Engineering Serverless Workflow Applications in Federated FaaS.pdf

Engineering Serverless Workflow Applications in
Federated FaaS
Sashko Ristov

Agenda
• About me
• Federated FaaS challenges
• Research questions
• Conclusion
S. Ristov: Engineering Serverless Workflow Applications in Federated FaaS | TEWI-Kolloquium, Uni-Klu | 2022-11-29 1/63

Positions at UIBK
• (Jun. 2022 - present) Assistant Professor
• Quality Engineering Group
• (Apr. 2016 - Apr. 2022) Postdoctoral University Assistant
• Distributed and Parallel Systems Group

Recent Publications in 2022
• Journals / Magazines ×4
• S. Ristov, D. Kimovski, T. Fahringer, "FaaScinating Resilience for Serverless Function
Choreographies in Federated Clouds," in IEEE Trans. on Network and Service Management, 2022.
• S. Ristov, S. Pedratscher, T. Fahringer, “xAFCL: Run scalable function choreographies across
multiple FaaS systems,” in IEEE Transactions on Services Computing, pp. 1–1, 2022
• S. Pedratscher, S. Ristov, T. Fahringer, "M2FaaS: Transparent and Fault Tolerant FaaSification of
Node.js Monolith Code Blocks", in Elsevier Fut. Gen. Comp. Sys., vol. 135, pp. 57–71, 2022.
• D. Kimovski, S. Ristov, and R. Prodan, “Decentralized machine learning for intelligent health-care
systems on the computing continuum,” Computer, vol. 55, no. 10, pp. 55–65, 2022.
• Top Conferences ×3
• S. Ristov, M. Hautz, C. Hollaus, R. Prodan. “Simulate Serverless Workflows and Their Twins and
Siblings in Federated FaaS,” ACM Symp. on Cloud Computing (SoCC’22), San Francisco, Nov. 2022
• S. Ristov and P. Gritsch. "FaaSt: Optimize makespan of serverless workflows in federated
commercial FaaS," (IEEE CLUSTER ’22). Heidelberg, Germany, 182–194, 2022.
• S. Ristov, S. Brandacher, M. Felderer, R. Breu, “GoDeploy: Portable Deployment of Serverless
Functions in Federated FaaS”, IEEE Cloud Summit 2022, Fairfax, Virginia, USA, Oct. 2022
• Workshops Papers on Top-tier conferences×2
• S. Ristov, C. Hollaus, and M. Hautz. "Colder Than the Warm Start and Warmer Than the Cold
Start! Experience the Spawn Start in FaaS Providers," in ApPLIED ’22). Salerno, Italy, 35–39, 2022.
• M. Gusev, S. Ristov, et al. „CardioHPC: Serverless Approaches for Real-Time Heart Monitoring of
Thousands of Patients“, WORKS’22 Workshop (Supercomputing) , Dallas, USA, Nov. 2022

Recent Collaborations with Uni-Klu
• IEEE Transactions / Magazines
• S. Ristov, D. Kimovski, T. Fahringer, "FaaScinating Resilience for Serverless Function
Choreographies in Federated Clouds," in IEEE Trans. on Network and Service Management, 2022.
• R. Matha, S. Ristov, T. Fahringer, and R. Prodan, “Simplified workflow simulation on clouds
based on computation and communication noisiness,” in IEEE Transactions on Parallel and
Distributed Systems, vol. 31, no. 7, pp. 1559–1574, 2020
• D. Kimovski, S. Ristov, and R. Prodan, “Decentralized machine learning for intelligent
health-care systems on the computing continuum,” Computer, vol. 55, no. 10, pp. 55–65, 2022.
• IEEE Magazines
• S. Ristov, M. Hautz, C. Hollaus, R. Prodan. “Simulate Serverless Workflows and Their Twins and
Siblings in Federated FaaS,” ACM Symp. on Cloud Computing (SoCC’22), San Francisco, Nov. 2022
• Top tier conferences
• M. Gusev, S. Ristov, A. Amza, A. Hohenegger, R. Prodan, D. Mileski, P. Gushev, G. Temelkov
"CardioHPC: Serverless Approaches for Real-Time Heart Monitoring of Thousands of Patients“,
WORKS’22 Workshop (Supercomputing) , Dallas, USA, Nov. 2022
• Project Collaboration - CardioHPC
• European High-Performance Computing Joint Undertaking, under grant agreement 951745
(FF4EuroHPC project and CardioHPC experiment)

Agenda
• About me
• Conclusion

Federated FaaS Heterogeneity
IBM
Google AWS
Alibaba

Free at last?
We can compute everywhere!

But, Are We Really Free?
Hurdles in all directions

Federated FaaS Challenges
Such a heterogeneous environment. How to:
• Deploy?
• Model?
• Orchestrate?
• Optimize?
• Run?

• Deploy? Ristov et al. 22 (GoDeploy - IEEE Cloud Summit BEST PAPER)
• Model?
• Orchestrate?
• Optimize?
• Run?

• Model? Ristov et al. (SimLess - ACM SoCC’22)
• Orchestrate?
• Optimize?
• Run?

• Orchestrate? Ristov et al. 2021 (AFCL - Elsev. FGCS)
• Optimize?
• Run?

• Optimize? Ristov & Gritsch. 2022 (FaaSt - IEEE Cluster ’22)
• Run?

• Run? Ristov et al. 2021 (xAFCL - IEEE TSC),

• Run? Ristov et al. 2021 (xAFCL - IEEE TSC), 2022 (rAFCL - IEEE TNSM)

Problem Statement 1: Vendor Lock-in
IBM
Google AWS
Alibaba

AWS

AWS
AWS Step Functions

AWS
AWS Step Functions
S3 Storage

Freedom?!

Freedom?!
Limited freedom (At Home) - The current state

Freedom?!
We want full freedom

Problem Statement 2: Limited Concurrency
xAFCL
UIBK
AWS Frankfurt
700 ms

xAFCL
UIBK
AWS Frankfurt
1,400 ms
700 ms

xAFCL
UIBK
AWS Frankfurt
1,400 ms
700 ms
AWS North Virginia
1,000 ms

xAFCL
UIBK
AWS Frankfurt
1,400 ms
700 ms
AWS North Virginia
1,000 ms
AWS Tokyo
1,200 ms

Google North Virginia
300 ms
xAFCL
UIBK
AWS Frankfurt
1,400 ms
700 ms
AWS North Virginia
1,000 ms
AWS Tokyo
1,200 ms

Agenda
• About me
• Conclusion

Research Goal RG1 - Abstraction
Research goal RG1 - Abstraction
Can we abstract the federated FaaS and develop distributed applications as they
run on a single computer?

Research question RQ1: Portability

• How to abstract heterogeneous resources and hide them from developers?

• How to abstract heterogeneous resources and hide them from developers?
• Can we easily migrate and run serverless functions in federated clouds?

Portability Levels

SOTA (b) vs xAFCL (c) Portable Approach (per function)
AWS
Lambda
IBM
Cloud
Functions
Google
Cloud
Functions
AWS
Step
Functions
IBM
Composer
Google
Cloud
Composer
IBM FC
Developer
AWS FC
Developer
Google FC
Developer
AWS
Lambda
IBM
Cloud
Functions
FaaS
systems
AFCL
AFCL FC
Developer
FC
system
a) c)
b)
AWS
Lambda
IBM
Cloud
Functions
Google
Cloud
Functions
xAFCL
xAFCL FC
Developer
…
b) S. Ristov, S. Pedratscher, T. Fahringer: AFCL: An Abstract Function Choreography Language for serverless
workflow specification. Future Generation Computer Systems 114: 368-382 (2021)
c) S. Ristov, S. Pedratscher and T. Fahringer, "xAFCL: Run Scalable Function Choreographies Across Multiple
FaaS Systems," in IEEE Transactions on Services Computing, 2021. doi: 10.1109/TSC.2021.3128137.

Resilient AFCL (per function) - rAFCL
AWS
Lambda
IBM
Cloud
Functions
Google
Cloud
Functions
FC
systems
FaaS
systems
AWS
Step
Functions
IBM
Composer
Google
Cloud
Composer
IBM FC
Developer
AWS FC
Developer
Google FC
Developer
AWS
Lambda
IBM
Cloud
Functions
Google
Cloud
Functions
FaaS
systems
ftAFCL
ftAFCL FC
Developer
FC
system
a) b)
x x
x
x
x x x
x x
x
x
x
x
S. Ristov, D. Kimovski, T. Fahringer, "FaaScinating Resilience for Serverless Function Choreographies in
Federated Clouds," in IEEE Transactions on Network and Service Management, 2022.

rAFCL Architecture
rAFCL
Environment
FCDev
rScheduler
resilience
model
AFCL Parser
fRepository
Log
AF
AFCL
alternative
rAFCL
Multi-FaaS
rEE
Invoke
Monitor
...
AWS Monitor
IBM Monitor
FaaS
Invoker
...

Example rAFCL Scheduler
AP0 AP1 AP2 AP3 Omitted
70
80
90
100 99.89 99.11
99.01 97.8
95.45
86.67
69.95
aj = 99.5%
99.89 99.99 99.9
93.35
82.88
87.89
75.77
99.89 98.75
Availability
(%)

AFCL Flexible Resources and Fault Tolerance
Flexible function deployment and alternative functions
properties:
- name: "resource"
value: "ARN..." #Run on AWS
# value: "https://europe-west6-xxx.cloudfunctions.net..." #Run on Google Zurich
constraints:
- name: "FT-Retries" #Retry 2 times alternatives of this function. Ignore provider's setup.
value: "2"
- name: "FT-AltPlan-0" #If AWS fails, retry two tasks on IBM Japan and London (duplicate)
value: "0.9879;https://jp-tok...;https://eu-gb...;"
- name: "FT-AltPlan-1" #If both tasks (jp and gb) fail, run alternatives on Fra. and Dallas
value: "0.9808;https://eu-de...;https://us-south...;"
S. Ristov, D. Kimovski, T. Fahringer, "FaaScinating Resilience for Serverless Function Choreographies in
Federated Clouds," in IEEE Trans. on Network and Service Management, 2022.

FaaS Concurrency Overhead Model
0 1 2 3 4 5 6 7
0
500
1,000
outliers
load burst release
Time (s)
Concurrent
functions
S. Ristov, S. Pedratscher and T. Fahringer, "xAFCL: Run Scalable Function Choreographies Across Multiple FaaS
Systems," in IEEE Transactions on Services Computing, 2021. doi: 10.1109/TSC.2021.3128137.

AFCL Example
- function:
properties:
- name: "resource"
value: "arn:aws:lambda:xxxx" # (AWS)
# value: "xxxx.functions.appdomain.cloud/xxxx" # (IBM)

AFCL Simple Example
S. Ristov, S. Pedratscher, T. Fahringer: AFCL: An Abstract Function Choreography Language for serverless
workflow specification. Future Generation Computer Systems 114: 368-382 (2021)

FCEditor: Complex Workflows
FCEditor: almost two years before AWS

Some Experiment Setup Example with xAFCL

What about functions with BaaS
f service dynamic URL
Region 1
Region 2

RQ2: Interoperability
Research question RQ2: Interoperability
• How to abstract cloud services within a serverless function, regardless
where it is hosted?

where it is hosted?
• Use Google Cloud Storage instead of AWS S3

where it is hosted?
• Use Google Cloud Storage instead of AWS S3
• Use Google Vision instead of AWS Rekognition, regardless where the image is
located (AWS S3 or Google Cloud Storage)

SOTA Approach: 4 Deployments for 2 Cloud Providers
access
deploy
access
deploy
deploy
access
deploy access

2 Chained Services = 8 Deployments
4 Lambda and 4 Google functions that:
• use AWS S3 and AWS Rekognition
• use AWS S3 and Google Vision
• use GCS and AWS Rekognition
• use GCS and Google Vision

Let’s start with some challenging game
Exchange places of black and white knights with minimum number of moves!

Let’s start with some challenging game
Exchange places of black and white knights with minimum number of moves!
40 moves are needed!

Let’s use Homothety
2
1
4
10
8
6
5 9
7
3

2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3
2
1 4 10
8
6
5
9
7
3

Long-term Goal
• Can we hide cloud providers, services, and resources?
• Can we abstract Federated FaaS heterogeneity as the Von Neumann’s
architecture?
• The current state-of-the-art approaches are towards "code once, run
everywhere"
• I plan to go even further: Can we code once, run everywhere, but also with
everything (every cloud service)?
• even during runtime, without code rewriting and redeployment!

Long-term Goal
• Can we hide cloud providers, services, and resources?
• Can we abstract Federated FaaS heterogeneity as the Von Neumann’s
architecture?
• The current state-of-the-art approaches are towards "code once, start
everywhere"
• I plan to go even further: Can we code once, run everywhere, but also with
everything (every cloud service)?
• even during runtime, without code rewriting and redeployment!

Homothety: Federated FaaS to Von Neumann
OS
Computer
App1 App2
APPs
NET
FS

fService Programming Model: fStorage
access
deploy
access
fStorage
{"source": "https://bucket.s3.region.amazonaws.../..."}
{"source": "https://storage.google.cloud.com/.../..."}
Supported languages: Java, Python, Go.

fService Architecture
fStorage.copy(bucketUrlAWS, "/tmp/destFile");
fStorage.copy("/tmp/sourceFile", bucketUrlGoogle);
fStorage.copy(bucketUrlGoogle, bucketUrlAWS);
fRecognition.detectFaces("path/to/image.jgp")
Unpublished work yet.

fService Chained Approach: fRecognition + fStorage
AWS Rekognition
Google Storage
AWS S3
Google Vision
The red arrows show the fastest runtime.
Supported languages: Java.

Results with fStorage + fScheduler
Europe
US
S3 O
AWS N
GCF F
GCS F
xAFCL UIBK
AWS O
S3 N
Asia
GCS S
GCF S
input files are scattered faster is to use GCF F and even AWS
N, although there is no input data.

Research Goal RG2 - Runtime System Optimization
Research goal RG2 - Runtime System Optimization
Can we create a "no-learning" unified system model and optimize serverless
applications in federated clouds?

Research question RQ3: FaaS system model

• Which parameters are important? Which features of functions, runtime
systems, and the underlying resources should be formalized to estimate
functions’ behavior?

• Which parameters are important? Which features of functions, runtime
systems, and the underlying resources should be formalized to estimate
functions’ behavior?
• Can we determine the behavior of not even deployed function in AWS
Frankfurt based on the same function in AWS Tokyo?

SOTA for RQ3
• Running benchmark applications
• Too expensive for the computing continuum
• ML approaches
• Maybe cheaper, but still too expensive
• Requires time series

Method: Twins and Siblings
Tokyo
128MB
SIBLING
512MB
TWIN
128MB
Frankfurt

Innovative Approach: RTT = ET + O
ET: Exec. time
O: Overhead
RTT: Round
Trip Time SO: Session
AO: Authentication
FO: FaaS
CO: Concurrency
NO: Network
O: Overhead

FC Simulation Model
RTT = O + ET (1)
O = SO + NO + AO + FO + CO (2)
AO =
(
cr + 3 · NO, 2-way authentication;
cr + 2 · NO, 1-way authentication.
(3)
CO = (k − 1) · d (4)

FC Simulation Model: Twins and Siblings
RTTS = (RTT-O) x s(m)/m + O
RTTT = RTT - O + OT
128MB
SIBLING
512MB
TWIN
128MB
xAFCL
SimLess
run
Tokyo
Frankfurt
O
ET
RTT
OS = O
ETS = ET x s(4)/4
ETT = ET
O
ET = RTT - O
OT > O
simulate
twins
simulate
siblings
OT

Cloud Regions for Learning and Validation
Goal Cont. AWS IBM Google
3*Learn EU1 AF IF GB
US1 AV IW GV
AP1 AT IT GT
3*Validate EU2 AL IL GL
US2 AC ID GI
AP2 AS IS GH
AF
MC10
MC1000
w Auth.
no-op
w Auth.
no-op
w/o Auth.
CloudPing
gcPing
xAFCL
NO
xAFCL
FO, SO,
xAFCL
AO, cr CO
SimLess
RTTT
O
RTT, M
~
AT AV
twins
siblings
RTTS
d
~
M

Parameter Setup
Region SO NO cr AO FO O e
O δO[%]
AF 3*550 30 3*76 166 3*30 226 189 19.6
AV 140 496 666 676 1.48
AT 267 877 1174 1200 2.17
IF 3*152 30 3*76 136 3*57 223 204 9.31
IW 140 356 553 577 4.16
IT 267 610 934 930 0.43
GB 3*112 59 3*– 3*– 3*23 82 95 13.7
GV 131 154 131 17.6
GT 275 298 307 2.93

Overhead Accuracy
AF AV AT AL AC AS IF IW IT IL ID IS GB GV GT GL GI GH
Region
0
200
400
600
800
1000
1200
1400
O
(ms)
0%
5%
10%
15%
20%
25%
30%
δ
O
O
f
O
δO

Evaluated Tools

Low Concurrency MC10 and BWA20
AF AL AV AC IF IL IW ID GB GL GV GI
Region
0
5
10
15
20
25
30
RTT
T
/
M
(s)
0%
5%
10%
15%
20%
25%
±
RTTT g
RTTT M f
M ±RTT ±M
Simulated versus measured RTT and
makespan of monteCarlo twins, and
their inaccuracy.
S
p
l
i
t
I
n
d
e
x
A
L
N
1
A
L
N
2
S
a
m
p
e
M
S
p
l
i
t
I
n
d
e
x
A
L
N
1
A
L
N
2
S
a
m
p
e
M
Functions and total Makespan
0
10
20
30
40
RTT
T
/
M
(s)
0%
5%
10%
15%
20%
25%
30%
±
RTT
AV from AF AT from AF
RTTT, M
g
RTTT, f
M
±RTT
Simulated versus measured RTT of
BWA20 twins and their inaccuracy

High Concurrency MC1,000 and BWA2,500
0 200 400 600 800 1000
functions
16
17
18
19
20
ct
(s)
5
6
7
±
ct
(%)
f
ct ct ±ct
Simulated versus measured completion
time of MC1,000 functions on AF and their
inaccuracy.
0 500 1000 1500 2000 2500
functions
10
20
30
40
ct
(s)
0
10
20
30
40
±
ct
(%)
f
ct ct ±ct
Simulated vs. measured completion
time ct of BWA2,500 functions on AF and
their inaccuracy.

Research Question RQ4
Research question RQ4
Can serverless workflow applications benefit from executing individual functions
across different FaaS systems in the federated FaaS?

SOTA1 for RQ4: Serverful Scheduling
• Scheduling of serverful workflow applications that run on VMs

SOTA1 for RQ4: Serverful Scheduling
• Scheduling of serverful workflow applications that run on VMs
• Not applicable for FaaS
• Resources are selected during deployment, not during runtime
• Iterate over functions, then look for possible function deployments, rather
than VMs.

SOTA2 for RQ4: FaaS Scheduling
• Many container-based algorithms schedule individual functions on
specific container (executor)

SOTA2 for RQ4: FaaS Scheduling
• Many container-based algorithms schedule individual functions on
specific container (executor)
• within a single region on open source FaaS platforms (e.g. OpenWhisk)

SOTA3 for RQ4: FC Distribution in Federated FaaS
federated clouds
single region
pFor
N = 1000
f2 f3
f4
f1
pFor
N = 600
f2 f3
f4
f1
pFor
N = 400
f2 f3
f4
f1
federated clouds
pFor
N = 500
f2 f3
f4
f1
pFor
N = 400
f2 f3
f4
f1
pFor
N = 100
f2 f3
f4
f1
a) per function
single region
b) per FC
federated clouds
c) per function
federated clouds
• a) only f1 and f4 can run 1,000 instances concurrently, but not f2 and f3
• b) cannot exploit f1 runs faster on blue, while f4 on the left (orange).
• c) schedules f1 on blue, f4 on orange.

FaaS Deployment Model
• Function type - WHAT

• high-level of abstraction
• data inputs and outputs, semantics
• FC dependencies (data-flow,
control-flow)
• YAML, JSON
pFor
N = 1000
f2 f3
f4
f1

• high-level of abstraction
• data inputs and outputs, semantics
• FC dependencies (data-flow,
control-flow)
• YAML, JSON
• Function deployment - HOW and
WHERE
• real application / code
• e.g. python, for AWS, algorithm
• code + resources
• e.g. region, memory, handler function
pFor
N = 1000
f2 f3
f4
f1
pFor
N = 500
f2 f3
f4
f1
pFor
N = 400
f2 f3
f4
f1
pFor
N = 100
f2 f3
f4
f1

FaaS Abstract Resource Model
• R regions
• cl - concurrency limit
• a user can run cl functions, regardless of assigned memory
• Abstract resource ar

FC Scheduler
r1
fd11
fd21
r2
fd22
fd12
FT1 (Split)
FT2 (Index)
FT f
f1
f2
f3
f4
f5
Split3
Split2
Split1
Index1
Index2
FDs & Regions AR
AR1
ar12
ar11
ar14
f1-> fd11
f2-> fd11
f5-> fd21
ar13
AR2
ar21
ar22
f3-> fd12
f4-> fd22
ar23

Algorithm 1: FC scheduling algorithm
Output: schedFC = {(fi, sched(fi, fdij, arrk))|∀fi ∈ F}; // FC schedule
1 Function (FC, FD, AR):
2 Rank ← B-Rank(F); // order the tasks according to B-rank
3 schedFC ← ∅ ; // Initialize FC schedule
4 for i ← 1 to N do // Iterate over each fi
5 Tmin ← ∞; fdmin ← 0; ct(fi) ← ∞; // Initialization
6 for j ← 1 to Mi do // Iterate over each fdij of fi
7 r ← reg(fdij); // find the region of fdij
8 for k ← 1 to clr do // Iterate over each arrk of rr
9 Tj ← EST(fdij, arrk); // Compute EST
10 ct(fi) ← Tj + RTTij; // compute completion time
11 if ct(fi) < Tmin then
12 Tmin ← ct(fi) ; // Save min. completion time
13 fdmin ← j; // Save the fdij for fi
14 armin ← k; // Save the arrk for fdij
15 end
16 end
17 end
18 schedFC ← schedFC ∪ sched(fi, fdmin, armin); // schedule fi
19 end
20 return

Agenda
• About me
• Conclusion

Summary - Scenarios

Summary - Approaches

Summary - SOTA

Further Investigation in 2022/2023
Storage Interoperability + Data-aware Scheduler
• copy(srcURL, destURL) (Java, Python, Go1
)
• Dynamically select storage during runtime, without redeployment
DeployLess Model and Scheduler
Functions Size AF AN AT
Split, Index, AlnR1, AlnR2 0.17 MB 1.8 s 2.7 s 4.5 s
Sampe, Merge, Sort 1.7 MB 2.0 s 3.1 s 5.6 s
1
https://github.com/FaaSTools/GoStorage

QUESTIONS?
Email: sashko.ristov@uibk.ac.at

Engineering Serverless Workflow Applications in Federated FaaS.pdf

Recommended

Recommended

More Related Content

Similar to Engineering Serverless Workflow Applications in Federated FaaS.pdf

Similar to Engineering Serverless Workflow Applications in Federated FaaS.pdf (20)

More from Förderverein Technische Fakultät

More from Förderverein Technische Fakultät (20)

Recently uploaded

Recently uploaded (20)

Engineering Serverless Workflow Applications in Federated FaaS.pdf