe e, by
se ing
g=3
, he pla o m only spawns 2 packs, each wi h
3 wo ke s. The highe
g
, he lowe
m
, educing he numbe
o en i onmen c ea ions, which is a c i ical pa o unc ion
in oca ion ime in FaaS. Then, wo ke code and dependencies
a e loaded only once pe pack and sha ed by all co-loca ed
wo ke s. This u he helps wi h ini ializa ion ime (especially
when dependencies a e la ge) and op imizes esou ce usage
(e.g., a oiding memo y duplica ion [
41
]). A simila easoning
applies o da a loading: wo ke s p ocessing he same da a
(like in hype pa ame e uning) download i jus once pe pack
and u ilize hei agg ega ed esou ces o speed up he ans e
(i.e., wi h pa allel downloads).
Choosing gis a ade-o be ween ease o sys em manage-
men and locali y maximiza ion. To illus a e ha , we iden i y
h ee s a egies
5
o wo ke packing: (i) he e ogeneous, whe e
wo ke s a e placed in con aine s as big as possible in he un-
de lying sys em machines; (ii) homogeneous, whe e wo ke s
a e placed in ixed-size con aine s; and (iii) mixed, whe e
wo ke s a e pu in ixed-size packs, bu i mul iple packs
all on o he same machine, hey a e me ged in o a single
con aine . The i s app oach maximizes locali y, bu i can
become a esou ce scheduling p oblem, as i is p one o ag-
men a ion. The homogeneous packing mi iga es ha issue,
bu i es ic s wo ke locali y. The hi d s a egy is he com-
p omise ha allows a as and lexible managemen while s ill
maximizing locali y (see §5). Gi en his complexi y, we a gue
ha he esponsibili y o se ing he g anula i y should lie
wi h he pla o m a he han he use , enabling be e con ol
5
S a egies mus conside how many esou ces we assign o each wo ke .
Fo simplici y, his pape conside s only CPUs and applies 1 CPU pe
wo ke , bu he s a egies wo k o any such assignmen .
o e esou ce scheduling and p o iding a mo e s eamlined
and use - iendly se ice.
Wo ke communica ion Bu s applica ions a e elas ically
dis ibu ed and collabo a i e. They a e coded as a single
unc ion un by all wo ke s ha accep s any wo ke mul iplic-
i y anspa en ly. Then, because wo ke s a e gua an eed o
be pa allel, hey may coo dina e synch onously by sending
messages and wi h common communica ion pa e ns.
To simpli y his, bu s compu ing includes a wo ke - o-
wo ke , message-passing communica ion middlewa e eadily
a ailable o wo ke s. The middlewa e seamlessly iden i ies
messages be ween wo ke s placed in he same pack o local
communica ion (ze o-copy). Only messages be ween packs
a e ans e ed emo ely, and he middlewa e op imizes hese
connec ions (e.g., a b oadcas only sends one message pe
pack). Remo e deli e y may be implemen ed wi h se e al
echnologies. Ou con ibu ions a e independen o his choice
because bu s compu ing educes any emo e communica ion
h ough packing. In his wo k, we ollow he usual app oach
in FaaS and only conside indi ec solu ions using an ex e nal
communica ion se e B3 .
4 Design and implemen a ion
We pu he abo e ideas in o a p o o ype bu s compu ing pla -
o m and communica ion middlewa e. He e we p o ide he
design de ails and implemen a ion. Fig. 4shows an o e iew
o he main componen s and hei in e ac ions.
The bu s pla o m ex ends he design o a FaaS pla o m
o implemen g oup in oca ion and wo ke packing. Buil
a op Apache OpenWhisk, ou pla o m sha es i s componen s
wi h impo an modi ica ions (see §4.4). The con olle man-
ages use in e ac ion wi h he pla o m, i handles inbound
HTTP eques s o deploy and in oke bu s s, o e sees sys-
em esou ces, and pe o ms wo ke packing. A da abase
s o es he bu s de ini ions and con igu a ion, as well as he
esul s and execu ion me ada a. Compu a ional esou ces in
he pla o m a e p o ided by he in oke s, a se o machines
wi h capaci y o bu s packs. Packs a e un in con aine s ha
isola e a cus om un ime en i onmen o un wo ke s.
Ou bu s communica ion middlewa e (BCM) has wo main
componen s: he co e communica ion lib a y and he emo e
backends. The lib a y exposes message-based communica-
ion o wo ke s, and i is ex ensible wi h backends o use
di e en emo e message deli e y solu ions.
4.1 Li e cycle o e iew
Fig. 4depic s he li e cycle o he sys em. To deploy a new
bu s de ini ion, he use i s sends
1
a
deploy
HTTP e-
ques . The con olle ecei es i and egis e s
2
he new
de ini ion in he da abase. La e , when he use desi es o ig-
ge he execu ion o he bu s , hey send
3
a
la e
HTTP
eques wi h speci ic pa ame e s. The con olle handles he
in oca ion and decides wo ke alloca ion
4
based on he
cu en s a e o he in oke machines. The a ec ed in oke s
42 2025 USENIX Annual Technical Con e ence USENIX Associa ion

Bu s pla o m
deploy()
Con olle
In oke s
Remo e communica ion backend
1
3fla e()
Pack i
Pack
(ano he
bu s )
BCM Bu s communica ion middlewa e
Sha ed memo y
Wo ke s
BCM
6
- Bu s defini ions
- Resul s
- Me ada a
2
5
4
Figu e 4: Bu s compu ing pla o m o e iew.
ecei e he ask o spawn he equi ed un ime en i onmen s
(packs) wi h space o as many wo ke s as needed. When
he en i onmen s boo , hei hos in oke ells hem which
bu s de ini ion and pa ame e s o load
5
om he da abase.
Then, each pack spawns i s wo ke s in e nally, which will
execu e he use -de ined unc ion (
wo k
in Table 2) in pa al-
lel. Wo ke s may use he BCM o coo dina e and sha e da a.
This seamlessly uses sha ed memo y o emo e connec ions
o communica e wo ke s
6
in he same o a di e en pack,
espec i ely. Addi ionally, wo ke s may ead o w i e da a
o ex e nal s o age sys ems (e.g., objec s o age) o p oduce
a esul ha is s o ed back o he da abase, whe e i may be
e ie ed la e by use s h ough ano he HTTP eques .
4.2 De eloping and unning bu s s
Use expe ience is key o bu s compu ing. As a se e less
se ice, all esou ce managemen emains hidden. Use s in-
e ac wi h he se ice h ough a simple in e ace ha allows
o de ine bu s s wi h esou ce-agnos ic code and o schedule
hei execu ion. This is simila o FaaS se ices ha allow
use s o upload hei unc ion de ini ions and hen se up ig-
ge s o in oke hem as needed. The in e ace and abs ac ions
a e summa ized in Table 2.
Deploymen Simila o unc ions in FaaS, de elope s pack-
age and upload hei bu s de ini ions (code) o he cloud,
gi ing hem a name and con igu a ion. The con igu a ion in-
cludes un ime pa ame e s and wo ke cha ac e is ics (such
as language and memo y size).
In oca ion Bu s de ini ions a e igge ed o execu ion like
unc ions in FaaS: an e en o HTTP eques no i ies he in en
o execu e a bu s wi h speci ic inpu pa ame e s. We call each
bu s in oca ion a bu s la e (Table 2). The main di e ence
wi h FaaS is ha a la e will spawn a g oup o pa allel wo ke s
(ins ead o a single unc ion ins ance). The se ice ensu es
ha all wo ke s un simul aneously and applies packing. In
ou p o o ype, he bu s size is explici on he size o he in-
pu Pa ams a ay. Hence, use s ha e di ec con ol o e i . We
belie e his o be impo an because pa allelism is s ic ly
applica ion-speci ic and depends on da a olume (e.g., ETL
Table 2: Bu s compu ing abs ac ions and API.
In e ace Func ions
Bu s deploy(de Name,package,con )
Se ice upload and deploy a bu s de ini ion
la e(de Name, [inpu Pa ams])
in okes a bu s
Bu s abs ac wo k(inpu Pa ams,bu s Con ex )
Func ion unc ion o un on each wo ke
Bu s wo ke ID →unique ID o his wo ke wi hin he la e
Con ex bu s Size →numbe o o al wo ke s in he la e
packID →unique ID o he cu en pack
packSize →numbe o wo ke s in he cu en pack
numPacks →numbe o packs wi hin he la e
belongToPack(wo ke ID)→packID
e u ns he pack ID o which a wo ke belongs o
isPackLeade () →bool
e u ns ue i his wo ke is i s pack’s leade
Comm. send(da a,des )→none
P imi i es ec (sou ce)→da a
b oadcas (da a, oo )→da a
allToAll([da a])→[da a]
educe(da a, (da a,da a)→da a)→da a
asks), da a con en (e.g., dimensionali y o spa si y), o algo-
i hm con igu a ion (e.g., he numbe o clus e s in
k
-means).
Sma bu s sizing is le o u u e wo k, i.e., he pla o m
may au oma ically calcula e he numbe o wo ke s based on
applica ion and da a in o ma ion.
Coding Bu s de ini ions a e coded as a single unc ion ha
is un by each wo ke in he bu s (
wo k
in Table 2). This
unc ion mus be p og ammed elas ically so ha i accep s
and uns co ec ly o any bu s size. The code is also agnos ic
o he packing pe o med by he se ice. To ha end, he
wo k
unc ion ecei es a bu s con ex objec h ough which each
wo ke may ob ain in o ma ion abou he wo ke dis ibu ion
wi hin he pa icula la e. Fo example, a wo ke can que y i s
own unique ID, he bu s size, g anula i y, o which wo ke s
belong o each pack (Bu s Con ex in Table 2). Wi h his
in o ma ion (p o ided by he pla o m in oke ), he code can
implemen logic o apply locali y op imiza ions a he pack
and bu s le els (see an example in §5.4.1). This con ex
objec also gi es access o he BCM.
Communica ion in e ace The BCM o e s simple ye pow-
e ul wo ke - o-wo ke communica ion h ough message pass-
ing simila o MPI. The abs ac ions a e elas ic (adap o he
bu s size) and a ailable h ough he bu s con ex . Bu s com-
pu ing p og ams make use o wo basic p imi i es o connec
wo ke s: send and ecei e. These p imi i es enable poin - o-
poin communica ion be ween wo ke s and a e designed o
send a bi a y olumes o da a e icien ly wi hin he bu s . To
acili a e common communica ion pa e ns in pa allel jobs,
bu s s may also use g oup collec i es. As lis ed in Table 2,
ou p o o ype implemen s b oadcas , all- o-all, and educe.
P imi i es and collec i es a e locali y-awa e, al hough he
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 43
n wo k(pa ams: Inpu , bu s : &Bu s Con ex ) -> Ou pu {
le num_nodes =pa ams.num_nodes;
le mu page_ anks = ec![1.0 /num_nodes; num_nodes];
le mu sum = ec![0.0; num_nodes];
le adjacency_ma ix =ge _adjacency_ma ix(¶ms);
while e <ERROR_THRESHOLD {
page_ anks =bu s .b oadcas (page_ anks, ROOT_WORKER);
o (node, links) in g aph {
o link in links {
sum[*link] += page_ anks[*node] /ou _links(*node);
}
}
le educed_ anks =bu s . educe(sum, | ec1, ec2|{
ec1.zip( ec2).map(|(a, b)|a+b).collec ()
});
i bu s .wo ke _id == ROOT_WORKER {
e =calcula e_e o (&page_ anks, & educed_ anks);
page_ anks = educed_ anks;
}
e =bu s .b oadcas (e , ROOT_WORKER);
ese _sums(&mu sum);
}
Ou pu { page_ anks }
}
Figu e 5: Simpli ied sou ce code o he PageRank wo k unc-
ion o bu s compu ing. The accesses o he bu s con ex o
ob ain he wo ke ID o communica e a e highligh ed.
p og ams emain agnos ic o i , i.e., co-loca ed wo ke s (same
pack) communica e on sha ed memo y and only emo e wo k-
e s hi he ne wo k.
4.3 Applica ion example
Fig. 5shows an example in Rus code (simpli ied) o he
wo k
unc ion ha implemen s he PageRank applica ion. The al-
go i hm consis s o an i e a i e p ocess in which each wo ke
holds a po ion o he adjacency g aph ( ela ing links be ween
web pages). In each i e a ion, he new global anks a e com-
pu ed in pa allel, agg ega ed, and educed in a ee s uc u e,
hen b oadcas ed om he oo wo ke o he es o hem. The
algo i hm uns un il i con e ges pas a h eshold o eaches a
limi o i e a ions.
Simila ly o he MPI compu ing model, all wo ke s execu e
he same code bu pe o m di e en logic based on he wo ke
ID ( he ank in MPI). The example highligh s he wo ke
accesses o he
Bu s Con ex
objec o pe o m collec i es
and ob ain in o ma ion abou he cu en la e. Fo example, i
is used o pe o m a collec i e b oadcas o sha e he upda ed
anks ec o , and la e a educe o agg ega e he pa ial anks
compu ed among he wo ke s. I also shows how a wo ke
checks i s ID when i needs o calcula e he con e gence,
since his is only done by he oo wo ke a e collec ing he
agg ega ed ec o in he educe.
4.4 Bu s pla o m implemen a ion
The p o o ype implemen a ion is buil on op o he popu-
la Apache OpenWhisk pla o m ( 1.0.0). We used Open-
Whisk as he basis because i is a well-known, open-sou ce,
p oduc ion- es ed FaaS implemen a ion and p o ides highe
bu s abili y han o he pla o ms like Kna i e (Table 1). Ou
changes amoun o app oxima ely
2K SLOC
. They a ec he
main componen s o he pla o m, including he con olle ,
he in oke , and he un ime en i onmen .
The con olle now suppo s wo new HTTP endpoin s o
bu s s:
deploy
and
la e
. I also implemen s he logic o
handle hem (§4.1). This includes he packing s a egy in he
h ee la o s (§3): he e ogeneous, homogeneous, and mixed.
G anula i y can be con igu ed. In any case, he con olle
calcula es he numbe and size o he packs based on he
speci ic bu s size and he esou ces a ailable in he in oke s.
In oke s un a new moni o ing logic ha can be adjus ed
o epo hei load o he con olle based on CPU ins ead
o RAM. Ou p o o ype is se o assign 1 CPU pe wo ke
because bu s s end o be compu e-in ensi e jobs and we do
no conside pa allelism wi hin a wo ke ,
6
bu o he con ig-
u a ions a e possible. In oke s also implemen new logic o
suppo he c ea ion and execu ion o packs, spawning Docke
con aine s o he app op ia e size o each bu s (by speci ying
esou ce limi s) and elling each con aine / un ime he num-
be o wo ke s o un, plus hei IDs and con ex . Con aine s
a e cu en ly no eused ac oss bu s s.
Fo he un ime, we adap ed he o icial OpenWhisk Rus
en i onmen , bu i is possible o suppo o he s. The new
logic allows o spawn mul iple wo ke s wi hin i as eques ed
by i s hos in oke . In pa icula , he Rus un ime spawns one
h ead pe wo ke o p o ide pa allelism. Finally, he un ime
also includes ou BCM buil -in.
4.5 BCM implemen a ion
The bu s communica ion middlewa e (BCM) is coded in
Rus in abou
5K SLOC
. I is eadily a ailable o ou cus om
Rus un ime and we a e wo king on a binding o Py hon.
7
I
enables he ansmission o in a-pack (ze o-copy) and in e -
pack ( ia emo e backend) messages. The BCM is ins an ia ed
by he un ime (once pe pack) and made a ailable o wo ke s
as a pa ame e (in he wo k unc ion as shown in Table 2).
Fo local communica ion, BCM uses in-memo y queues
o send and ecei e da a be ween wo ke s in he same pack.
In he Rus un ime, wo ke s a e h eads and eside in he
same memo y space, so sha ed memo y mechanisms a e no
necessa y (e.g.,
shm_open
o
mmap
). Ins ead, wo ke s jus pass
memo y poin e s be ween hem. Thanks o Rus ’s memo y
sa e y gua an ees, access o sha ed da a is h ead-sa e. Rus
also p o ides a e e ence-coun ing mechanism o immu able
da a, so sha ed da a is eleased when i is no longe used a
un ime. Fo example, he oo wo ke in a b oadcas sends a
ead-only memo y poin e o i s local wo ke s, and hey sa ely
access he message concu en ly. To modi y he da a, one may
use mechanisms such as copy-on-w i e.
Fo emo e communica ion, each pack has a sha ed connec-
ion pool o he emo e backend, which allows each wo ke
wi hin he pack o send and ecei e messages concu en ly,
wi h he goal o maximizing he con aine ’s bandwid h. This
6The bu s size (numbe o wo ke s) de e mines o al job pa allelism.
7
O he languages may be suppo ed h ough bindings (Ja a, C++, Go. . . ).
44 2025 USENIX Annual Technical Con e ence USENIX Associa ion
is especially use ul in p imi i es like all- o-all, whe e all wo k-
e s mus open channels wi h all he o he s. Fo la ge messages,
he da a is spli in o smalle chunks ha a e sen and ecei ed
concu en ly. This maximizes ne wo k u iliza ion and allows
eade s o s a ecei ing da a om he i s chunk, ins ead o
wai ing o he ull message o be a ailable a he backend.
The BCM is ex ensible, allowing he implemen a ion o
mo e emo e backends. Cu en ly, we suppo Redis, D ag-
on lyDB, Rabbi MQ, and S3. The backend in e ace di e -
en ia es be ween sending di ec messages (one- o-one) and
b oadcas messages (one- o-many). The eason is ha di ec
messages a e ead only once, while b oadcas messages mul-
iple imes, so we wan o op imize his pa icula case. Fo
ins ance, in Rabbi MQ, one- o-one messages use di ec b o-
ke s, while one- o-many use an-ou b oke s.
To ensu e ha no messages a e los (a -leas -once deli -
e y seman ics), he BCM elies i s on he backend deli e y
gua an ees (e.g., Rabbi MQ uses du able queues o a oid
d opping messages). Addi ionally, he BCM keeps a coun o
di ec messages sen be ween each pai o wo ke s, and o
each collec i e ope a ion. The middlewa e handles duplica e
and/o ou -o -o de messages. Fo ha , messages include a
heade wi h he sou ce and des ina ion wo ke , collec i e ype,
coun e , and, i chunked, he numbe o chunks and chunk
numbe . Messages wi h a coun e lowe han he expec ed
alue a e igno ed and assumed as al eady p ocessed. Those
wi h a coun e g ea e han expec ed a e cached locally un il
needed. Fo chunked messages ecei ed ou -o -o de , a mem-
o y egion is ese ed o he o al payload and chunks a e
w i en o hei espec i e o se as hey come in.
5 E alua ion
Ou e alua ion aims o assess bu s compu ing agains cu en
FaaS on he h ee ic ion poin s desc ibed in §2.1. Impo -
an ly, we show and analyze he e ec s o wo ke packing
and locali y. All expe imen s un on Amazon Web Se ices
(AWS) in he us-eas -1 egion.
5.1 Bu s g oup in oca ion
G oup in oca ion is he key elemen agains ic ion F1. He e
we e alua e how job-le el isola ion imp o es wo ke eadi-
ness ime (in oca ion la ency), ensu es hei simul anei y, and
p o ides locali y o collabo a i e code and da a loading.
Se up: The bu s pla o m uns on an Amazon EKS clus e ,
wi h he con ol plane on a
4i.xla ge
VM (4 CPUs and
16 GB RAM), and he in oke s on up o 20
c7i.12xla ge
VMs (48 CPUs and 96 GB RAM). This gi es us space o
accommoda e up o 960 wo ke s wi h 1 CPU each.
Impac on bu s in oca ion la ency Fi s , we use he homo-
geneous packing policy o e alua e how assigning di e en
g anula i y (
g
) a ec s bu s in oca ion la ency. The explo-
a ion is depic ed in Fig. 6 o wo bu s s o sizes 48 (le ) and
960 ( igh ).
8
I is quickly appa en ha as
g
inc eases (up o
8
We conduc ed expe imen s wi h simila esul s o bu s sizes in-be ween.
FaaS
2
4
6
12
24
48
0 5 10 15 20
FaaS
2
4
6
12
24
48
0 5 10 15 20
S a -up ime (s)
G anula i y
Figu e 6: Wo ke s a -up la ency dis ibu ion wi hin one job
o bu s compu ing wi h di e en packing g anula i y and
FaaS (equi alen o
g=1
). Le and igh show, espec i ely,
bu s sizes o 48 and 960.
Las wo ke ini Simul anei y Wo ke li e- ime
0
250
500
750
1000
0 5 10 15 20 25
0
250
500
750
1000
0 5 10 15 20 25
Time (s)
# ac i a ion
Figu e 7: Simul anei y (numbe o wo ke s unning a an
ins an ) in FaaS (le ) and Bu s wi h
g=48
( igh ). Each ba
ep esen s he li e- ime o a wo ke .
48 in bo h cases), he s a -up ime dec eases, and gene ally
becomes mo e consis en ac oss wo ke s o all bu s sizes.
Fo ins ance, he la ency o ha ing all wo ke s eady in a bu s
o size 960 educes by 11.5
×
om
g=1
(FaaS) o
g=48
. We
ound ha con aine c ea ion domina es in oca ion la ency,
hence highe
g
pe o ms bes . This p o es ha c ea ing he
bigges possible con aine s, and hus he less amoun o hem
(he e ogeneous packing), achie es he bes s a -up la ency,
since i c ea es a single con aine pe in oke pe la e. By ex-
ension, he mixed packing s a egy exhibi s he same esul s,
bu allows he sys em o manage esou ces mo e e ec i ely
in small po ions o acili a e alloca ion and a oid esou ce
agmen a ion. To assess he impac o g anula i y, he es o
he e alua ion uses homogeneous packing.
Impac on wo ke simul anei y We un a bu s wi h size
960 on FaaS agains bu s compu ing wi h
g=48
. Fo demon-
s a ion pu poses, each wo ke pe o ms a 5-second sleep and
we plo hei execu ion imeline in Fig. 7. The plo shows
ha bu s compu ing achie es as e esou ce alloca ion and
quicke eadiness o wo ke s. This ensu es wo ke pa allelism.
Analyzing dispe si y o wo ke s a -up ime (also in Fig. 6),
he FaaS execu ion e inces a ange o
18.8 s
be ween he
s a o he i s wo ke and ha o he las one, wi h a median
absolu e de ia ion (MAD) o
2.65 s
. In con as , he ange
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 45
FaaS
2
4
6
12
24
48
100GB 50GB 0 5s 10s 15s 20s
Download size Download ime
G anula i y
Figu e 8: A bu s o 96 wo ke s loading he same
1 GiB
objec
om S3 wi h di e en g anula i y.
wi h
g=48
is jus
0.44 s
(MAD is
0.1 s
). Compa ed, he
ange is 43
×
lowe in bu s compu ing, wi h MAD show-
ing 26.5
×
lowe dispe si y han FaaS. Dispe si y in wo ke
s a -up la ency p ecludes FaaS o achie e ull pa allelism (all
wo ke s unning simul aneously om s a o inish), while
bu s gua an ees i .
Impac on da a loading Bu s compu ing mi iga es he FaaS
p oblem o loading he same da a on all unc ions (§2.1), e.g.,
in a g id sea ch. We can le e age wo ke access o locali y
in o ma ion o op imize his p oblem and download he da a
only once pe pack, i ially educing da a inges ion. Speci i-
cally, each wo ke in a pack e ie es a pa o he da a based
on calcula ions om pack in o ma ion in he bu s con ex
(Table 2). Then hey ec ea e he ull da a in a local sha ed
memo y egion. This allows o pa allelize he download and
comple e he p ocess as e han choosing a leade o pe o m
i (also possible h ough he
isPackLeade
unc ion in he
con ex , which e u ns ue o he wo ke wi h lowes ID
wi hin i s pack). We e alua e his app oach on mul iple
g
and
p esen i in Fig. 8. Bu s op imiza ions achie e a download
ime speed-up o 32.6×wi h g=48 compa ed o FaaS.
Takeaway Fla es elimina e ic ion F1 h ough as e wo ke
g oup ini ializa ion (11.5
×
) and ensu ed simul anei y (43
×
less dispe sed wo ke s) ha enables locali y wi h packing. In
u n, locali y may accele a e da a download in applica ions
(32.6×), ackling ic ion F3.
5.2 Bu s in e -pack communica ion
Be o e we e alua e he e ec s o he BCM on ic ions F2 and
F3, we wan o ensu e ha an indi ec communica ion model
is easible and o ind a backend ha sus ains he load o
bu s s a scale. Fo his, we measu e he h oughpu o se e al
indi ec communica ion backends. Speci ically, we es Redis,
D agon lyDB (a Redis-compa ible mul i- h eaded al e na i e),
Rabbi MQ, and S3. Redis and D agon lyDB e alua e wo
la o s: using lis s o s eams.
Message chunk size The BCM chunks messages in o se -
e al blocks o op imize ne wo k u iliza ion and allow pa allel
ead/w i e. The op imal chunk size is a ade-o be ween
la ency o i s by e and ope a ion o e head, and i a ies
64 KiB 1 MiB 64 MiB 128 MiB 256 MiB
Chunk Size
0
100
200
300
400
Th oughpu (MiB/s)
Rabbi MQ
Redis Lis
D agon lyDB Lis
Redis S eam
D agon lyDB S eam
S3
1
(a) Th oughpu be ween wo emo e wo ke s sending a
1 GiB
pay-
load chunked in di e en sizes.
8 16 32 96 192 384
Bu s Size
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Agg ega ed
Th oughpu (GiB/s)
1
(b) Agg ega e h oughpu o wo emo e packs, A and B, o a ying
size (
g=
bu s size
/2
), whe e each wo ke om pack A sends a
256 MiB payload o ano he wo ke om emo e pack B.
Figu e 9: Th oughpu expe imen s o he di e en BCM
backends. Median alues wi h s anda d de ia ion (10 uns).
o each communica ion backend. To ind he op imal con-
igu a ion, we measu e he h oughpu o sending a
1 GiB
message be ween wo emo e wo ke s. The wo ke s un on
wo
c7i.la ge
machines (4 CPUs, 8 GB) and we deploy
a
c7i.16xla ge
(64 CPUs, 128 GB) o he in e media e
se e . Fig. 9a plo s he esul s. Rabbi MQ o e s a cons an
h oughpu o la ge chunk sizes, bu does no allow payloads
la ge han
128 MiB
due o AMQP p o ocol limi a ions. Re-
dis and D agon lyDB wo k bes a
1 MiB
, he la e being
sligh ly supe io . S3 o e s he lowes h oughpu because
objec s o es a e no designed o small iles (
1 MiB
o less
exceeds he allowed se ice eques a e limi s).
Maximum h oughpu To unde s and how he di e en
backends scale unde pa allel load, we measu e he agg e-
ga ed h oughpu be ween se e al pai s o wo ke s commu-
nica ing simul aneously. In his expe imen , we launch a
g oup o wo ke s (bu s size om 8 o 384) spli in o wo
emo e g oups. Each wo ke in a g oup A sends a ixed mes-
sage (
256 MiB
) o a wo ke in he o he , emo e g oup B.
As he bu s size inc eases, so does he o al da a olume
sen . Each backend uses he op imal chunk size assessed
in he mic o-benchma k abo e. Wo ke s un on wo VMs
scaled o he bu s size ( om
c7i.xla ge
o 8 wo ke s o
46 2025 USENIX Annual Technical Con e ence USENIX Associa ion
and Kei h Wins ein. 2017. Encoding, Fas and
Slow: Low-La ency Video P ocessing Using Thou-
sands o Tiny Th eads. In 14 h USENIX Sympo-
sium on Ne wo ked Sys ems Design and Implemen a-
ion (NSDI 17). USENIX Associa ion, Bos on, MA,
363–376. h ps://www.usenix.o g/con e ence/nsdi17/
echnical-sessions/p esen a ion/ ouladi
[17]
Ped o Ga cia Lopez, Aleksande Slominski, Be na d
Me zle , Michael Be hend , and Simon Shillake . 2024.
Se e less End Game: Disagg ega ion enabling T ans-
pa ency. In P oceedings o he 2nd Wo kshop on SE e -
less Sys ems, Applica ions and ME hodologies (A hens,
G eece) (SESAME ’24). Associa ion o Compu ing Ma-
chine y, New Yo k, NY, USA, 9–14. h ps://doi.o g/10.
1145/3642977.3652094
[18]
Joseph M. Helle s ein, Jose Falei o, Joseph E. Gonza-
lez, Johann Schleie -Smi h, Vik am S eekan i, Alexey
Tumano , and Chenggang Wu. 2018. Se e less Com-
pu ing: One S ep Fo wa d, Two S eps Back. h ps:
//doi.o g/10.48550/ARXIV.1812.03651
[19]
Bo Huang, Shengsheng Huang, Jinquan Dai, Jie Huang,
and Tao Xie. 2010. The HiBench benchma k sui e:
Cha ac e iza ion o he MapReduce-based da a analysis.
In 2010 IEEE 26 h In e na ional Con e ence on Da a
Enginee ing Wo kshops (ICDEW 2010). IEEE Com-
pu e Socie y, Los Alami os, CA, USA, 41–51. h ps:
//doi.o g/10.1109/ICDEW.2010.5452747
[20]
Aman Jain, A a F. Baa zi, Geo ge Kesidis, Bhu an U -
gaonka , Nade Al a es, and Mahmu Kandemi . 2020.
Spli Se e: E icien ly Spli ing Apache Spa k Jobs
Ac oss FaaS and IaaS. In P oceedings o he 21s In e -
na ional Middlewa e Con e ence (Del , Ne he lands)
(Middlewa e ’20). Associa ion o Compu ing Machin-
e y, New Yo k, NY, USA, 236–250. h ps://doi.o g/10.
1145/3423211.3425695
[21]
Zhipeng Jia and Emme Wi chel. 2021. Nigh co e:
e icien and scalable se e less compu ing o la ency-
sensi i e, in e ac i e mic ose ices. In P oceedings o
he 26 h ACM In e na ional Con e ence on A chi ec-
u al Suppo o P og amming Languages and Ope -
a ing Sys ems (ASPLOS ’21). Associa ion o Compu -
ing Machine y, New Yo k, NY, USA, 152–166. h ps:
//doi.o g/10.1145/3445814.3446701
[22]
Chao Jin, Zili Zhang, Xingyu Xiang, Songyun Zou,
Gang Huang, Xuanzhe Liu, and Xin Jin. 2023. Di o: E -
icien Se e less Analy ics wi h Elas ic Pa allelism. In
P oceedings o he ACM SIGCOMM 2023 Con e ence
(New Yo k, NY, USA) (ACM SIGCOMM ’23). Associ-
a ion o Compu ing Machine y, New Yo k, NY, USA,
406–419. h ps://doi.o g/10.1145/3603269.3604816
[23]
E ic Jonas, Qi an Pu, Shi a am Venka a aman, Ion S o-
ica, and Benjamin Rech . 2017. Occupy he cloud:
dis ibu ed compu ing o he 99%. In P oceedings
o he 2017 Symposium on Cloud Compu ing (SoCC
’17). Associa ion o Compu ing Machine y, New Yo k,
NY, USA, 445–451. h ps://doi.o g/10.1145/3127479.
3128601
[24]
E ic Jonas, Johann Schleie -Smi h, Vik am S eekan i,
Chia-Che Tsai, Anu ag Khandelwal, Qi an Pu, Vaishaal
Shanka , Joao Menezes Ca ei a, Ka l K au h, Nee -
aja Yadwadka , Joseph Gonzalez, Raluca Ada Popa,
Ion S oica, and Da id A. Pa e son. 2019. Cloud
P og amming Simpli ied: A Be keley View on Se e -
less Compu ing. Technical Repo UCB/EECS-2019-
3. EECS Depa men , Uni e si y o Cali o nia, Be ke-
ley. h ps://www2.eecs.be keley.edu/Pubs/TechRp s/
2019/EECS-2019-3.pd
[25]
Kos is Ka es, Nee aja J. Yadwadka , and Ch is os
Kozy akis. 2019. Cen alized Co e-g anula Scheduling
o Se e less Func ions. In P oceedings o he ACM
Symposium on Cloud Compu ing (SoCC ’19). Associ-
a ion o Compu ing Machine y, New Yo k, NY, USA,
158–164. h ps://doi.o g/10.1145/3357223.3362709
[26]
Nima Ka iani, Dmi iy Kalinin, and Michael Maximi-
lien. 2019. Towa ds Se e less as Commodi y: a case
o Kna i e. In P oceedings o he 5 h In e na ional
Wo kshop on Se e less Compu ing (Da is, CA, USA)
(WOSC ’19). Associa ion o Compu ing Machine y,
New Yo k, NY, USA, 13–18. h ps://doi.o g/10.1145/
3366623.3368135
[27]
Jeongchul Kim and Kyungyong Lee. 2019. Func ion-
Bench: A Sui e o Wo kloads o Se e less Cloud Func-
ion Se ice. In 2019 IEEE 12 h In e na ional Con-
e ence on Cloud Compu ing (CLOUD). IEEE Com-
pu e Socie y, Los Alami os, CA, USA, 502–504. h ps:
//doi.o g/10.1109/CLOUD.2019.00091
[28]
Ana Klimo ic, Yawen Wang, Pa ick S uedi, Animesh
T i edi, Jonas P e e le, and Ch is os Kozy akis. 2018.
Pocke : Elas ic Epheme al S o age o Se e less Ana-
ly ics. In 13 h USENIX Symposium on Ope a ing Sys-
ems Design and Implemen a ion (OSDI 18). USENIX
Associa ion, Ca lsbad, CA, 427–444. h ps://www.
usenix.o g/con e ence/osdi18/p esen a ion/klimo ic
[29]
Swa oop Ko ni, Ajay Nayak, Vinod Ganapa hy, and
A kap a a Basu. 2021. Faas lane: Accele a ing
Func ion-as-a-Se ice Wo k lows. In 2021 USENIX An-
nual Technical Con e ence (USENIX ATC 21). USENIX
Associa ion, Bos on, MA, 805–820. h ps://www.
usenix.o g/con e ence/a c21/p esen a ion/ko ni
[30]
Collin Lee and John Ous e hou . 2019. G anula Com-
pu ing. In P oceedings o he Wo kshop on Ho Topics in
Ope a ing Sys ems (Be ino o, I aly) (Ho OS ’19). Asso-
cia ion o Compu ing Machine y, New Yo k, NY, USA,
149–154. h ps://doi.o g/10.1145/3317550.3321447
[31]
Yilong Li, Seo Jin Pa k, and John Ous e hou . 2021.
MilliSo and MilliQue y: La ge-Scale Da a-In ensi e
Compu ing in Milliseconds. In 18 h USENIX Sympo-
sium on Ne wo ked Sys ems Design and Implemen a-
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 53

ion (NSDI 21). USENIX Associa ion, Bos on, MA,
593–611. h ps://www.usenix.o g/con e ence/nsdi21/
p esen a ion/li-yilong
[32]
Yiming Li, Laiping Zhao, Yanan Yang, and Wenyu Qu.
2023. Re hinking Deploymen o Se e less Func-
ions: A Pe o mance-Fi s Pe spec i e. In P oceedings
o he In e na ional Con e ence o High Pe o mance
Compu ing, Ne wo king, S o age and Analysis (Den-
e , CO, USA) (SC ’23). Associa ion o Compu ing
Machine y, New Yo k, NY, USA, A icle 67, 14 pages.
h ps://doi.o g/10.1145/3581784.3613211
[33]
B yan Lis on. 2016. Ad Hoc Big Da a P o-
cessing Made Simple wi h Se e less MapRe-
duce. Re ie ed May 02, 2024 om h ps:
//aws.amazon.com/blogs/compu e/ad-hoc-big-da a-
p ocessing-made-simple-wi h-se e less-map educe/
[34]
Da id H. Liu, Ami Le y, Shadi Noghabi, and Sebas ian
Bu ckha d . 2023. Doing Mo e wi h Less: O ches a -
ing Se e less Applica ions wi hou an O ches a o . In
20 h USENIX Symposium on Ne wo ked Sys ems De-
sign and Implemen a ion (NSDI 23). USENIX Associa-
ion, Bos on, MA, 1505–1519. h ps://www.usenix.o g/
con e ence/nsdi23/p esen a ion/liu-da id
[35]
Fangming Lu, Xingda Wei, Zhuobin Huang,
Rong Chen, Minyu Wu, and Haibo Chen. 2024.
Se ializa ion/Dese ializa ion- ee S a e T ans e in
Se e less Wo k lows. In P oceedings o he Nine-
een h Eu opean Con e ence on Compu e Sys ems
(A hens, G eece) (Eu oSys ’24). Associa ion o
Compu ing Machine y, New Yo k, NY, USA, 132–147.
h ps://doi.o g/10.1145/3627703.3629568
[36]
Ash a Mahgoub, Edga do Ba sallo Yi, Ka hick
Shanka , Sameh Elnike y, Somali Cha e ji, and Sau abh
Bagchi. 2022. ORION and he Th ee Righ s: Sizing,
Bundling, and P ewa ming o Se e less DAGs. In 16 h
USENIX Symposium on Ope a ing Sys ems Design and
Implemen a ion (OSDI 22). USENIX Associa ion, Ca ls-
bad, CA, 303–320. h ps://www.usenix.o g/con e ence/
osdi22/p esen a ion/mahgoub
[37]
Sh u i Mohan y, Vi ek M. Bhasi, Myungjun Son, Mah-
mu Taylan Kandemi , and Chi a Das. 2024. FAAS loop:
Op imizing Loop-Based Applica ions o Se e less
Compu ing. In P oceedings o he 2024 ACM Sympo-
sium on Cloud Compu ing (Redmond, WA, USA) (SoCC
’24). Associa ion o Compu ing Machine y, New Yo k,
NY, USA, 943–960. h ps://doi.o g/10.1145/3698038.
3698560
[38]
Ingo Mülle , Rod igo B uno, Ana Klimo ic, John
Wilkes, E ic Sedla , and Gus a o Alonso. 2020. Se e -
less Clus e s: The Missing Piece o In e ac i e Ba ch
Applica ions? (Ap il 2020), 3 pages. h ps://doi.o g/10.
3929/e hz-b-000405616 P esen ed a 10 h Wo kshop on
Sys ems o Pos -Moo e A chi ec u es (SPMA 2020).
[39]
Ge a d Pa ís, Ped o Ga cía-López, and Ma c Sánchez-
A igas. 2020. Se e less Elas ic Explo a ion o Un-
balanced Algo i hms. In 2020 IEEE 13 h In e na ional
Con e ence on Cloud Compu ing (CLOUD). IEEE Com-
pu e Socie y, Los Alami os, CA, USA, 149–157. h ps:
//doi.o g/10.1109/CLOUD49709.2020.00033 ISSN:
2159-6190.
[40]
Qi an Pu, Shi a am Venka a aman, and Ion S oica. 2019.
Shu ling, Fas and Slow: Scalable Analy ics on Se e -
less In as uc u e. In 16 h USENIX Symposium on Ne -
wo ked Sys ems Design and Implemen a ion (NSDI 19).
USENIX Associa ion, Bos on, MA, 193–206. h ps:
//www.usenix.o g/con e ence/nsdi19/p esen a ion/pu
[41]
Wei Qiu, Ma cin Copik, Yun Wang, Alexand u Calo oiu,
and To s en Hoe le . 2023. Use -guided Page Me ging
o Memo y Deduplica ion in Se e less Sys ems. In
2023 IEEE In e na ional Con e ence on Big Da a (Big-
Da a). IEEE Compu e Socie y, Los Alami os, CA, USA,
159–169. h ps://doi.o g/10.1109/BigDa a59044.2023.
10386487
[42]
Zhenyuan Ruan, Shihang Li, Kaiyan Fan, Seo Jin
Pa k, Ma cos K. Aguile a, Adam Belay, and Mal e
Schwa zkop . 2025. Quicksand: Ha nessing S anded
Da acen e Resou ces wi h G anula Compu ing. In
22nd USENIX Symposium on Ne wo ked Sys ems De-
sign and Implemen a ion (NSDI 25). USENIX Associa-
ion, Philadelphia, PA, 147–165. h ps://www.usenix.
o g/con e ence/nsdi25/p esen a ion/ uan
[43]
Josep Sampe, Ma c Sanchez-A igas, Gil Ve nik, Ido
Yehekzel, and Ped o Ga cia-Lopez. 2023. Ou sou cing
Da a P ocessing Jobs Wi h Li hops. IEEE T ansac ions
on Cloud Compu ing 11, 01 (Jan. 2023), 1026–1037.
h ps://doi.o g/10.1109/TCC.2021.3129000
[44]
Josep Sampé, Gil Ve nik, Ma c Sánchez-A igas, and
Ped o Ga cía-López. 2018. Se e less Da a Analy -
ics in he IBM Cloud. In P oceedings o he 19 h In-
e na ional Middlewa e Con e ence Indus y (Rennes,
F ance) (Middlewa e ’18). Associa ion o Compu ing
Machine y, New Yo k, NY, USA, 1–8. h ps://doi.o g/
10.1145/3284028.3284029
[45]
Josep Sampé, Ma c Sánchez-A igas, Ped o Ga cía-
López, and Ge a d Pa ís. 2017. Da a-d i en se e -
less unc ions o objec s o age. In P oceedings o he
18 h ACM/IFIP/USENIX Middlewa e Con e ence (Mid-
dlewa e ’17). Associa ion o Compu ing Machine y,
New Yo k, NY, USA, 121–133. h ps://doi.o g/10.1145/
3135974.3135980
[46]
Ma c Sánchez-A igas, Ge mán T. Eizagui e, Gil
Ve nik, Lachlan S ua , and Ped o Ga cía-López. 2020.
P imula: a P ac ical Shu le/So Ope a o o Se e -
less Compu ing. In P oceedings o he 21s In e na-
ional Middlewa e Con e ence Indus ial T ack (Del ,
Ne he lands) (Middlewa e ’20). Associa ion o Com-
pu ing Machine y, New Yo k, NY, USA, 31–37. h ps:
//doi.o g/10.1145/3429357.3430522
54 2025 USENIX Annual Technical Con e ence USENIX Associa ion
[47]
T e e Schi me , Joel Scheune , Tobias P andzel e , and
Da id Be mbach. 2022. Fusionize: Imp o ing Se e -
less Applica ion Pe o mance h ough Feedback-D i en
Func ion Fusion. In 2022 IEEE In e na ional Con e -
ence on Cloud Enginee ing (IC2E). IEEE Compu e
Socie y, Los Alami os, CA, USA, 85–95. h ps://doi.
o g/10.1109/IC2E55432.2022.00017
[48]
Ca los Sega a, Simon Shillake , Guo Li, Ele he ia
Mappou a, Rod igo B uno, Lluís Vilano a, and Pe e
Pie zuch. 2025. GRANNY: G anula Managemen
o Compu e-In ensi e Applica ions in he Cloud. In
22nd USENIX Symposium on Ne wo ked Sys ems De-
sign and Implemen a ion (NSDI 25). USENIX Associa-
ion, Philadelphia, PA, 205–218. h ps://www.usenix.
o g/con e ence/nsdi25/p esen a ion/sega a
[49]
Vaishaal Shanka , Ka l K au h, Kailas Vod ahalli, Qi-
an Pu, Benjamin Rech , Ion S oica, Jona han Ragan-
Kelley, E ic Jonas, and Shi a am Venka a aman. 2020.
Se e less linea algeb a. In P oceedings o he 11 h
ACM Symposium on Cloud Compu ing (Vi ual E en ,
USA) (SoCC ’20). Associa ion o Compu ing Machin-
e y, New Yo k, NY, USA, 281–295. h ps://doi.o g/10.
1145/3419111.3421287
[50]
Simon Shillake and Pe e Pie zuch. 2020. Faasm:
Ligh weigh Isola ion o E icien S a e ul Se e less
Compu ing. In 2020 USENIX Annual Technical Con e -
ence (USENIX ATC 20). USENIX Associa ion, Bos on,
MA, 419–433. h ps://www.usenix.o g/con e ence/
a c20/p esen a ion/shillake
[51]
Won Wook Song, Taegeon Um, Sameh Elnike y,
Myeongjae Jeon, and Byung-Gon Chun. 2023. Sponge:
Fas Reac i e Scaling o S eam P ocessing wi h Se e -
less F amewo ks. In 2023 USENIX Annual Techni-
cal Con e ence (USENIX ATC 23). USENIX Associ-
a ion, Bos on, MA, 301–314. h ps://www.usenix.o g/
con e ence/a c23/p esen a ion/song
[52]
Vik am S eekan i, Chenggang Wu, Xiayue Cha les Lin,
Johann Schleie -Smi h, Joseph E. Gonzalez, Joseph M.
Helle s ein, and Alexey Tumano . 2020. Cloudbu s :
s a e ul unc ions-as-a-se ice. P oceedings o he
VLDB Endowmen 13, 12 (July 2020), 2438–2452.
h ps://doi.o g/10.14778/3407790.3407836
[53]
Jo an S ojko ic, Tianyin Xu, Hube us F anke, and
Josep To ellas. 2023. MXFaaS: Resou ce Sha ing
in Se e less En i onmen s o Pa allelism and E i-
ciency. In P oceedings o he 50 h Annual In e na-
ional Symposium on Compu e A chi ec u e (O lando,
FL, USA) (ISCA ’23). Associa ion o Compu ing Ma-
chine y, New Yo k, NY, USA, A icle 34, 15 pages.
h ps://doi.o g/10.1145/3579371.3589069
[54]
Ali Ta iq, Aus in Pahl, Sha a Nimmagadda, E ic Rozne ,
and Siddha h Lanka. 2020. Sequoia: enabling quali y-
o -se ice in se e less compu ing. In P oceedings o
he 11 h ACM Symposium on Cloud Compu ing (SoCC
’20). Associa ion o Compu ing Machine y, New Yo k,
NY, USA, 311–327. h ps://doi.o g/10.1145/3419111.
3421306
[55]
Shelby Thomas, Lixiang Ao, Geo ey M. Voelke , and
Geo ge Po e . 2020. Pa icle: epheme al endpoin s o
se e less ne wo king. In P oceedings o he 11 h ACM
Symposium on Cloud Compu ing (SoCC ’20). Associ-
a ion o Compu ing Machine y, New Yo k, NY, USA,
16–29. h ps://doi.o g/10.1145/3419111.3421275
[56]
Bishoy Wadie, Lachlan S ua , Ch is ophe M. Ra h,
Be nha d D o le , Se gii Mamedo , and Theodo e
Alexand o . 2024. METASPACE-ML: Con ex -speci ic
me aboli e anno a ion o imaging mass spec ome y
using machine lea ning. Na u e Communica ions 15,
9110 (2024), 16 pages. h ps://doi.o g/10.1038/s41467-
024-52213-9
[57]
Michael Waw zoniak, Gianluca Mo o, Rod igo B uno,
Ana Klimo ic, and Gus a o Alonso. 2024. O - he-shel
Da a Analy ics on Se e less. In P oceedings o he
14 h Con e ence on Inno a i e Da a Sys ems Resea ch,
CIDR 2024. CIDR, Chaminade, USA, 10 pages. h ps:
//doi.o g/10.3929/e hz-b-000646171
[58]
Michal Waw zoniak, Ingo Mülle , Gus a o Alonso, and
Rod igo B uno. 2021. Boxe : Da a Analy ics on
Ne wo k-enabled Se e less Pla o ms. In 11 h Con-
e ence on Inno a i e Da a Sys ems Resea ch, CIDR
2021. CIDR, Chaminade, USA, 8 pages. h ps://doi.
o g/10.3929/e hz-b-000456492
[59]
Xingda Wei, Fangming Lu, Tianxia Wang, Jinyu Gu,
Yuhan Yang, Rong Chen, and Haibo Chen. 2023. No P o-
isioned Concu ency: Fas RDMA-codesigned Remo e
Fo k o Se e less Compu ing. In 17 h USENIX Sym-
posium on Ope a ing Sys ems Design and Implemen-
a ion (OSDI 23). USENIX Associa ion, Bos on, MA,
497–517. h ps://www.usenix.o g/con e ence/osdi23/
p esen a ion/wei- dma
[60]
Sebas ian We ne and S e an Tai. 2024. A e e ence
a chi ec u e o se e less big da a p ocessing. Fu u e
Gene a ion Compu e Sys ems 155 (2024), 179–192.
h ps://doi.o g/10.1016/j. u u e.2024.01.029
[61]
Zhao ui Wu, Yuhui Deng, Yi Zhou, Jie Li, Shujie Pang,
and Xiao Qin. 2024. FaaSBa ch: Boos ing Se e less
E iciency Wi h In-Con aine Pa allelism and Resou ce
Mul iplexing . IEEE T ans. Compu . 73, 04 (Ap il 2024),
1071–1085. h ps://doi.o g/10.1109/TC.2024.3352834
[62]
Minchen Yu, Tingjia Cao, Wei Wang, and Ruichuan
Chen. 2023. Following he Da a, No he Func ion: Re-
hinking Func ion O ches a ion in Se e less Compu -
ing. In 20 h USENIX Symposium on Ne wo ked Sys ems
Design and Implemen a ion (NSDI 23). USENIX Asso-
cia ion, Bos on, MA, 1489–1504. h ps://www.usenix.
o g/con e ence/nsdi23/p esen a ion/yu
[63]
Hong Zhang, Yupeng Tang, Anu ag Khandelwal, Jin-
g ong Chen, and Ion S oica. 2021. Cae us: NIM-
USENIX Associa ion 2025 USENIX Annual Technical Con e ence 55
BLE Task Scheduling o Se e less Analy ics. In 18 h
USENIX Symposium on Ne wo ked Sys ems Design
and Implemen a ion (NSDI 21). USENIX Associa-
ion, Bos on, MA, 653–669. h ps://www.usenix.o g/
con e ence/nsdi21/p esen a ion/zhang-hong
56 2025 USENIX Annual Technical Con e ence USENIX Associa ion