SlideShare a Scribd company logo
1 of 66
Download to read offline
Traffic Optimization in Data Center and
Software-Defined Programmable Networks
GERMAN SVIRIDOV
Doctoral Program in Electrical, Electronics and Communications
Engineering
(XXXIII cycle)
Supervisors:
Prof. Paolo Giaccone
Prof. Andrea Bianco
2
Stateful Software-Defined
Networking: enabling replicated
network applications in
programmable data planes
Flow scheduling in data center
networks: optimizing flow
performance by minimizing flow
completion time
Cloud gaming and game engines:
Quality of Experience improvement in
cloud gaming
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
Blockchains for vehicular
applications
2
3
Stateful Software-Defined
Networking: enabling replicated
network applications in
programmable data planes
Flow scheduling in data center
networks: optimizing flow
performance by minimizing flow
completion time
Cloud gaming and game engines:
Quality of Experience improvement in
cloud gaming
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
Blockchains for vehicular
applications
3
Enabling replicated network applications in
programmable data planes
03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 4
Classic Software Defined Networks (SDN)
Control
Plane
Data
plane
Physical separation of control and data plane
A single entity (controller) for the control plane
Data plane operations are now centrally managed
Key enabler for novel applications
03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 5
Issues with classic SDN
Control
Plane
Data
plane
• Physical latency between switches
and the controller
• Software processing latency
Added latency
• Per flow processing only
• Support of legacy protocols only
Inflexible
commercial
hardware
• Local in-switch processing
• Simplified switch-local control logic
for latency-sensitive application
Need to take a
step back
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 6
Programmable data planes
• User-defined packet parsing
• Support for virtually any data protocol
• Modification to any packet field at line rate
Programmable
parsers
• Programmable match-action rules
• Access to persistent states 𝒔𝒊
• Packet recirculation/cloning
Programmable
packet
processing
𝑠𝑖
03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 7
Enabling new applications and
accelerating old ones
𝑠𝑖
Financial applications
Machine learning
MQTT brokers
DBMS Acceleration
Content caching
Overlay protocols
Load balancers
Packet scheduling
Network telemetry
Firewall/IDS
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 8
Limits of data plane-assisted network
applications
𝑠1
𝑠2
𝑠3
𝑠4
•Network applications are local to a single switch
•Independent behavior in the network
Locality constraints
•DDoS detection
•Distributed application/link-aware load balancers
•Network-wide rate limiters
•...
Precluded applications
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 9
Limits of data plane-assisted network
applications
𝑠1
𝑠2
𝑠3
𝑠4
•Network applications are local to a single switch
•Independent behavior in the network
Locality constraints
•DDoS detection
•Distributed application/link-aware load balancers
•Network-wide rate limiters
•...
Precluded applications
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 10
How to implement logically centralized network-
wide network applications?
Logically centralized network applications
𝑠1
𝑠2
𝑠3
𝑠4
Unique network application
Centralized
state with
global view
of the
network
Traffic
reroute
Data
overhead
Bandwidth
constraints
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 11
𝑠1
𝑠2
𝑠3
𝑠4
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
State replication in programmable data
planes
12
Replicated network application
Centralized
state with
global view
of the
network
No traffic
reroute
No data
overhead
Sync
overhead
𝑠1
𝑠3
𝑠4
𝑠2
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
State replication in programmable data
planes
13
Replicated network application
Centralized
state with
global view
of the
network
No traffic
reroute
No data
overhead
Sync
overhead
A recipe for state replication in
programmable data planes
LODGE
Suitable
replication
algorithm
Need to convey
state-updates to
other replicas
Need to define a
suitable
transport
protocol
Need to be able
to generate new
packets
𝑠2
𝑠3
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 14
A recipe for state replication in
programmable data planes
LODGE
Eventual
consistency
Need to convey
state-updates to
other replicas
Need to define a
suitable
transport
protocol
Need to be able
to generate new
packets
𝑠2
𝑠3
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 15
A recipe for state replication in
programmable data planes
LODGE
Eventual
consistency
Controller-
constructed
distribution tree
Need to define a
suitable
transport
protocol
Need to be able
to generate new
packets
𝑠2
𝑠3
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 16
A recipe for state replication in
programmable data planes
LODGE
Eventual
consistency
Controller-
constructed
distribution tree
Tailored
P4/OPP-enabled
transport
protocol
Need to be able
to generate new
packets
𝑠2
𝑠3
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 17
A recipe for state replication in
programmable data planes
LODGE
Eventual
consistency
Controller-
constructed
distribution tree
Tailored
P4/OPP-enabled
transport
protocol
Traffic-triggered
packet cloning
and
recirculation
𝑠2
𝑠3
𝑠2
1
𝑠3
1
D
S
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 18
A recipe for state replication in
programmable data planes
LODGE
Eventual
consistency
Controller-
constructed
distribution tree
Tailored
P4/OPP-enabled
transport
protocol
Traffic-triggered
packet cloning
and
recirculation
𝑠2
𝑠3
𝑠2
1
𝑠3
1
D
S
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 19
LODGE-Enabled Applications
server cluster 1
server cluster 3
server cluster 2
server cluster 4
ASN1 ASN2
ASN3
ASN4
R2
R1
R3
R4
SW1
SW2
SW3
SW4
Realistic scenario
Emulated
testbed
LODGE
implementation
•WAN Autonomous System
(AS) with 4 neighbours
•Incoming traffic monitoring
and reaction
•V1 Model with BMV
•Emulated network
•P4 14/16
•Open Packet Processor (OPP)
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 20
LODGE-Enabled Applications
• Distributed DDoS from ASN#
• LODGE logic embedded at
SW1 and SW3
• Simultaneous detection of the
DDoS at all switches without
traffic rerouting
• Distributed rate limiting
• LODGE application embedded
at border routers
• Simultaneous drop of the rate
for both flows
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 21
server cluster 1
server cluster 3
server cluster 2
server cluster 4
ASN1 ASN2
ASN3
ASN4
R2
R1
R3
R4
SW1
SW2
SW3
SW4
LODGE-Enabled Applications
• Distributed DDoS from ASN#
• LODGE logic embedded at
SW1 and SW3
• Simultaneous detection of the
DDoS at all switches without
traffic rerouting
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
Why SW1 & SW3 or R# in
particular?
22
server cluster 1
server cluster 3
server cluster 2
server cluster 4
ASN1 ASN2
ASN3
ASN4
R2
R1
R3
R4
SW1
SW2
SW3
SW4
• Distributed rate limiting
• LODGE application embedded
at border routers
• Simultaneous drop of the rate
for both flows
Where to place the state replicas?
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 23
State replicas placement: data vs
synchronization traffic
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 24
Data traffic
Sync traffic
𝑠1
1
and 𝑠1
2
: replicas of the same state 𝑠1
X2
X1
s s
s s
- More data traffic overhead
- Less sync traffic overhead
- Less data traffic overhead
- More sync traffic overhead
Number of state replicas
𝑠1
1
𝑠1
2
𝑠1
State replicas placement: data vs
synchronization traffic
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 25
Number of state replicas • Data vs synchronization traffic trade-off?
• How many replicas to use?
• Where to place the replicas?
Data traffic
Sync traffic
𝑠1
1
and 𝑠1
2
: replicas of the same state 𝑠1
X2
X1
s s
s s
𝑠1
1
𝑠1
2
𝑠1
- More data traffic overhead
- Less sync traffic overhead
- Less data traffic overhead
- More sync traffic overhead
Optimal state replication in stateful data
planes
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 26
MILP formulation for optimal state replication
Input
• Network topology
• Traffic matrix
• Network applications
• Per-flow application
requirements
• Amount of sync traffic ෡
𝜆𝑠
Constrained by
• Bandwidth availability
• Switch resources
availability
• Traffic matrix
Objective
• Minimize maximum
congestion
• Minimize total traffic
Output
• Number of replicas
• Replicas placement
• Data traffic routing
• Sync traffic routing
Optimal state replication in stateful data
planes
Difficult to apply to big topologies!
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 27
MILP formulation for optimal state replication
Input
• Network topology
• Traffic matrix
• Network applications
• Per-flow application
requirements
• Amount of sync traffic ෡
𝜆𝑠
Constrained by
• Bandwidth availability
• Switch resources
availability
• Traffic matrix
Objective
• Minimize maximum
congestion
• Minimize total traffic
Output
• Number of replicas
• Replicas placement
• Data traffic routing
• Sync traffic routing
Optimal state replication –
approximated solution
28
Algorithm for replica placement
• Based on betweenness centrality
• Solution perturbation
Optimal number of replicas
• Analytical expression
• Asymptotic optimal for Manhattan topology
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 28
How to put everything together?
Network application
abstracion
Application
decompositi
on
Consistency
constraints
LODGE dataplane
implementation
P4 OPP
Optimal state
placement
Number of
states
State
position
Traffic routing
Data traffic
Synchroniza-
tion traffic
How to define distributed
network applications?
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 29
LOADER: An abstraction for replicated network
applications
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 30
31
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 31
The goals of LOADER
Define an abstraction model for replicated
network applications
Enable efficient translation to basic network
primitives
Combine efforts of optimal state replication
and LODGE into a unified framework
32
Programming
abstraction for
network applications:
•Exposure of
replicated states to
the programmer
•Management of state
inconsistency
•Definition of 𝜆𝑠
Compiler
•Automatic policy
decomposition and
translation to target-
specific language
Embedder
•Automatic state
embedding
•Based on heuristic for
optimal state
replication
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 32
33
Programming
abstraction for
network applications:
•Exposure of
replicated states to
the programmer
•Management of state
inconsistency
•Definition of 𝜆𝑠
Compiler
•Automatic policy
decomposition and
translation to target-
specific language
Embedder
•Automatic state
embedding
•Based on heuristic for
optimal state
replication
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 33
34
Programming
abstraction for
network applications:
•Exposure of
replicated states to
the programmer
•Management of state
inconsistency
•Definition of 𝜆𝑠
Compiler
•Automatic policy
decomposition and
translation to target-
specific language
Embedder
•Automatic state
placement
•Based on heuristic for
optimal state
replication
𝑠1
𝑠2
𝑠3
𝑠4
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 34
35
Programming
abstraction for
network applications:
•Exposure of
replicated states to
the programmer
•Management of state
inconsistency
•Definition of 𝜆𝑠
Compiler
•Automatic policy
decomposition and
translation to target-
specific language
Embedder
•Automatic state
placement
•Based on heuristic for
optimal state
replication
𝑠1
𝑠2
𝑠3
𝑠4
𝑠2
1
𝑠3
1
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 35
LOADER
Network application
abstracion
Application
decompositi
on
Consistency
constraints
LODGE dataplane
implementation
P4 OPP
Optimal state
placement
Number of
states
State
position
Traffic routing
Data traffic
Synchroniza-
tion traffic
[1] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LODGE: LOcal Decisions on
Global statEs in programmable data planes, in: IEEE NetSoft, 2018
[2] Muqaddas, Abubakar Sidique; Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Optimal state replication in stateful dataplanes, in:
IEEE JSAC, 2020
[3] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LOcAl DEcisions on Replicated
States (LOADER) in programmable data planes: programming abstraction and experimental evaluation, in: Computer Networks 2021
[1] [2]
[3]
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 36
37
Stateful Software-Defined
Networking: enabling replicated
network applications in
programmable data planes
Flow scheduling in data center
networks: optimizing flow
performance by minimizing flow
completion time
Cloud gaming and game engines:
Quality of Experience improvement in
cloud gaming
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
Blockchains for vehicular
applications
37
Automating game QoE assessment in cloud gaming
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 38
The next generation of gaming:
Cloud Gaming
• Games run on a remote server
• Video stream of the game is sent to the users
Remote game
rendering
• Users buy the game and pay the subscription fee
• Play directly on a TV or phone
No need for
hardware
03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 39
The next generation of gaming:
Cloud Gaming
• Games run on a remote server
• Video stream of the game is sent to the users
Remote game
rendering
• Users buy the game and pay the subscription fee
• Play directly on a TV or phone
No need for
hardware
03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 40
Video streaming
(High BW, High Delay)
Online gaming
(Low BW, Low Delay)
Input lag
(High BW, Low Delay)
Not all games are made equal
Different impact
of latency
Different genres
Multiple stages
•Action/FPS
•Strategy, card,
interactive fiction
•Action stages
•Exploration/narration
stages
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 41
Not all games are made equal
Different impact
of latency
Different genres
Multiple stages
•Accuracy decrease in
FPS
•Smaller crop yield in
farming simulator?
•Action/FPS
•Strategy, card,
interactive fiction
•Action stages
•Exploration/narration
stages
Different network requirements
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 42
Not all games are made equal
Different impact
of latency
Different genres
Multiple stages
•Accuracy decrease in
FPS
•Smaller crop yield in
farming simulator?
•Action/FPS
•Strategy, card,
interactive fiction
•Action stages
•Exploration/narration
stages
Different network requirements
Need for fine-grained game QoE
assessment
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 43
Game Quality of Experience (QoE) assessment in
practice
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 44
Game
- 25 new game releases
per day on Steam
- Impossible to keep up
with new releases
WAN DC
379 565
1771
2964
4207
7049
9050
0
2000
4000
6000
8000
10000
2012 2013 2014 2015 2016 2017 2018
Number of games released on steam
Human subjects
- Difficult and
expensive to find
enough subjects
- No diversity
guarantee
Controlled environment
- Need to bring subjects to
the lab
- Difficult to achieve in the
wild
Outcome
- Noisy experimental data
- Coarse grained
information
- No repeatability
- Outdated information
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 45
Game
- 25 new game releases
per day on Steam
- Impossible to keep up
with new releases
WAN DC
379 565
1771
2964
4207
7049
9050
0
2000
4000
6000
8000
10000
2012 2013 2014 2015 2016 2017 2018
Number of games released on steam
Guilty
Human subjects
- Difficult and
expensive to find
enough subjects
- No diversity
guarantee
Controlled environment
- Need to bring subjects to
the lab
- Difficult to achieve in the
wild
Outcome
- Noisy experimental data
- Coarse grained
information
- No repeatability
- Outdated information
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 46
WAN DC
379 565
1771
2964
4207
7049
9050
0
2000
4000
6000
8000
10000
2012 2013 2014 2015 2016 2017 2018
Number of games released on steam
Game
- 25 new game releases
per day on Steam
- Impossible to keep up
with new releases
Bots
- Diversity and
experience can be
programmed
- Still expensive
Controlled environment
- Can be easily simulated
- Fine tuning is easily
achieved
Outcome
+ Deterministic results
+ Fine grained information
+ Repeatability
- Outdated information
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 47
WAN DC
379 565
1771
2964
4207
7049
9050
0
2000
4000
6000
8000
10000
2012 2013 2014 2015 2016 2017 2018
Number of games released on steam
Game
- 25 new game releases
per day on Steam
- Impossible to keep up
with new releases?
Bots
- Diversity and
experience can be
programmed
- Still expensive?
Controlled environment
- Can be easily simulated
- Fine tuning is easily
achieved
Outcome
- Deterministic results
- Fine grained information
- Repeatability
- Outdated information
AI?
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 48
Removing humans from the loop:
Putting the AI to work for QoE assessment
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 49
Deep reinforcement learning (DRL) for
video games
Learning to play from raw video pixels
Trial and error approach
No supervision Good adaptability
Active exploration
Reward shaping
Tunable behavior
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 50
Learning to play Atari
• Deep Q Networks
• Simple, yet powerful model for DRL
One
algorithm
• Seaquest, Beam Rider and Breakout
• Taken from the Atari game catalogue
Three
games
• (Super) human level achieved in less
than 5 hours of training
• Ready to be deployed in the testbed
Three AI
agents
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 51
Measuring the gaming performance
• Per-frame lag 𝑙
• Per-frame lag probability
𝑝𝑙𝑎𝑔
• Per-keystroke drop
probability 𝑝𝑑𝑟𝑜𝑝
Emulated
network
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 52
Measuring the gaming performance
• Per-frame lag 𝑙
• Per-frame lag probability
𝑝𝑙𝑎𝑔
• Per-keystroke drop
probability 𝑝𝑑𝑟𝑜𝑝
Emulated
network
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 53
Maximizing average game score
Different network requirements
Need for fine-grained game QoE
assessment
Without traffic
prioritization
With traffic
prioritization
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 54
Moving to more complex games
Moving to more complex games
55
Classic Doom
Similar algorithm for training
Different
maps
Different
game modes
Network
parameter
perturbation
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 55
Automating game QoE assessment
Cloud gaming
scenario
Emulated network
with perturbation
Artificial players trained with DRL
Atari Doom
[4] Sviridov, German; Beliar, Cedric; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Removing human players from the loop: AI-assisted
assessment of Gaming QoE, in: IEEE Infocom NI Workshop, 2020
[5] Sviridov, German; Beliar, Cedric; Simon, Gwendal; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Leveraging AI players for QoE estimation
in cloud gaming, in IEEE Infocom Demo Session, 2020
[5]
[4-5]
Per-game QoS mechanism which
maximizes average game QoE
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 56
Past works
57
02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
•Is centralized flow scheduling feasible in practice?
•Can we reach close to state of the art performance
while employing commodity switches for flow
scheduling?
Flow scheduling
in data center
networks
•Blockchain as an enabler for mobility verification
•How to manage trust for sensitive data?
•Which scale private blockchains can reach?
Blockchains for
vehicular
applications
[6-7]
[8]
[6] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Low-Complexity Flow Scheduling for Commodity Switches in Data Center Networks, in: IEEE Globecom,
2019
[7] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, To Sync or Not to Sync: Why Asynchronous Traffic Control Is Good Enough for Your Data Center, in: IEEE
Globecom, 2018
[8] Chiasserini, Carla Fabiana; Giaccone, Paolo; Malnati, Giovanni; Macagno, Michele; Sviridov, German, Blockchain-based mobility verification of connected cars,
in IEEE CCNC, 2020
57
Current work – (Cloud) Gaming
58
•Games are made of different stages
•Each stage requires different QoS
•Can we achieve fine-grained QoS control?
Game stage
classification
•Training time of a given game is typically
unknown
•Can we understand the training difficulty before
even training?
Estimating
training
complexity
•Game engines are monolithic
•Opposite of the cloud-based applications
•Can we make cloud gaming truly cloud-based?
Distributed
gaming
engines
Graphics
Front-end
A.I. Physics
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 58
References
59
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV
[1] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LODGE: LOcal Decisions on Global statEs in
programmable data planes, in: IEEE NetSoft, 2018
[2] Muqaddas, Abubakar Sidique; Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Optimal state replication in stateful dataplanes, in: IEEE JSAC, 2020
[3] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LOcAl DEcisions on Replicated States
(LOADER) in programmable data planes: programming abstraction and experimental evaluation, in: Computer Networks 2021
[4] Sviridov, German; Beliar, Cedric; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Removing human players from the loop: AI-assisted assessment of
Gaming QoE, in: IEEE Infocom NI Workshop, 2020
[5] Sviridov, German; Beliar, Cedric; Simon, Gwendal; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Leveraging AI players for QoE estimation in cloud
gaming, in IEEE Infocom Demo Session, 2020
[6] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Low-Complexity Flow Scheduling for Commodity Switches in Data Center Networks, in: IEEE
Globecom, 2019
[7] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, To Sync or Not to Sync: Why Asynchronous Traffic Control Is Good Enough for Your Data Center, in:
IEEE Globecom, 2018
[8] Chiasserini, Carla Fabiana; Giaccone, Paolo; Malnati, Giovanni; Macagno, Michele; Sviridov, German, Blockchain-based mobility verification of
connected cars, in IEEE CCNC, 2020
59
Backup slides
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 60
Centralized explicit rate assignment
Taking inspiration from Faspass:
◦ Centralized per packet DC-wide scheduling
Per-flow rate control:
◦ Given the actual number of flows...
◦ ...allocate a rate to transmit data for each pair of
source/destination servers
◦ Rate re-computation occurs whenever a flow starts/ends
Two step algorithms
1. Rate assignment
2. Routing
Grant:
SRCi -> DSTj
Through Sk
With rate Rij
Centralized Controller
At each
new/ended
flow
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 62
Example of ASY algorithms
Rate assignment algorithm
◦ rij = number of flows from server i to server j
◦ Normalize the matrix between pairs of servers in order
to be double-stochastic
◦ Avoid to overload servers
◦ Possible parallel implementation
Routing
◦ Exploit the available parallel paths to reduce output
contentions
◦ Randomly assign a different path for each flow (e.g.,
ECMP)
r11 r12 r13 . . . r1N
r21 . . . .
r31 . . . .
. . . . .
. . . .
rN1 rN2 . . . rNN
Normalize
Normalize
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 63
SYN vs ASY - schedulers
Scenario
◦ Realistic protocol stack
◦ incast + fixed packet size
SYN schedulers
◦ MMF: Max-Min Fair
◦ OCF: Oldest Cell First
◦ SRJF: Shortest Remaining Job First
Average Flow Completion Time (AFCT)
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 64
Minimizing FCT in DCN
Optimal flow scheduling in DCN requires knowledge about
individual flow length.
◦ Difficult (impossible) to know in advance
◦ Fully information-agnostic schedulers penalize long flows
◦ Knowledge about flow length distribution may help in improving the
scheduling
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 65
Minimizing FCT
with average flow
length distribution
Flow prioritization flows without knowledge about
their length.
◦ Use of average flow length distribution
◦ Fine grained scheduling at host
◦ Coarse scheduling at switches
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 66
Minimizing FCT with average flow length
distribution
Mice Medium Elephant
Average
Realistic packet-level network simulator
◦ Real implementation of TCP and DCTCP
◦ Realistic flow length distributions
◦ Mixed and uniform flow length scenarios
◦ Robustness analysis to CDF underestimation
01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 67

More Related Content

What's hot

Traffic Engineering in Software-Defined Networks
Traffic Engineering in Software-Defined NetworksTraffic Engineering in Software-Defined Networks
Traffic Engineering in Software-Defined NetworksHai Dinh Tuan
 
First Contact: Can Switching to RINA save the Internet?
First Contact: Can Switching to RINA save the Internet?First Contact: Can Switching to RINA save the Internet?
First Contact: Can Switching to RINA save the Internet?ARCFIRE ICT
 
An overview of SDN & Openflow
An overview of SDN & OpenflowAn overview of SDN & Openflow
An overview of SDN & OpenflowPeyman Faizian
 
The hague rina-workshop-intro-eduard
The hague rina-workshop-intro-eduardThe hague rina-workshop-intro-eduard
The hague rina-workshop-intro-eduardICT PRISTINE
 
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay Networking
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay NetworkingMulti-operator "IPC" VPN Slices: Applying RINA to Overlay Networking
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay NetworkingARCFIRE ICT
 
Introduction to SDN: Software Defined Networking
Introduction to SDN: Software Defined NetworkingIntroduction to SDN: Software Defined Networking
Introduction to SDN: Software Defined NetworkingAnkita Mahajan
 
The hague rina-workshop-interop-deployment_vincenzo
The hague rina-workshop-interop-deployment_vincenzoThe hague rina-workshop-interop-deployment_vincenzo
The hague rina-workshop-interop-deployment_vincenzoICT PRISTINE
 
The hague rina-workshop-nfv-diego
The hague rina-workshop-nfv-diegoThe hague rina-workshop-nfv-diego
The hague rina-workshop-nfv-diegoICT PRISTINE
 
RINA Distributed Mobility Management over WiFi
RINA Distributed Mobility Management over WiFiRINA Distributed Mobility Management over WiFi
RINA Distributed Mobility Management over WiFiARCFIRE ICT
 
The hague rina-workshop-mobility-eduard
The hague rina-workshop-mobility-eduardThe hague rina-workshop-mobility-eduard
The hague rina-workshop-mobility-eduardICT PRISTINE
 
The Challenges of SDN/OpenFlow in an Operational and Large-scale Network
The Challenges of SDN/OpenFlow in an Operational and Large-scale NetworkThe Challenges of SDN/OpenFlow in an Operational and Large-scale Network
The Challenges of SDN/OpenFlow in an Operational and Large-scale NetworkOpen Networking Summits
 
Next-gen Network Telemetry is Within Your Packets: In-band OAM
Next-gen Network Telemetry is Within Your Packets: In-band OAMNext-gen Network Telemetry is Within Your Packets: In-band OAM
Next-gen Network Telemetry is Within Your Packets: In-band OAMOpen Networking Summit
 
Deep Packet Inspection technology evolution
Deep Packet Inspection technology evolutionDeep Packet Inspection technology evolution
Deep Packet Inspection technology evolutionDaniel Vinyar
 
Rina p4 rina workshop
Rina p4   rina workshopRina p4   rina workshop
Rina p4 rina workshopEduard Grasa
 
SDN Architecture & Ecosystem
SDN Architecture & EcosystemSDN Architecture & Ecosystem
SDN Architecture & EcosystemKingston Smiler
 
Software Defined Networking - 1
Software Defined Networking - 1Software Defined Networking - 1
Software Defined Networking - 1Pradeep Kumar TS
 

What's hot (20)

Traffic Engineering in Software-Defined Networks
Traffic Engineering in Software-Defined NetworksTraffic Engineering in Software-Defined Networks
Traffic Engineering in Software-Defined Networks
 
First Contact: Can Switching to RINA save the Internet?
First Contact: Can Switching to RINA save the Internet?First Contact: Can Switching to RINA save the Internet?
First Contact: Can Switching to RINA save the Internet?
 
An overview of SDN & Openflow
An overview of SDN & OpenflowAn overview of SDN & Openflow
An overview of SDN & Openflow
 
The hague rina-workshop-intro-eduard
The hague rina-workshop-intro-eduardThe hague rina-workshop-intro-eduard
The hague rina-workshop-intro-eduard
 
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay Networking
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay NetworkingMulti-operator "IPC" VPN Slices: Applying RINA to Overlay Networking
Multi-operator "IPC" VPN Slices: Applying RINA to Overlay Networking
 
Introduction to SDN: Software Defined Networking
Introduction to SDN: Software Defined NetworkingIntroduction to SDN: Software Defined Networking
Introduction to SDN: Software Defined Networking
 
Chapter05
Chapter05Chapter05
Chapter05
 
The hague rina-workshop-interop-deployment_vincenzo
The hague rina-workshop-interop-deployment_vincenzoThe hague rina-workshop-interop-deployment_vincenzo
The hague rina-workshop-interop-deployment_vincenzo
 
The hague rina-workshop-nfv-diego
The hague rina-workshop-nfv-diegoThe hague rina-workshop-nfv-diego
The hague rina-workshop-nfv-diego
 
Chapter13
Chapter13Chapter13
Chapter13
 
RINA Distributed Mobility Management over WiFi
RINA Distributed Mobility Management over WiFiRINA Distributed Mobility Management over WiFi
RINA Distributed Mobility Management over WiFi
 
The hague rina-workshop-mobility-eduard
The hague rina-workshop-mobility-eduardThe hague rina-workshop-mobility-eduard
The hague rina-workshop-mobility-eduard
 
The Challenges of SDN/OpenFlow in an Operational and Large-scale Network
The Challenges of SDN/OpenFlow in an Operational and Large-scale NetworkThe Challenges of SDN/OpenFlow in an Operational and Large-scale Network
The Challenges of SDN/OpenFlow in an Operational and Large-scale Network
 
Link_NwkingforDevOps
Link_NwkingforDevOpsLink_NwkingforDevOps
Link_NwkingforDevOps
 
Next-gen Network Telemetry is Within Your Packets: In-band OAM
Next-gen Network Telemetry is Within Your Packets: In-band OAMNext-gen Network Telemetry is Within Your Packets: In-band OAM
Next-gen Network Telemetry is Within Your Packets: In-band OAM
 
Deep Packet Inspection technology evolution
Deep Packet Inspection technology evolutionDeep Packet Inspection technology evolution
Deep Packet Inspection technology evolution
 
Rina p4 rina workshop
Rina p4   rina workshopRina p4   rina workshop
Rina p4 rina workshop
 
Rina2020 michal
Rina2020 michalRina2020 michal
Rina2020 michal
 
SDN Architecture & Ecosystem
SDN Architecture & EcosystemSDN Architecture & Ecosystem
SDN Architecture & Ecosystem
 
Software Defined Networking - 1
Software Defined Networking - 1Software Defined Networking - 1
Software Defined Networking - 1
 

Similar to German Sviridov - PhD defense

Edge virtualisation for Carrier Networks
Edge virtualisation for Carrier NetworksEdge virtualisation for Carrier Networks
Edge virtualisation for Carrier NetworksMyNOG
 
Multicloud as the Next Generation of Cloud Infrastructure
Multicloud as the Next Generation of Cloud Infrastructure Multicloud as the Next Generation of Cloud Infrastructure
Multicloud as the Next Generation of Cloud Infrastructure Brad Eckert
 
ONF & iSDX Webinar
ONF & iSDX WebinarONF & iSDX Webinar
ONF & iSDX WebinarKatie Hyman
 
Packet Optical SDN Field Trial for Multi-Layer Network Optimization
Packet Optical SDN Field Trial for Multi-Layer Network OptimizationPacket Optical SDN Field Trial for Multi-Layer Network Optimization
Packet Optical SDN Field Trial for Multi-Layer Network OptimizationADVA
 
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...Radisys Corporation
 
Business Models for Dynamically Provisioned Optical Networks
Business Models for Dynamically Provisioned Optical NetworksBusiness Models for Dynamically Provisioned Optical Networks
Business Models for Dynamically Provisioned Optical NetworksTal Lavian Ph.D.
 
Next Generation Optical Networking: Software-Defined Optical Networking
Next Generation Optical Networking: Software-Defined Optical NetworkingNext Generation Optical Networking: Software-Defined Optical Networking
Next Generation Optical Networking: Software-Defined Optical NetworkingADVA
 
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...Tal Lavian Ph.D.
 
Transport SDN Overview and Standards Update: Industry Perspectives
Transport SDN Overview and Standards Update: Industry PerspectivesTransport SDN Overview and Standards Update: Industry Perspectives
Transport SDN Overview and Standards Update: Industry PerspectivesInfinera
 
cncf overview and building edge computing using kubernetes
cncf overview and building edge computing using kubernetescncf overview and building edge computing using kubernetes
cncf overview and building edge computing using kubernetesKrishna-Kumar
 
Software Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaSoftware Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaCPqD
 
Software Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaSoftware Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaCPqD
 
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...Tal Lavian Ph.D.
 
Necos keynote ii_mobislice
Necos keynote ii_mobisliceNecos keynote ii_mobislice
Necos keynote ii_mobisliceAugusto Neto
 
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...Lionel Briand
 
Edge Device Multi-unicasting for Video Streaming
Edge Device Multi-unicasting for Video StreamingEdge Device Multi-unicasting for Video Streaming
Edge Device Multi-unicasting for Video StreamingTal Lavian Ph.D.
 
Cloud interconnection networks basic .pptx
Cloud interconnection networks basic .pptxCloud interconnection networks basic .pptx
Cloud interconnection networks basic .pptxRahulBhole12
 

Similar to German Sviridov - PhD defense (20)

Simplifying Wired Network Deployments with Software-Defined Networking (SDN)
Simplifying Wired Network Deployments with Software-Defined Networking (SDN)Simplifying Wired Network Deployments with Software-Defined Networking (SDN)
Simplifying Wired Network Deployments with Software-Defined Networking (SDN)
 
Edge virtualisation for Carrier Networks
Edge virtualisation for Carrier NetworksEdge virtualisation for Carrier Networks
Edge virtualisation for Carrier Networks
 
Multicloud as the Next Generation of Cloud Infrastructure
Multicloud as the Next Generation of Cloud Infrastructure Multicloud as the Next Generation of Cloud Infrastructure
Multicloud as the Next Generation of Cloud Infrastructure
 
ONF & iSDX Webinar
ONF & iSDX WebinarONF & iSDX Webinar
ONF & iSDX Webinar
 
Packet Optical SDN Field Trial for Multi-Layer Network Optimization
Packet Optical SDN Field Trial for Multi-Layer Network OptimizationPacket Optical SDN Field Trial for Multi-Layer Network Optimization
Packet Optical SDN Field Trial for Multi-Layer Network Optimization
 
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...
Radisys/Wind River: The Telcom Cloud - Deployment Strategies: SDN/NFV and Vir...
 
Business Models for Dynamically Provisioned Optical Networks
Business Models for Dynamically Provisioned Optical NetworksBusiness Models for Dynamically Provisioned Optical Networks
Business Models for Dynamically Provisioned Optical Networks
 
Next Generation Optical Networking: Software-Defined Optical Networking
Next Generation Optical Networking: Software-Defined Optical NetworkingNext Generation Optical Networking: Software-Defined Optical Networking
Next Generation Optical Networking: Software-Defined Optical Networking
 
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
 
Transport SDN Overview and Standards Update: Industry Perspectives
Transport SDN Overview and Standards Update: Industry PerspectivesTransport SDN Overview and Standards Update: Industry Perspectives
Transport SDN Overview and Standards Update: Industry Perspectives
 
cncf overview and building edge computing using kubernetes
cncf overview and building edge computing using kubernetescncf overview and building edge computing using kubernetes
cncf overview and building edge computing using kubernetes
 
Verizon Managed SD-WAN with Cisco IWAN
Verizon Managed SD-WAN with Cisco IWAN Verizon Managed SD-WAN with Cisco IWAN
Verizon Managed SD-WAN with Cisco IWAN
 
Software Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaSoftware Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur Channegowda
 
Software Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur ChannegowdaSoftware Defined Optical Networks - Mayur Channegowda
Software Defined Optical Networks - Mayur Channegowda
 
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
DWDM-RAM: DARPA-Sponsored Research for Data Intensive Service-on-Demand Advan...
 
Necos keynote ii_mobislice
Necos keynote ii_mobisliceNecos keynote ii_mobislice
Necos keynote ii_mobislice
 
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...
Dynamic Adaptation of Software-defined Networks for IoT Systems: A Search-bas...
 
Evolution of internet by Ali Kashif
Evolution of internet  by Ali KashifEvolution of internet  by Ali Kashif
Evolution of internet by Ali Kashif
 
Edge Device Multi-unicasting for Video Streaming
Edge Device Multi-unicasting for Video StreamingEdge Device Multi-unicasting for Video Streaming
Edge Device Multi-unicasting for Video Streaming
 
Cloud interconnection networks basic .pptx
Cloud interconnection networks basic .pptxCloud interconnection networks basic .pptx
Cloud interconnection networks basic .pptx
 

Recently uploaded

2024 May Patch Tuesday
2024 May Patch Tuesday2024 May Patch Tuesday
2024 May Patch TuesdayIvanti
 
AI in Action: Real World Use Cases by Anitaraj
AI in Action: Real World Use Cases by AnitarajAI in Action: Real World Use Cases by Anitaraj
AI in Action: Real World Use Cases by AnitarajAnitaRaj43
 
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....rightmanforbloodline
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontologyjohnbeverley2021
 
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)Samir Dash
 
WebRTC and SIP not just audio and video @ OpenSIPS 2024
WebRTC and SIP not just audio and video @ OpenSIPS 2024WebRTC and SIP not just audio and video @ OpenSIPS 2024
WebRTC and SIP not just audio and video @ OpenSIPS 2024Lorenzo Miniero
 
Decarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceDecarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceIES VE
 
Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Zilliz
 
Quantum Leap in Next-Generation Computing
Quantum Leap in Next-Generation ComputingQuantum Leap in Next-Generation Computing
Quantum Leap in Next-Generation ComputingWSO2
 
Event-Driven Architecture Masterclass: Challenges in Stream Processing
Event-Driven Architecture Masterclass: Challenges in Stream ProcessingEvent-Driven Architecture Masterclass: Challenges in Stream Processing
Event-Driven Architecture Masterclass: Challenges in Stream ProcessingScyllaDB
 
Modernizing Legacy Systems Using Ballerina
Modernizing Legacy Systems Using BallerinaModernizing Legacy Systems Using Ballerina
Modernizing Legacy Systems Using BallerinaWSO2
 
API Governance and Monetization - The evolution of API governance
API Governance and Monetization -  The evolution of API governanceAPI Governance and Monetization -  The evolution of API governance
API Governance and Monetization - The evolution of API governanceWSO2
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Victor Rentea
 
Stronger Together: Developing an Organizational Strategy for Accessible Desig...
Stronger Together: Developing an Organizational Strategy for Accessible Desig...Stronger Together: Developing an Organizational Strategy for Accessible Desig...
Stronger Together: Developing an Organizational Strategy for Accessible Desig...caitlingebhard1
 
Portal Kombat : extension du réseau de propagande russe
Portal Kombat : extension du réseau de propagande russePortal Kombat : extension du réseau de propagande russe
Portal Kombat : extension du réseau de propagande russe中 央社
 
Intro to Passkeys and the State of Passwordless.pptx
Intro to Passkeys and the State of Passwordless.pptxIntro to Passkeys and the State of Passwordless.pptx
Intro to Passkeys and the State of Passwordless.pptxFIDO Alliance
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDropbox
 
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptxHarnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptxFIDO Alliance
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FMESafe Software
 
UiPath manufacturing technology benefits and AI overview
UiPath manufacturing technology benefits and AI overviewUiPath manufacturing technology benefits and AI overview
UiPath manufacturing technology benefits and AI overviewDianaGray10
 

Recently uploaded (20)

2024 May Patch Tuesday
2024 May Patch Tuesday2024 May Patch Tuesday
2024 May Patch Tuesday
 
AI in Action: Real World Use Cases by Anitaraj
AI in Action: Real World Use Cases by AnitarajAI in Action: Real World Use Cases by Anitaraj
AI in Action: Real World Use Cases by Anitaraj
 
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontology
 
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)
AI+A11Y 11MAY2024 HYDERBAD GAAD 2024 - HelloA11Y (11 May 2024)
 
WebRTC and SIP not just audio and video @ OpenSIPS 2024
WebRTC and SIP not just audio and video @ OpenSIPS 2024WebRTC and SIP not just audio and video @ OpenSIPS 2024
WebRTC and SIP not just audio and video @ OpenSIPS 2024
 
Decarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceDecarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational Performance
 
Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)
 
Quantum Leap in Next-Generation Computing
Quantum Leap in Next-Generation ComputingQuantum Leap in Next-Generation Computing
Quantum Leap in Next-Generation Computing
 
Event-Driven Architecture Masterclass: Challenges in Stream Processing
Event-Driven Architecture Masterclass: Challenges in Stream ProcessingEvent-Driven Architecture Masterclass: Challenges in Stream Processing
Event-Driven Architecture Masterclass: Challenges in Stream Processing
 
Modernizing Legacy Systems Using Ballerina
Modernizing Legacy Systems Using BallerinaModernizing Legacy Systems Using Ballerina
Modernizing Legacy Systems Using Ballerina
 
API Governance and Monetization - The evolution of API governance
API Governance and Monetization -  The evolution of API governanceAPI Governance and Monetization -  The evolution of API governance
API Governance and Monetization - The evolution of API governance
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
 
Stronger Together: Developing an Organizational Strategy for Accessible Desig...
Stronger Together: Developing an Organizational Strategy for Accessible Desig...Stronger Together: Developing an Organizational Strategy for Accessible Desig...
Stronger Together: Developing an Organizational Strategy for Accessible Desig...
 
Portal Kombat : extension du réseau de propagande russe
Portal Kombat : extension du réseau de propagande russePortal Kombat : extension du réseau de propagande russe
Portal Kombat : extension du réseau de propagande russe
 
Intro to Passkeys and the State of Passwordless.pptx
Intro to Passkeys and the State of Passwordless.pptxIntro to Passkeys and the State of Passwordless.pptx
Intro to Passkeys and the State of Passwordless.pptx
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor Presentation
 
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptxHarnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx
Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
 
UiPath manufacturing technology benefits and AI overview
UiPath manufacturing technology benefits and AI overviewUiPath manufacturing technology benefits and AI overview
UiPath manufacturing technology benefits and AI overview
 

German Sviridov - PhD defense

  • 1. Traffic Optimization in Data Center and Software-Defined Programmable Networks GERMAN SVIRIDOV Doctoral Program in Electrical, Electronics and Communications Engineering (XXXIII cycle) Supervisors: Prof. Paolo Giaccone Prof. Andrea Bianco
  • 2. 2 Stateful Software-Defined Networking: enabling replicated network applications in programmable data planes Flow scheduling in data center networks: optimizing flow performance by minimizing flow completion time Cloud gaming and game engines: Quality of Experience improvement in cloud gaming 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV Blockchains for vehicular applications 2
  • 3. 3 Stateful Software-Defined Networking: enabling replicated network applications in programmable data planes Flow scheduling in data center networks: optimizing flow performance by minimizing flow completion time Cloud gaming and game engines: Quality of Experience improvement in cloud gaming 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV Blockchains for vehicular applications 3
  • 4. Enabling replicated network applications in programmable data planes 03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 4
  • 5. Classic Software Defined Networks (SDN) Control Plane Data plane Physical separation of control and data plane A single entity (controller) for the control plane Data plane operations are now centrally managed Key enabler for novel applications 03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 5
  • 6. Issues with classic SDN Control Plane Data plane • Physical latency between switches and the controller • Software processing latency Added latency • Per flow processing only • Support of legacy protocols only Inflexible commercial hardware • Local in-switch processing • Simplified switch-local control logic for latency-sensitive application Need to take a step back 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 6
  • 7. Programmable data planes • User-defined packet parsing • Support for virtually any data protocol • Modification to any packet field at line rate Programmable parsers • Programmable match-action rules • Access to persistent states 𝒔𝒊 • Packet recirculation/cloning Programmable packet processing 𝑠𝑖 03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 7
  • 8. Enabling new applications and accelerating old ones 𝑠𝑖 Financial applications Machine learning MQTT brokers DBMS Acceleration Content caching Overlay protocols Load balancers Packet scheduling Network telemetry Firewall/IDS 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 8
  • 9. Limits of data plane-assisted network applications 𝑠1 𝑠2 𝑠3 𝑠4 •Network applications are local to a single switch •Independent behavior in the network Locality constraints •DDoS detection •Distributed application/link-aware load balancers •Network-wide rate limiters •... Precluded applications 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 9
  • 10. Limits of data plane-assisted network applications 𝑠1 𝑠2 𝑠3 𝑠4 •Network applications are local to a single switch •Independent behavior in the network Locality constraints •DDoS detection •Distributed application/link-aware load balancers •Network-wide rate limiters •... Precluded applications 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 10 How to implement logically centralized network- wide network applications?
  • 11. Logically centralized network applications 𝑠1 𝑠2 𝑠3 𝑠4 Unique network application Centralized state with global view of the network Traffic reroute Data overhead Bandwidth constraints 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 11
  • 12. 𝑠1 𝑠2 𝑠3 𝑠4 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV State replication in programmable data planes 12 Replicated network application Centralized state with global view of the network No traffic reroute No data overhead Sync overhead
  • 13. 𝑠1 𝑠3 𝑠4 𝑠2 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV State replication in programmable data planes 13 Replicated network application Centralized state with global view of the network No traffic reroute No data overhead Sync overhead
  • 14. A recipe for state replication in programmable data planes LODGE Suitable replication algorithm Need to convey state-updates to other replicas Need to define a suitable transport protocol Need to be able to generate new packets 𝑠2 𝑠3 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 14
  • 15. A recipe for state replication in programmable data planes LODGE Eventual consistency Need to convey state-updates to other replicas Need to define a suitable transport protocol Need to be able to generate new packets 𝑠2 𝑠3 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 15
  • 16. A recipe for state replication in programmable data planes LODGE Eventual consistency Controller- constructed distribution tree Need to define a suitable transport protocol Need to be able to generate new packets 𝑠2 𝑠3 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 16
  • 17. A recipe for state replication in programmable data planes LODGE Eventual consistency Controller- constructed distribution tree Tailored P4/OPP-enabled transport protocol Need to be able to generate new packets 𝑠2 𝑠3 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 17
  • 18. A recipe for state replication in programmable data planes LODGE Eventual consistency Controller- constructed distribution tree Tailored P4/OPP-enabled transport protocol Traffic-triggered packet cloning and recirculation 𝑠2 𝑠3 𝑠2 1 𝑠3 1 D S 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 18
  • 19. A recipe for state replication in programmable data planes LODGE Eventual consistency Controller- constructed distribution tree Tailored P4/OPP-enabled transport protocol Traffic-triggered packet cloning and recirculation 𝑠2 𝑠3 𝑠2 1 𝑠3 1 D S 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 19
  • 20. LODGE-Enabled Applications server cluster 1 server cluster 3 server cluster 2 server cluster 4 ASN1 ASN2 ASN3 ASN4 R2 R1 R3 R4 SW1 SW2 SW3 SW4 Realistic scenario Emulated testbed LODGE implementation •WAN Autonomous System (AS) with 4 neighbours •Incoming traffic monitoring and reaction •V1 Model with BMV •Emulated network •P4 14/16 •Open Packet Processor (OPP) 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 20
  • 21. LODGE-Enabled Applications • Distributed DDoS from ASN# • LODGE logic embedded at SW1 and SW3 • Simultaneous detection of the DDoS at all switches without traffic rerouting • Distributed rate limiting • LODGE application embedded at border routers • Simultaneous drop of the rate for both flows 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 21 server cluster 1 server cluster 3 server cluster 2 server cluster 4 ASN1 ASN2 ASN3 ASN4 R2 R1 R3 R4 SW1 SW2 SW3 SW4
  • 22. LODGE-Enabled Applications • Distributed DDoS from ASN# • LODGE logic embedded at SW1 and SW3 • Simultaneous detection of the DDoS at all switches without traffic rerouting 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV Why SW1 & SW3 or R# in particular? 22 server cluster 1 server cluster 3 server cluster 2 server cluster 4 ASN1 ASN2 ASN3 ASN4 R2 R1 R3 R4 SW1 SW2 SW3 SW4 • Distributed rate limiting • LODGE application embedded at border routers • Simultaneous drop of the rate for both flows
  • 23. Where to place the state replicas? 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 23
  • 24. State replicas placement: data vs synchronization traffic 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 24 Data traffic Sync traffic 𝑠1 1 and 𝑠1 2 : replicas of the same state 𝑠1 X2 X1 s s s s - More data traffic overhead - Less sync traffic overhead - Less data traffic overhead - More sync traffic overhead Number of state replicas 𝑠1 1 𝑠1 2 𝑠1
  • 25. State replicas placement: data vs synchronization traffic 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 25 Number of state replicas • Data vs synchronization traffic trade-off? • How many replicas to use? • Where to place the replicas? Data traffic Sync traffic 𝑠1 1 and 𝑠1 2 : replicas of the same state 𝑠1 X2 X1 s s s s 𝑠1 1 𝑠1 2 𝑠1 - More data traffic overhead - Less sync traffic overhead - Less data traffic overhead - More sync traffic overhead
  • 26. Optimal state replication in stateful data planes 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 26 MILP formulation for optimal state replication Input • Network topology • Traffic matrix • Network applications • Per-flow application requirements • Amount of sync traffic ෡ 𝜆𝑠 Constrained by • Bandwidth availability • Switch resources availability • Traffic matrix Objective • Minimize maximum congestion • Minimize total traffic Output • Number of replicas • Replicas placement • Data traffic routing • Sync traffic routing
  • 27. Optimal state replication in stateful data planes Difficult to apply to big topologies! 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 27 MILP formulation for optimal state replication Input • Network topology • Traffic matrix • Network applications • Per-flow application requirements • Amount of sync traffic ෡ 𝜆𝑠 Constrained by • Bandwidth availability • Switch resources availability • Traffic matrix Objective • Minimize maximum congestion • Minimize total traffic Output • Number of replicas • Replicas placement • Data traffic routing • Sync traffic routing
  • 28. Optimal state replication – approximated solution 28 Algorithm for replica placement • Based on betweenness centrality • Solution perturbation Optimal number of replicas • Analytical expression • Asymptotic optimal for Manhattan topology 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 28
  • 29. How to put everything together? Network application abstracion Application decompositi on Consistency constraints LODGE dataplane implementation P4 OPP Optimal state placement Number of states State position Traffic routing Data traffic Synchroniza- tion traffic How to define distributed network applications? 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 29
  • 30. LOADER: An abstraction for replicated network applications 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 30
  • 31. 31 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 31 The goals of LOADER Define an abstraction model for replicated network applications Enable efficient translation to basic network primitives Combine efforts of optimal state replication and LODGE into a unified framework
  • 32. 32 Programming abstraction for network applications: •Exposure of replicated states to the programmer •Management of state inconsistency •Definition of 𝜆𝑠 Compiler •Automatic policy decomposition and translation to target- specific language Embedder •Automatic state embedding •Based on heuristic for optimal state replication 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 32
  • 33. 33 Programming abstraction for network applications: •Exposure of replicated states to the programmer •Management of state inconsistency •Definition of 𝜆𝑠 Compiler •Automatic policy decomposition and translation to target- specific language Embedder •Automatic state embedding •Based on heuristic for optimal state replication 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 33
  • 34. 34 Programming abstraction for network applications: •Exposure of replicated states to the programmer •Management of state inconsistency •Definition of 𝜆𝑠 Compiler •Automatic policy decomposition and translation to target- specific language Embedder •Automatic state placement •Based on heuristic for optimal state replication 𝑠1 𝑠2 𝑠3 𝑠4 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 34
  • 35. 35 Programming abstraction for network applications: •Exposure of replicated states to the programmer •Management of state inconsistency •Definition of 𝜆𝑠 Compiler •Automatic policy decomposition and translation to target- specific language Embedder •Automatic state placement •Based on heuristic for optimal state replication 𝑠1 𝑠2 𝑠3 𝑠4 𝑠2 1 𝑠3 1 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 35
  • 36. LOADER Network application abstracion Application decompositi on Consistency constraints LODGE dataplane implementation P4 OPP Optimal state placement Number of states State position Traffic routing Data traffic Synchroniza- tion traffic [1] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LODGE: LOcal Decisions on Global statEs in programmable data planes, in: IEEE NetSoft, 2018 [2] Muqaddas, Abubakar Sidique; Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Optimal state replication in stateful dataplanes, in: IEEE JSAC, 2020 [3] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LOcAl DEcisions on Replicated States (LOADER) in programmable data planes: programming abstraction and experimental evaluation, in: Computer Networks 2021 [1] [2] [3] 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 36
  • 37. 37 Stateful Software-Defined Networking: enabling replicated network applications in programmable data planes Flow scheduling in data center networks: optimizing flow performance by minimizing flow completion time Cloud gaming and game engines: Quality of Experience improvement in cloud gaming 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV Blockchains for vehicular applications 37
  • 38. Automating game QoE assessment in cloud gaming 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 38
  • 39. The next generation of gaming: Cloud Gaming • Games run on a remote server • Video stream of the game is sent to the users Remote game rendering • Users buy the game and pay the subscription fee • Play directly on a TV or phone No need for hardware 03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 39
  • 40. The next generation of gaming: Cloud Gaming • Games run on a remote server • Video stream of the game is sent to the users Remote game rendering • Users buy the game and pay the subscription fee • Play directly on a TV or phone No need for hardware 03/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 40 Video streaming (High BW, High Delay) Online gaming (Low BW, Low Delay) Input lag (High BW, Low Delay)
  • 41. Not all games are made equal Different impact of latency Different genres Multiple stages •Action/FPS •Strategy, card, interactive fiction •Action stages •Exploration/narration stages 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 41
  • 42. Not all games are made equal Different impact of latency Different genres Multiple stages •Accuracy decrease in FPS •Smaller crop yield in farming simulator? •Action/FPS •Strategy, card, interactive fiction •Action stages •Exploration/narration stages Different network requirements 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 42
  • 43. Not all games are made equal Different impact of latency Different genres Multiple stages •Accuracy decrease in FPS •Smaller crop yield in farming simulator? •Action/FPS •Strategy, card, interactive fiction •Action stages •Exploration/narration stages Different network requirements Need for fine-grained game QoE assessment 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 43
  • 44. Game Quality of Experience (QoE) assessment in practice 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 44
  • 45. Game - 25 new game releases per day on Steam - Impossible to keep up with new releases WAN DC 379 565 1771 2964 4207 7049 9050 0 2000 4000 6000 8000 10000 2012 2013 2014 2015 2016 2017 2018 Number of games released on steam Human subjects - Difficult and expensive to find enough subjects - No diversity guarantee Controlled environment - Need to bring subjects to the lab - Difficult to achieve in the wild Outcome - Noisy experimental data - Coarse grained information - No repeatability - Outdated information 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 45
  • 46. Game - 25 new game releases per day on Steam - Impossible to keep up with new releases WAN DC 379 565 1771 2964 4207 7049 9050 0 2000 4000 6000 8000 10000 2012 2013 2014 2015 2016 2017 2018 Number of games released on steam Guilty Human subjects - Difficult and expensive to find enough subjects - No diversity guarantee Controlled environment - Need to bring subjects to the lab - Difficult to achieve in the wild Outcome - Noisy experimental data - Coarse grained information - No repeatability - Outdated information 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 46
  • 47. WAN DC 379 565 1771 2964 4207 7049 9050 0 2000 4000 6000 8000 10000 2012 2013 2014 2015 2016 2017 2018 Number of games released on steam Game - 25 new game releases per day on Steam - Impossible to keep up with new releases Bots - Diversity and experience can be programmed - Still expensive Controlled environment - Can be easily simulated - Fine tuning is easily achieved Outcome + Deterministic results + Fine grained information + Repeatability - Outdated information 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 47
  • 48. WAN DC 379 565 1771 2964 4207 7049 9050 0 2000 4000 6000 8000 10000 2012 2013 2014 2015 2016 2017 2018 Number of games released on steam Game - 25 new game releases per day on Steam - Impossible to keep up with new releases? Bots - Diversity and experience can be programmed - Still expensive? Controlled environment - Can be easily simulated - Fine tuning is easily achieved Outcome - Deterministic results - Fine grained information - Repeatability - Outdated information AI? 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 48
  • 49. Removing humans from the loop: Putting the AI to work for QoE assessment 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 49
  • 50. Deep reinforcement learning (DRL) for video games Learning to play from raw video pixels Trial and error approach No supervision Good adaptability Active exploration Reward shaping Tunable behavior 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 50
  • 51. Learning to play Atari • Deep Q Networks • Simple, yet powerful model for DRL One algorithm • Seaquest, Beam Rider and Breakout • Taken from the Atari game catalogue Three games • (Super) human level achieved in less than 5 hours of training • Ready to be deployed in the testbed Three AI agents 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 51
  • 52. Measuring the gaming performance • Per-frame lag 𝑙 • Per-frame lag probability 𝑝𝑙𝑎𝑔 • Per-keystroke drop probability 𝑝𝑑𝑟𝑜𝑝 Emulated network 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 52
  • 53. Measuring the gaming performance • Per-frame lag 𝑙 • Per-frame lag probability 𝑝𝑙𝑎𝑔 • Per-keystroke drop probability 𝑝𝑑𝑟𝑜𝑝 Emulated network 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 53
  • 54. Maximizing average game score Different network requirements Need for fine-grained game QoE assessment Without traffic prioritization With traffic prioritization 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 54
  • 55. Moving to more complex games Moving to more complex games 55 Classic Doom Similar algorithm for training Different maps Different game modes Network parameter perturbation 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 55
  • 56. Automating game QoE assessment Cloud gaming scenario Emulated network with perturbation Artificial players trained with DRL Atari Doom [4] Sviridov, German; Beliar, Cedric; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Removing human players from the loop: AI-assisted assessment of Gaming QoE, in: IEEE Infocom NI Workshop, 2020 [5] Sviridov, German; Beliar, Cedric; Simon, Gwendal; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Leveraging AI players for QoE estimation in cloud gaming, in IEEE Infocom Demo Session, 2020 [5] [4-5] Per-game QoS mechanism which maximizes average game QoE 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 56
  • 57. Past works 57 02/03/2021 PHD DEFENSE – GERMAN SVIRIDOV •Is centralized flow scheduling feasible in practice? •Can we reach close to state of the art performance while employing commodity switches for flow scheduling? Flow scheduling in data center networks •Blockchain as an enabler for mobility verification •How to manage trust for sensitive data? •Which scale private blockchains can reach? Blockchains for vehicular applications [6-7] [8] [6] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Low-Complexity Flow Scheduling for Commodity Switches in Data Center Networks, in: IEEE Globecom, 2019 [7] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, To Sync or Not to Sync: Why Asynchronous Traffic Control Is Good Enough for Your Data Center, in: IEEE Globecom, 2018 [8] Chiasserini, Carla Fabiana; Giaccone, Paolo; Malnati, Giovanni; Macagno, Michele; Sviridov, German, Blockchain-based mobility verification of connected cars, in IEEE CCNC, 2020 57
  • 58. Current work – (Cloud) Gaming 58 •Games are made of different stages •Each stage requires different QoS •Can we achieve fine-grained QoS control? Game stage classification •Training time of a given game is typically unknown •Can we understand the training difficulty before even training? Estimating training complexity •Game engines are monolithic •Opposite of the cloud-based applications •Can we make cloud gaming truly cloud-based? Distributed gaming engines Graphics Front-end A.I. Physics 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 58
  • 59. References 59 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV [1] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LODGE: LOcal Decisions on Global statEs in programmable data planes, in: IEEE NetSoft, 2018 [2] Muqaddas, Abubakar Sidique; Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Optimal state replication in stateful dataplanes, in: IEEE JSAC, 2020 [3] Sviridov, German; Bonola, Marco; Tulumello, Angelo; Giaccone, Paolo; Bianco, Andrea; Bianchi, Giuseppe, LOcAl DEcisions on Replicated States (LOADER) in programmable data planes: programming abstraction and experimental evaluation, in: Computer Networks 2021 [4] Sviridov, German; Beliar, Cedric; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Removing human players from the loop: AI-assisted assessment of Gaming QoE, in: IEEE Infocom NI Workshop, 2020 [5] Sviridov, German; Beliar, Cedric; Simon, Gwendal; Bianco, Andrea; Giaccone, Paolo; Rossi, Dario, Leveraging AI players for QoE estimation in cloud gaming, in IEEE Infocom Demo Session, 2020 [6] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, Low-Complexity Flow Scheduling for Commodity Switches in Data Center Networks, in: IEEE Globecom, 2019 [7] Sviridov, German; Giaccone, Paolo; Bianco, Andrea, To Sync or Not to Sync: Why Asynchronous Traffic Control Is Good Enough for Your Data Center, in: IEEE Globecom, 2018 [8] Chiasserini, Carla Fabiana; Giaccone, Paolo; Malnati, Giovanni; Macagno, Michele; Sviridov, German, Blockchain-based mobility verification of connected cars, in IEEE CCNC, 2020 59
  • 60. Backup slides 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 60
  • 61. Centralized explicit rate assignment Taking inspiration from Faspass: ◦ Centralized per packet DC-wide scheduling Per-flow rate control: ◦ Given the actual number of flows... ◦ ...allocate a rate to transmit data for each pair of source/destination servers ◦ Rate re-computation occurs whenever a flow starts/ends Two step algorithms 1. Rate assignment 2. Routing Grant: SRCi -> DSTj Through Sk With rate Rij Centralized Controller At each new/ended flow 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 62
  • 62. Example of ASY algorithms Rate assignment algorithm ◦ rij = number of flows from server i to server j ◦ Normalize the matrix between pairs of servers in order to be double-stochastic ◦ Avoid to overload servers ◦ Possible parallel implementation Routing ◦ Exploit the available parallel paths to reduce output contentions ◦ Randomly assign a different path for each flow (e.g., ECMP) r11 r12 r13 . . . r1N r21 . . . . r31 . . . . . . . . . . . . . rN1 rN2 . . . rNN Normalize Normalize 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 63
  • 63. SYN vs ASY - schedulers Scenario ◦ Realistic protocol stack ◦ incast + fixed packet size SYN schedulers ◦ MMF: Max-Min Fair ◦ OCF: Oldest Cell First ◦ SRJF: Shortest Remaining Job First Average Flow Completion Time (AFCT) 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 64
  • 64. Minimizing FCT in DCN Optimal flow scheduling in DCN requires knowledge about individual flow length. ◦ Difficult (impossible) to know in advance ◦ Fully information-agnostic schedulers penalize long flows ◦ Knowledge about flow length distribution may help in improving the scheduling 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 65
  • 65. Minimizing FCT with average flow length distribution Flow prioritization flows without knowledge about their length. ◦ Use of average flow length distribution ◦ Fine grained scheduling at host ◦ Coarse scheduling at switches 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 66
  • 66. Minimizing FCT with average flow length distribution Mice Medium Elephant Average Realistic packet-level network simulator ◦ Real implementation of TCP and DCTCP ◦ Realistic flow length distributions ◦ Mixed and uniform flow length scenarios ◦ Robustness analysis to CDF underestimation 01/03/2021 PHD DEFENSE – GERMAN SVIRIDOV 67