SlideShare a Scribd company logo
Online Learning for Low-latency Streaming
Grand Challenge on Adaptation Algorithms for Near-Second Latency
Theo Karagkioules, Rufael Mekuria, Dirk Griffioen, Arjen Wagenaar
Unified Streaming, Amsterdam, Netherlands
Low-latency adaptive streaming
Motivation
HTTP Adaptive Streaming (HAS) is now the dominant source of data traffic.
• It is projected to reach 82% of the global Internet traffic, by 20221
• OTT services (Netflix, YouTube etc.) account for > 50% of the global ⇓ traffic2
• Cord-cutters and live applications are expedited → rise of low-latency streaming
Some current challenges of bitrate adaptation:
• Channel instability (physical phenomena)
• QoE management
• Reliable low-latency video streaming (1-5 s) . Need for robust, lightweight and
universal solutions ← our contribution
1
Cisco Visual Networking Index, “Forecast and Trends, 2017-2022” - Cisco ’19
2
Global Internet Phenomena - Sandvine 19’
Low-latency adaptive streaming 1 / 11
Adaptive streaming media format
Low-latency bitrate adaptation is facilitated by:
• Multiple qualities={resolutions, encoding rates} served from a regular origin.
• By requiring commodity web infrastructure → operational costs ↓
• Fragments - smaller (independently decodable) files
• Independent fragment requests (multiple servers) maintain session state
• Data segmentation → finer granularity of bitrate selection → channel utilization ↑
• A client initially fetches an XML file, which contains content information
• Control logic, a.k.a a bitrate adaptation algorithm→ bitrate indicates quality
• Logic that selects the appropriate encoding bitrate ∀ fragments, given network
conditions
• Essentially matches the streaming rate to the channel rate, via means of adaptation
• CMAF in combination with the supported chunked transfer mechanism of HTTP 1.1
and beyond, have set the stage for low-latency streaming
Low-latency adaptive streaming 2 / 11
Bitrate adaptation via learning
Motivation
Insights from earlier research (MMSys ’17)3:
• Existing schemes → parameter tuning according to network or app. scenario
• Difficulty in generalizing well beyond a certain scope of usage
• Optimal adaptation over fluctuating channels remains a challenging operation
• Low-latency bitrate adaptation still relies on throughput estimation
To overcome this limitation → resort to learning or control theory
• Practical implementation may be hindered
• by energy-demanding architectures (Deep Learning)
• or by the complexity of exploring the complete optimization space (MDPs)
Proposal: Novel adaptation algorithm based on Online Convex Optimization (OCO)
• OCO emerged in 2003 (Zinkevich) as a very effective online learning framework
• Does not rely on estimations (throughput). Only on historical values.
• “Model-free”, as no assumption for the statistical model of the channel is required
• Independent of any parameter selection concerning the streaming environment
• It provides tractable feasibility and performance guarantees
3
T. Karagkioules et al. A Comparative Case Study of HTTP Adaptive Streaming Algorithms in Mobile Networks
Bitrate adaptation via learning 3 / 11
Problem definition
An adaptation algorithm is an optimization solution with the objective of maximizing
the video bitrate, while ensuring stable and continuous low-latency streaming.
However, the application of OCO in HAS is not a straightforward task
• OCO requires convex decision space and constraints
• HAS has a discrete decision space (set of qualities for every fragment)
• Instantaneous state-dependent constraints (finite-sized buffer)
Modelling assumptions:
• We model the adaptive streaming client by a learning agent
• Agent minimizes latency, s.t. scheduling constraints of the buffer
• We model the channel rate evolution by an adversary
• Adversary decides the cost of each decision only after it has been taken
• Adversarial setting is general enough to include any potential time-variant distribution
• Learning agent minimizes an adversarial loss function, unknown at decision time.
We fulfill the OCO requirement (convex decision set and constraints):
1. Agent decides video quality, according to a probability distribution for the
appropriate bitrate, given network and latency conditions
2. Relaxation to unbounded buffer → adheres to time-averaging constraints
Bitrate adaptation via learning 4 / 11
Relaxations
Relaxation 1: Convexification of decision set by randomization
∀ fragment t ∈ {1, . . . , T}, the client selects quality xt ∈ X = {1, . . . N}
• upon bitrate indication rxt ∈ {r1, . . . , rN} of the bitrate adaptation algorithm
Consider the probability simplex:
Ω = {ω ∈ RN
: ω ≥ 0 ∧ ∥ω∥1 = 1}
→ Instead of xt, client learns the optimal probability ωt = (ωt,n)n=1,...,N of picking xt ∈ X
Relaxation 2: Convexification of buffer constraints
Buffer evolution:
Bt+1 =










Bt − ρt
Playback rate
Seg. size≈V·rxt
St,xt
Ct
Throughput
⇓ time










+
+ V
Seg. duration
−
Inter-request delay
∆t
keeps Bt+1<Bmax
∈ [0, Bmax]
→ We treat the buffer as an infinite queue, i.e. Bt ∈ (−∞, ∞). ∆t is never imposed
• We allow instantaneous violation of the buffer budget
• But utilize a penalty to maintain 0 < Bt < Bmax on average
Bitrate adaptation via learning 5 / 11
Problem formulation
Consider the following functions =⇒ random processes:
˜ft(xt) ≜ V −
ρtSt,xt
Ct
Loss
=⇒ ft(ωt) ≜ E V −
ρtSt,xt
Ct
˜gt(xt)
Underflow
≜
ρtSt,xt
Ct
adversary
− V
Bt>0 on avg.
=⇒ gt(ωt) ≜ E
ρtSt,xt
Ct
− V
Given the loss function and constraints above, we formulate the constrained OCO:
min
ω∈Ω
T
t=1
ft(ω) s.t.
T
t=1
gt(ω) ≤ 0
At every decision epoch t the following events occur in succession:
(a) Agent computes ωt ∈ Ω → xt ∈ arg minx∈X |rx − N
n=1 ωt,nrn|
(b) Adversary decides jointly Ct, ρt, and ˜ft(xt). ˜gt(xt) → actual buffer displacement
(c) Feedback is provided to the agent
Bitrate adaptation via learning 6 / 11
OCO solution for adaptation (Learn2Adapt)
Challenge in adversarial problems: gt(ωt) are unknown at decision time
Solution: prediction around ωt−1 evaluated at ωt (Taylor)
ˆgt(ωt) ≜ gt−1(ωt−1) + ⟨∇gt−1(ωt−1), ωt − ωt−1⟩
We combine the objective and the constraint function in a regularized Lagrangian:
Lt(ω, Q(t)) = Q(t)ˆgt(ω)
Buffer deviation
+ VL
ˆft(ω)
Latency
+ α||ωt − ωt−1||2
Smoothness
• VL and α are convergence parameters
• ||ωt − ωt−1||2 is a regularization term that smooths the decisions
• Q(t) is a multiplier → accumulates the constraint deviations:
Q(t + 1) = [Q(t) + ˆgt(ωt)]+
,
→ We have accounted for latency, smoothness, and stalling, in terms of QoE provisions
Bitrate adaptation via learning 7 / 11
Learn2Adapt-LowLatency (L2A-LL)
→ Learn2Adapt-LL takes a step in the direction of the sub-gradient of the Lagrangian
Algorithm 1 Learn2Adapt-LowLatency (L2A-LL)
Initialize: Q(1) = 0, ω0 ∈ S
Parameters: cautiousness parameter VL, step size α
1: for all t ∈ {1, 2, . . . , T} do
2: ωt = projΩ ωt−1 −
VL∇ft−1(ωt−1)+Q(t)∇gt−1(ωt−1)
2α
3: Q(t + 1) = [Q(t) + ˆgt(ωt)]+
4: end for
Bitrate adaptation via learning 8 / 11
Implementation in dash.js
Open source:
Learn2Adapt-LowLatency (L2A-LL) has been implemented in ‘dash.js’
It is is publicly available for experimentation at Unified Streaming GitHub.
Experimental setup:
5 network profiles: Cascade, Intra-Cascade, Spike, Slow and Fast jitters.
Video sequence (encoded at R = 0.3, 0.6, 1.0 Mbps) organized in fragments
(V = 0.5s)
We allow a variable ‘catch-up’ playback rate pt but left its selection to dash.js.
Implementation location:
Modified dash.js instance now includes L2A-LL + all server and orchestration nodes.
The implemented instance of L2A-LL in dash.js exists in:
‘Learn2Adapt-LowLatency/dash.js/src/streaming/rules/abr/L2ARule.js’
All required player configuration can be found in:
‘Learn2Adapt-LowLatency/dash.js/samples/low-latency/index.html’.
Bitrate adaptation via learning 9 / 11
Experimental evaluation
Table 1: Preliminary experimental results for L2A-LL
Network profile Avg Bitrate (Mbps) Avg Buffer length (s) Latency (s) Stall duration (s) Num. Switches
Cascade 0.58 0.46 1.7 28 10
Intra-Cascade 0.35 0.4 2.5 38 7
Spike 0.59 0.4 1.5 6 3
Slow jitter 0.33 0.47 1.3 7 4
Fast jitter 0.33 0.57 1.2 0.7 2
General remarks:
• Our algorithmic approach allows to adjust QoE priorities
• Modular Lagrangian accounts for: latency, stalling and smoothness
• Of course is all a matter of trade-offs
• L2A-LL achieves relatively high avg. bitrate, and low latency (<2 s)
• L2A-LL adheres to theoretical performance guarantees (Theorem 4.14)
• Verified experimentally (also in the paper)
4
T. Karagkioules et al. Online learning for low-latency bitrate adaptation, MMSys ’20 - Challenges
Bitrate adaptation via learning 10 / 11
Summary
Learn2Adapt-LowLatency:
• is a novel bitrate adaptation algorithm, based on online learning
• does not require modifications according to application type
• requires no statistical assumptions for the channel
• performs well in a wide spectrum of possible scenarios
• robust, due to design principle; its ability to learn
• allows parameter selection according to QoE prioritization
• facilitates effective bitrate adaptation in low-latency mode when combined with
CMAF and chunked transfer
All these properties are significantly relevant to the field of modern HAS
• where OTT providers and broadcasters are continuously expanding their services
• to include more: diverse user classes, network scenarios and streaming applications
Bitrate adaptation via learning 11 / 11
For any questions, feel free to contact me at:
theo@unified-streaming.com
Thank you.
Bitrate adaptation via learning 11 / 11

More Related Content

What's hot

Diseases of ear, nose and throat
Diseases of ear, nose and throatDiseases of ear, nose and throat
Diseases of ear, nose and throat
Binod Chaudhary
 
Wound management
Wound managementWound management
Wound management
Sumer Yadav
 
acquired heart disease
acquired heart diseaseacquired heart disease
acquired heart disease
Engidaw Ambelu
 
Skin ulcers
Skin ulcersSkin ulcers
Skin ulcers
EnFerMeriithhaa !!!
 
Clubbing
ClubbingClubbing
Clubbing
khushyy
 
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
Welcome Cure LLP
 
Cellulitis
CellulitisCellulitis
Cellulitis
Dr Subodh Shah
 
Fever in children by dr Hussein Abdeldayem
Fever in children by dr Hussein AbdeldayemFever in children by dr Hussein Abdeldayem
Fever in children by dr Hussein Abdeldayem
Hussein Abdeldayem
 
Varicose veins
Varicose  veinsVaricose  veins
Varicose veins
Daniel Augustine
 
Abdominal examination byMuhamad Fathy (MD)
Abdominal examination byMuhamad Fathy (MD)Abdominal examination byMuhamad Fathy (MD)
Abdominal examination byMuhamad Fathy (MD)
Muhamad Zaidan
 
Cold Injuries
Cold InjuriesCold Injuries
Cold Injuries
Daniel Licardo
 
Venous ulcer
Venous ulcerVenous ulcer
Venous ulcer
KIST Surgery
 
Bacterial skin infection
Bacterial skin infectionBacterial skin infection
Bacterial skin infectionHabrol Afzam
 
Carbuncle
CarbuncleCarbuncle
Diabetic foot disease ‫‬
Diabetic foot disease ‫‬Diabetic foot disease ‫‬
Diabetic foot disease ‫‬
ismail naameh
 

What's hot (17)

Diseases of ear, nose and throat
Diseases of ear, nose and throatDiseases of ear, nose and throat
Diseases of ear, nose and throat
 
Wound management
Wound managementWound management
Wound management
 
acquired heart disease
acquired heart diseaseacquired heart disease
acquired heart disease
 
Skin ulcers
Skin ulcersSkin ulcers
Skin ulcers
 
Clubbing
ClubbingClubbing
Clubbing
 
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
Are Corns Leaving You Sore? Then Its Time To Go For Homeopathy!
 
Cellulitis
CellulitisCellulitis
Cellulitis
 
Fever in children by dr Hussein Abdeldayem
Fever in children by dr Hussein AbdeldayemFever in children by dr Hussein Abdeldayem
Fever in children by dr Hussein Abdeldayem
 
Varicose veins
Varicose  veinsVaricose  veins
Varicose veins
 
Abdominal examination byMuhamad Fathy (MD)
Abdominal examination byMuhamad Fathy (MD)Abdominal examination byMuhamad Fathy (MD)
Abdominal examination byMuhamad Fathy (MD)
 
Cold Injuries
Cold InjuriesCold Injuries
Cold Injuries
 
Deep Vein Thrombosis
Deep Vein ThrombosisDeep Vein Thrombosis
Deep Vein Thrombosis
 
Venous ulcer
Venous ulcerVenous ulcer
Venous ulcer
 
Bacterial skin infection
Bacterial skin infectionBacterial skin infection
Bacterial skin infection
 
Clubbing
ClubbingClubbing
Clubbing
 
Carbuncle
CarbuncleCarbuncle
Carbuncle
 
Diabetic foot disease ‫‬
Diabetic foot disease ‫‬Diabetic foot disease ‫‬
Diabetic foot disease ‫‬
 

Similar to Online learning for low-latency streaming

Presentation - Dynamic Tuning of RTS Threshold
Presentation - Dynamic Tuning of RTS ThresholdPresentation - Dynamic Tuning of RTS Threshold
Presentation - Dynamic Tuning of RTS ThresholdMohammad Saiful Islam
 
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
Huawei Huang
 
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
Alpen-Adria-Universität
 
UDT
UDTUDT
UDT
lilyco
 
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
Berna Bulut
 
Pushing the limits of Controller Area Network (CAN)
Pushing the limits of Controller Area Network (CAN)Pushing the limits of Controller Area Network (CAN)
Pushing the limits of Controller Area Network (CAN)
RealTime-at-Work (RTaW)
 
Optimizing QoE and Latency of Live Video Streaming Using Edge Computing a...
Optimizing  QoE and Latency of  Live Video Streaming Using  Edge Computing  a...Optimizing  QoE and Latency of  Live Video Streaming Using  Edge Computing  a...
Optimizing QoE and Latency of Live Video Streaming Using Edge Computing a...
Alpen-Adria-Universität
 
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
Minh Nguyen
 
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-StudyDo We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
RealTime-at-Work (RTaW)
 
24-02-18 Rejender pratap.pdf
24-02-18 Rejender pratap.pdf24-02-18 Rejender pratap.pdf
24-02-18 Rejender pratap.pdf
FrangoCamila
 
Cvpr 2018 papers review (efficient computing)
Cvpr 2018 papers review (efficient computing)Cvpr 2018 papers review (efficient computing)
Cvpr 2018 papers review (efficient computing)
DonghyunKang12
 
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
MLconf
 
Variational quantum gate optimization on superconducting qubit system
Variational quantum gate optimization on superconducting qubit systemVariational quantum gate optimization on superconducting qubit system
Variational quantum gate optimization on superconducting qubit system
HeyaKentaro
 
Analytical Modeling of End-to-End Delay in OpenFlow Based Networks
Analytical Modeling of End-to-End Delay in OpenFlow Based NetworksAnalytical Modeling of End-to-End Delay in OpenFlow Based Networks
Analytical Modeling of End-to-End Delay in OpenFlow Based Networks
Azeem Iqbal
 
Congestion control in tcp
Congestion control in tcpCongestion control in tcp
Congestion control in tcp
samarai_apoc
 
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
Steven Wallach
 
LwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
LwTE-Live: Light-weight Transcoding at the Edge for Live StreamingLwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
LwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
Alpen-Adria-Universität
 
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
CSCJournals
 

Similar to Online learning for low-latency streaming (20)

Presentation - Dynamic Tuning of RTS Threshold
Presentation - Dynamic Tuning of RTS ThresholdPresentation - Dynamic Tuning of RTS Threshold
Presentation - Dynamic Tuning of RTS Threshold
 
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
Cost-Efficient Rule Management and Traffic Engineering for Software Defined N...
 
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
LLL-CAdViSE: Live Low-Latency Cloud-based Adaptive Video Streaming Evaluation...
 
UDT
UDTUDT
UDT
 
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
Cross-Layer Design of Raptor Codes for Video Multicast over 802.11n MIMO Chan...
 
Pushing the limits of Controller Area Network (CAN)
Pushing the limits of Controller Area Network (CAN)Pushing the limits of Controller Area Network (CAN)
Pushing the limits of Controller Area Network (CAN)
 
Optimizing QoE and Latency of Live Video Streaming Using Edge Computing a...
Optimizing  QoE and Latency of  Live Video Streaming Using  Edge Computing  a...Optimizing  QoE and Latency of  Live Video Streaming Using  Edge Computing  a...
Optimizing QoE and Latency of Live Video Streaming Using Edge Computing a...
 
UDT
UDTUDT
UDT
 
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
EPIQ'21: Days of Future Past: An Optimization-based Adaptive Bitrate Algorith...
 
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-StudyDo We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
Do We Really Need TSN in Next-Generation Helicopters? Insights From a Case-Study
 
Ijetcas14 400
Ijetcas14 400Ijetcas14 400
Ijetcas14 400
 
24-02-18 Rejender pratap.pdf
24-02-18 Rejender pratap.pdf24-02-18 Rejender pratap.pdf
24-02-18 Rejender pratap.pdf
 
Cvpr 2018 papers review (efficient computing)
Cvpr 2018 papers review (efficient computing)Cvpr 2018 papers review (efficient computing)
Cvpr 2018 papers review (efficient computing)
 
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon Uni...
 
Variational quantum gate optimization on superconducting qubit system
Variational quantum gate optimization on superconducting qubit systemVariational quantum gate optimization on superconducting qubit system
Variational quantum gate optimization on superconducting qubit system
 
Analytical Modeling of End-to-End Delay in OpenFlow Based Networks
Analytical Modeling of End-to-End Delay in OpenFlow Based NetworksAnalytical Modeling of End-to-End Delay in OpenFlow Based Networks
Analytical Modeling of End-to-End Delay in OpenFlow Based Networks
 
Congestion control in tcp
Congestion control in tcpCongestion control in tcp
Congestion control in tcp
 
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
A New Multi-Channel MAC Protocol With On-Demand Channel Assignment For Multi-...
 
LwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
LwTE-Live: Light-weight Transcoding at the Edge for Live StreamingLwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
LwTE-Live: Light-weight Transcoding at the Edge for Live Streaming
 
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
Soft Real-Time Guarantee for Control Applications Using Both Measurement and ...
 

Recently uploaded

UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
DianaGray10
 
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
Neo4j
 
UiPath Test Automation using UiPath Test Suite series, part 5
UiPath Test Automation using UiPath Test Suite series, part 5UiPath Test Automation using UiPath Test Suite series, part 5
UiPath Test Automation using UiPath Test Suite series, part 5
DianaGray10
 
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
James Anderson
 
Uni Systems Copilot event_05062024_C.Vlachos.pdf
Uni Systems Copilot event_05062024_C.Vlachos.pdfUni Systems Copilot event_05062024_C.Vlachos.pdf
Uni Systems Copilot event_05062024_C.Vlachos.pdf
Uni Systems S.M.S.A.
 
20240609 QFM020 Irresponsible AI Reading List May 2024
20240609 QFM020 Irresponsible AI Reading List May 202420240609 QFM020 Irresponsible AI Reading List May 2024
20240609 QFM020 Irresponsible AI Reading List May 2024
Matthew Sinclair
 
PCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase TeamPCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase Team
ControlCase
 
How to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptxHow to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptx
danishmna97
 
Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !
KatiaHIMEUR1
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
DanBrown980551
 
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
SOFTTECHHUB
 
GraphSummit Singapore | The Art of the Possible with Graph - Q2 2024
GraphSummit Singapore | The Art of the  Possible with Graph - Q2 2024GraphSummit Singapore | The Art of the  Possible with Graph - Q2 2024
GraphSummit Singapore | The Art of the Possible with Graph - Q2 2024
Neo4j
 
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
SOFTTECHHUB
 
Essentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FMEEssentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FME
Safe Software
 
Microsoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdfMicrosoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdf
Uni Systems S.M.S.A.
 
Generative AI Deep Dive: Advancing from Proof of Concept to Production
Generative AI Deep Dive: Advancing from Proof of Concept to ProductionGenerative AI Deep Dive: Advancing from Proof of Concept to Production
Generative AI Deep Dive: Advancing from Proof of Concept to Production
Aggregage
 
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
Neo4j
 
Large Language Model (LLM) and it’s Geospatial Applications
Large Language Model (LLM) and it’s Geospatial ApplicationsLarge Language Model (LLM) and it’s Geospatial Applications
Large Language Model (LLM) and it’s Geospatial Applications
Rohit Gautam
 
Monitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR EventsMonitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR Events
Ana-Maria Mihalceanu
 
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
名前 です男
 

Recently uploaded (20)

UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
 
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
GraphSummit Singapore | Enhancing Changi Airport Group's Passenger Experience...
 
UiPath Test Automation using UiPath Test Suite series, part 5
UiPath Test Automation using UiPath Test Suite series, part 5UiPath Test Automation using UiPath Test Suite series, part 5
UiPath Test Automation using UiPath Test Suite series, part 5
 
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
 
Uni Systems Copilot event_05062024_C.Vlachos.pdf
Uni Systems Copilot event_05062024_C.Vlachos.pdfUni Systems Copilot event_05062024_C.Vlachos.pdf
Uni Systems Copilot event_05062024_C.Vlachos.pdf
 
20240609 QFM020 Irresponsible AI Reading List May 2024
20240609 QFM020 Irresponsible AI Reading List May 202420240609 QFM020 Irresponsible AI Reading List May 2024
20240609 QFM020 Irresponsible AI Reading List May 2024
 
PCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase TeamPCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase Team
 
How to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptxHow to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptx
 
Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
 
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
 
GraphSummit Singapore | The Art of the Possible with Graph - Q2 2024
GraphSummit Singapore | The Art of the  Possible with Graph - Q2 2024GraphSummit Singapore | The Art of the  Possible with Graph - Q2 2024
GraphSummit Singapore | The Art of the Possible with Graph - Q2 2024
 
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
Goodbye Windows 11: Make Way for Nitrux Linux 3.5.0!
 
Essentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FMEEssentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FME
 
Microsoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdfMicrosoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdf
 
Generative AI Deep Dive: Advancing from Proof of Concept to Production
Generative AI Deep Dive: Advancing from Proof of Concept to ProductionGenerative AI Deep Dive: Advancing from Proof of Concept to Production
Generative AI Deep Dive: Advancing from Proof of Concept to Production
 
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
GraphSummit Singapore | Graphing Success: Revolutionising Organisational Stru...
 
Large Language Model (LLM) and it’s Geospatial Applications
Large Language Model (LLM) and it’s Geospatial ApplicationsLarge Language Model (LLM) and it’s Geospatial Applications
Large Language Model (LLM) and it’s Geospatial Applications
 
Monitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR EventsMonitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR Events
 
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
 

Online learning for low-latency streaming

  • 1. Online Learning for Low-latency Streaming Grand Challenge on Adaptation Algorithms for Near-Second Latency Theo Karagkioules, Rufael Mekuria, Dirk Griffioen, Arjen Wagenaar Unified Streaming, Amsterdam, Netherlands
  • 3. Motivation HTTP Adaptive Streaming (HAS) is now the dominant source of data traffic. • It is projected to reach 82% of the global Internet traffic, by 20221 • OTT services (Netflix, YouTube etc.) account for > 50% of the global ⇓ traffic2 • Cord-cutters and live applications are expedited → rise of low-latency streaming Some current challenges of bitrate adaptation: • Channel instability (physical phenomena) • QoE management • Reliable low-latency video streaming (1-5 s) . Need for robust, lightweight and universal solutions ← our contribution 1 Cisco Visual Networking Index, “Forecast and Trends, 2017-2022” - Cisco ’19 2 Global Internet Phenomena - Sandvine 19’ Low-latency adaptive streaming 1 / 11
  • 4. Adaptive streaming media format Low-latency bitrate adaptation is facilitated by: • Multiple qualities={resolutions, encoding rates} served from a regular origin. • By requiring commodity web infrastructure → operational costs ↓ • Fragments - smaller (independently decodable) files • Independent fragment requests (multiple servers) maintain session state • Data segmentation → finer granularity of bitrate selection → channel utilization ↑ • A client initially fetches an XML file, which contains content information • Control logic, a.k.a a bitrate adaptation algorithm→ bitrate indicates quality • Logic that selects the appropriate encoding bitrate ∀ fragments, given network conditions • Essentially matches the streaming rate to the channel rate, via means of adaptation • CMAF in combination with the supported chunked transfer mechanism of HTTP 1.1 and beyond, have set the stage for low-latency streaming Low-latency adaptive streaming 2 / 11
  • 6. Motivation Insights from earlier research (MMSys ’17)3: • Existing schemes → parameter tuning according to network or app. scenario • Difficulty in generalizing well beyond a certain scope of usage • Optimal adaptation over fluctuating channels remains a challenging operation • Low-latency bitrate adaptation still relies on throughput estimation To overcome this limitation → resort to learning or control theory • Practical implementation may be hindered • by energy-demanding architectures (Deep Learning) • or by the complexity of exploring the complete optimization space (MDPs) Proposal: Novel adaptation algorithm based on Online Convex Optimization (OCO) • OCO emerged in 2003 (Zinkevich) as a very effective online learning framework • Does not rely on estimations (throughput). Only on historical values. • “Model-free”, as no assumption for the statistical model of the channel is required • Independent of any parameter selection concerning the streaming environment • It provides tractable feasibility and performance guarantees 3 T. Karagkioules et al. A Comparative Case Study of HTTP Adaptive Streaming Algorithms in Mobile Networks Bitrate adaptation via learning 3 / 11
  • 7. Problem definition An adaptation algorithm is an optimization solution with the objective of maximizing the video bitrate, while ensuring stable and continuous low-latency streaming. However, the application of OCO in HAS is not a straightforward task • OCO requires convex decision space and constraints • HAS has a discrete decision space (set of qualities for every fragment) • Instantaneous state-dependent constraints (finite-sized buffer) Modelling assumptions: • We model the adaptive streaming client by a learning agent • Agent minimizes latency, s.t. scheduling constraints of the buffer • We model the channel rate evolution by an adversary • Adversary decides the cost of each decision only after it has been taken • Adversarial setting is general enough to include any potential time-variant distribution • Learning agent minimizes an adversarial loss function, unknown at decision time. We fulfill the OCO requirement (convex decision set and constraints): 1. Agent decides video quality, according to a probability distribution for the appropriate bitrate, given network and latency conditions 2. Relaxation to unbounded buffer → adheres to time-averaging constraints Bitrate adaptation via learning 4 / 11
  • 8. Relaxations Relaxation 1: Convexification of decision set by randomization ∀ fragment t ∈ {1, . . . , T}, the client selects quality xt ∈ X = {1, . . . N} • upon bitrate indication rxt ∈ {r1, . . . , rN} of the bitrate adaptation algorithm Consider the probability simplex: Ω = {ω ∈ RN : ω ≥ 0 ∧ ∥ω∥1 = 1} → Instead of xt, client learns the optimal probability ωt = (ωt,n)n=1,...,N of picking xt ∈ X Relaxation 2: Convexification of buffer constraints Buffer evolution: Bt+1 =           Bt − ρt Playback rate Seg. size≈V·rxt St,xt Ct Throughput ⇓ time           + + V Seg. duration − Inter-request delay ∆t keeps Bt+1<Bmax ∈ [0, Bmax] → We treat the buffer as an infinite queue, i.e. Bt ∈ (−∞, ∞). ∆t is never imposed • We allow instantaneous violation of the buffer budget • But utilize a penalty to maintain 0 < Bt < Bmax on average Bitrate adaptation via learning 5 / 11
  • 9. Problem formulation Consider the following functions =⇒ random processes: ˜ft(xt) ≜ V − ρtSt,xt Ct Loss =⇒ ft(ωt) ≜ E V − ρtSt,xt Ct ˜gt(xt) Underflow ≜ ρtSt,xt Ct adversary − V Bt>0 on avg. =⇒ gt(ωt) ≜ E ρtSt,xt Ct − V Given the loss function and constraints above, we formulate the constrained OCO: min ω∈Ω T t=1 ft(ω) s.t. T t=1 gt(ω) ≤ 0 At every decision epoch t the following events occur in succession: (a) Agent computes ωt ∈ Ω → xt ∈ arg minx∈X |rx − N n=1 ωt,nrn| (b) Adversary decides jointly Ct, ρt, and ˜ft(xt). ˜gt(xt) → actual buffer displacement (c) Feedback is provided to the agent Bitrate adaptation via learning 6 / 11
  • 10. OCO solution for adaptation (Learn2Adapt) Challenge in adversarial problems: gt(ωt) are unknown at decision time Solution: prediction around ωt−1 evaluated at ωt (Taylor) ˆgt(ωt) ≜ gt−1(ωt−1) + ⟨∇gt−1(ωt−1), ωt − ωt−1⟩ We combine the objective and the constraint function in a regularized Lagrangian: Lt(ω, Q(t)) = Q(t)ˆgt(ω) Buffer deviation + VL ˆft(ω) Latency + α||ωt − ωt−1||2 Smoothness • VL and α are convergence parameters • ||ωt − ωt−1||2 is a regularization term that smooths the decisions • Q(t) is a multiplier → accumulates the constraint deviations: Q(t + 1) = [Q(t) + ˆgt(ωt)]+ , → We have accounted for latency, smoothness, and stalling, in terms of QoE provisions Bitrate adaptation via learning 7 / 11
  • 11. Learn2Adapt-LowLatency (L2A-LL) → Learn2Adapt-LL takes a step in the direction of the sub-gradient of the Lagrangian Algorithm 1 Learn2Adapt-LowLatency (L2A-LL) Initialize: Q(1) = 0, ω0 ∈ S Parameters: cautiousness parameter VL, step size α 1: for all t ∈ {1, 2, . . . , T} do 2: ωt = projΩ ωt−1 − VL∇ft−1(ωt−1)+Q(t)∇gt−1(ωt−1) 2α 3: Q(t + 1) = [Q(t) + ˆgt(ωt)]+ 4: end for Bitrate adaptation via learning 8 / 11
  • 12. Implementation in dash.js Open source: Learn2Adapt-LowLatency (L2A-LL) has been implemented in ‘dash.js’ It is is publicly available for experimentation at Unified Streaming GitHub. Experimental setup: 5 network profiles: Cascade, Intra-Cascade, Spike, Slow and Fast jitters. Video sequence (encoded at R = 0.3, 0.6, 1.0 Mbps) organized in fragments (V = 0.5s) We allow a variable ‘catch-up’ playback rate pt but left its selection to dash.js. Implementation location: Modified dash.js instance now includes L2A-LL + all server and orchestration nodes. The implemented instance of L2A-LL in dash.js exists in: ‘Learn2Adapt-LowLatency/dash.js/src/streaming/rules/abr/L2ARule.js’ All required player configuration can be found in: ‘Learn2Adapt-LowLatency/dash.js/samples/low-latency/index.html’. Bitrate adaptation via learning 9 / 11
  • 13. Experimental evaluation Table 1: Preliminary experimental results for L2A-LL Network profile Avg Bitrate (Mbps) Avg Buffer length (s) Latency (s) Stall duration (s) Num. Switches Cascade 0.58 0.46 1.7 28 10 Intra-Cascade 0.35 0.4 2.5 38 7 Spike 0.59 0.4 1.5 6 3 Slow jitter 0.33 0.47 1.3 7 4 Fast jitter 0.33 0.57 1.2 0.7 2 General remarks: • Our algorithmic approach allows to adjust QoE priorities • Modular Lagrangian accounts for: latency, stalling and smoothness • Of course is all a matter of trade-offs • L2A-LL achieves relatively high avg. bitrate, and low latency (<2 s) • L2A-LL adheres to theoretical performance guarantees (Theorem 4.14) • Verified experimentally (also in the paper) 4 T. Karagkioules et al. Online learning for low-latency bitrate adaptation, MMSys ’20 - Challenges Bitrate adaptation via learning 10 / 11
  • 14. Summary Learn2Adapt-LowLatency: • is a novel bitrate adaptation algorithm, based on online learning • does not require modifications according to application type • requires no statistical assumptions for the channel • performs well in a wide spectrum of possible scenarios • robust, due to design principle; its ability to learn • allows parameter selection according to QoE prioritization • facilitates effective bitrate adaptation in low-latency mode when combined with CMAF and chunked transfer All these properties are significantly relevant to the field of modern HAS • where OTT providers and broadcasters are continuously expanding their services • to include more: diverse user classes, network scenarios and streaming applications Bitrate adaptation via learning 11 / 11
  • 15. For any questions, feel free to contact me at: theo@unified-streaming.com Thank you. Bitrate adaptation via learning 11 / 11