The document discusses research on using multi-access edge computing (MEC) to improve adaptive video streaming. It presents several contributions, including developing a MEC and HAS simulator called ANGELA, proposing dynamic segment repackaging at the edge to increase bandwidth savings, and designing edge-assisted adaptation schemes (EADAS and ECAS-ML) that leverage edge resources to guide client-based ABR algorithms and improve QoE and fairness. It also investigates segment prefetching policies at the edge such as ones based on last quality, Markov models, transrating, machine learning, and super-resolution.
Multi-access Edge Computing for Adaptive Video Streaming
1. Multi-access Edge
Computing for Adaptive
Video Streaming
Jesús Aguilar Armijo
Klagenfurt, 10/07/2023
Supervisor: Prof. Hermann Hellwagner
Co-supervisor: Prof. Christian Timmerer
3. Motivation
Video streaming is the most widely used service in mobile networks, with a
69%* of traffic data share in 2022.
3
*Ericsson. Ericsson Mobility Report. https:/
/www.ericsson.com/en/reports-and-papers/mobility-report/reports/june-2022, 2022.
Edge computing is a key technology for 5G networks that enables
storage and computing power close to the final clients.
This thesis aims to investigate how the video streaming process can
be improved with edge computing assistance.
Client Client Client Client
Edge
Edge
Server
QoE
5G
Fairness
ML
Latency
HAS
4. Multi-access Edge Computing
Multi-access Edge Computing (MEC) has several
important characteristics:
● Owned by the mobile network operator
⤍ Third-party applications
● Computation power and storage available
⤍ Mechanisms deployment
● Closer location to the client
⤍ Low latency (10-50 ms)
● Access to radio metrics through Radio Network
Information Service (RNIS)
⤍ Better adaptation to radio conditions
● Access to all clients’ requests
⤍ Improve fairness
4
MEC architecture
5. #2 - How can Radio Access Networks (RAN)
and HAS client context information enhance
bandwidth resource management?
The edge is aware of each client's requirements and
radio network metrics. It can coordinate the video
streaming process among all the clients improving the
QoE and fairness.
#4 - How can collaboration between different
edge computing nodes assist the video
streaming process?
Storage and computing resources can be shared
among different edge computing nodes.
#1 - How can network assistance for HTTP
Adaptive Streaming (HAS) be realized by
MEC functionality?
New mechanisms that leverage MEC functionalities
can be created and deployed at the edge nodes to
guide the video streaming process.
#3 - How to perform radio metrics and
client behavior analysis and predictions to
improve content delivery?
The computing power available at the edge can be
used to predict future content requests and
network conditions with machine learning
techniques.
5
Research Questions
6. Research Overview
Assessing the lack of tools to test edge
computing mechanisms for adaptive
multimedia with the creation of a MEC and
HAS simulator.
6
ANGELA: HTTP Adaptive Streaming and Edge
Computing Simulator¹
Performing repackaging of media delivery
formats at the edge to increase bandwidth
savings and improve cache efficiency.
Dynamic Segment Repackaging at the Edge
for HTTP Adaptive Streaming²
1) Jesús Aguilar Armijo, Christian Timmerer, and Hermann Hellwagner. "ANGELA: HTTP Adaptive Streaming and Edge Computing Simulator." In 10th IFIP International
Conference on Performance Evaluation and Modeling in Wireless and Wired Networks (PEMWN), 2021.
2) Jesús Aguilar Armijo, Babak Taraghi, Christian Timmerer, and Hermann Hellwagner. "Dynamic Segment Repackaging at the Edge for HTTP Adaptive Streaming." In
IEEE International Symposium on Multimedia (ISM), 2020.
7. Research Overview
Moving the ABR logic to the edge to perform
better adaptive decisions and distribute the
content according to each device’s
requirements, achieving better QoE and
fairness.
7
EADAS: Edge Assisted Adaptation Scheme for
HTTP Adaptive Streaming3
ECAS-ML: Edge Computing Assisted
Adaptation Scheme with Machine Learning
for HTTP Adaptive Streaming4
3) Jesús Aguilar Armijo, Christian Timmerer, and Hermann Hellwagner. "EADAS: edge assisted adaptation scheme for HTTP adaptive streaming." In IEEE 46th Conference
on Local Computer Networks (LCN), 2021.
4) Jesús Aguilar Armijo, Ekrem Çetinkaya, Christian Timmerer, and Hermann Hellwagner. "ECAS-ML: Edge Computing Assisted Adaptation Scheme with Machine
Learning for HTTP Adaptive Streaming." In MultiMedia Modeling: 28th International Conference (MMM), 2022
8. Research Overview
Different segment prefetching techniques
with different usage of amounts and types of
resources are proposed, analyzed and
compared.
8
Segment Prefetching at the Edge for Adaptive
Video Streaming5
SPACE: Segment Prefetching and Caching at
the Edge for Adaptive Video Streaming6
5) Jesús Aguilar Armijo, Christian Timmerer, and Hermann Hellwagner. "Segment Prefetching at the Edge for Adaptive Video Streaming." In 18th International Conference
on Wireless and Mobile Computing, Networking and Communications (WiMob), 2022.
6) Jesús Aguilar Armijo, Christian Timmerer, and Hermann Hellwagner. "SPACE: Segment Prefetching and Caching at the Edge for Adaptive Video Streaming." IEEE
Access, 2023.
9. ANGELA: HTTP Adaptive Streaming and Edge Computing Simulator
● ANGELA focuses on testing the performance of edge computing mechanisms to support video
streaming.
● Written in Python.
● Lightweight.
● Can be executed on Linux, Windows and Mac OS.
● Imports radio layer data to reduce the simulation time drastically without losing accuracy
○ The focus is on high layers, not on low layers.
○ Simulating a simple scenario with one client connected to a 4G network consuming video
streaming services:
■ ANGELA imports real 4G radio traces from a dataset ⤍ 2.1 seconds
■ ns-3 uses its LTE model ⤍ 764 seconds
9
10. ANGELA: HTTP Adaptive Streaming and Edge Computing Simulator
● MEC mechanisms have access to:
○ Radio metrics through RNIS
○ Player metrics
○ All clients’ requests
● Different ABR algorithms are available.
● Performance evaluation:
○ Customizable metrics such as mean bitrate, fairness index,
segment switches, stalls, etc.
○ QoE according to ITU-T P.1203 evaluation model*
● ANGELA is used for obtaining the results of the contributions:
○ Edge Assisted Adaptation Schemes
○ Segment Prefetching and Caching at the Edge
10
ANGELA architecture
*Steve Göring, Werner Robitza et al., “ITU-T Rec. P.1203 Standalone Implementation,” https:/
/github.com/itu-p1203/itu-p1203
11. Dynamic Segment Repackaging at the Edge for HTTP Adaptive Streaming
11
Classic scenario with four media formats
Credit: Akamai
Credit: Akamai
CMAF scenario Scenario of the proposed solution
● Using CMAF in the backhaul provides
bandwidth savings, less storage needed and
more cache efficiency.
● CMAF is converted to the desired media format
on the fly at the edge.
● The proposed solution reduces delivery latency
if the edge has more than 1.64 times the
compute power per segment than the origin
server.
12. Edge Assisted Adaptation Schemes
12
EADAS ECAS-ML
● On-the-fly
● Deployed at the
edge
● Lightweight
● Intercepts clients’
requests
● Player metrics
● Radio metrics
● Support
client-based ABR
algorithms
● Focus on QoE and
fairness
● Clustering per
subscription
● Substitute
client-based ABR
algorithms
● Focus on QoE
● Consider screen
display
● LSTM-based model
13. EADAS: Edge Assisted Adaptation Scheme for HTTP Adaptive Streaming
● Improves client-based ABR algorithms’ decisions.
● Compatible with segment prefetching (solves radio throughput miscalculations).
● The 𝛼 value can prioritize QoE or fairness according to content provider’s preferences:
○ Lower 𝛼 values prioritize fairness, higher alpha values prioritize QoE.
○ final score = 𝛼 ✕ quality score + (1 - 𝛼 ) ✕ fairness score
13
EADAS performance with different 𝛼 values
14. EADAS: Edge Assisted Adaptation Scheme for HTTP Adaptive Streaming
● Testing throughput-based (TBA)¹, buffer-based (BBA)², and hybrid-based (SARA)3
ABR algorithms
with and without EADAS support.
● Mean values of 25 clients, real 4G traces with different mobility patterns.
● EADAS improves the QoE and the fairness for three different types of ABR algorithms:
client-based, throughput-based and hybrid.
14
1) Duc V. Nguyen, Hung T. Le, Pham Ngoc Nam, Anh T. Pham, and Truong Cong Thang. "Adaptation method for video streaming over HTTP/2." In IEICE communications
Express, 2016.
2) Te-Yuan Huang, Ramesh Johari, Nick McKeown, Matthew Trunnell, and Mark Watson. "A buffer-based approach to rate adaptation: Evidence from a large video
streaming service." In Proceedings of the ACM conference on SIGCOMM, 2014.
3) Parikshit Juluri, Venkatesh Tamarapalli, and Deep Medhi. "SARA: Segment aware rate adaptation algorithm for dynamic adaptive streaming over HTTP." In IEEE
International Conference on Communication Workshop (ICCW), 2015.
EADAS performance evaluation
15. ECAS-ML: Edge Assisted Adaptation Scheme with ML for HAS
15
● Manage the tradeoff among bitrate, segment switches and stalls with four parameters:
○ Switches penalty factor
○ Stalls penalty factor
○ Buffer threshold 1
○ Buffer threshold 2
● Considers screen display, often ignored.
● A Long Short-Term Memory network (LSTM) model predicts the best set of parameters for the
current radio network conditions.
Risk areas of the player buffer in ECAS-ML
16. ECAS-ML: Edge Assisted Adaptation Scheme with ML for HAS
16
● Example of ECAS-ML buffer behavior:
○ High-risk area: from 0 seconds to 6 seconds.
○ Medium-risk area: from 6 seconds to 12 seconds.
○ Low-risk area: from 12 seconds to 20 seconds (maximum buffer size).
● It provides the highest quality that maintains the buffer level above the first threshold,
preventing stalls even when in unstable bandwidth scenarios.
17. ECAS-ML: Edge Assisted Adaptation Scheme with ML for HAS
17
● Testing BBA, TBA, SARA, GBBA*, SARA-EADAS and ECAS-ML.
● Mean values of 25 clients, real 4G traces with different mobility patterns.
● ECAS-ML outperforms other client-based and edge-based ABR algorithms.
* Minsu Kim and Kwangsue Chung. "Edge computing assisted adaptive streaming scheme for mobile networks." IEEE Access, 2020
ECAS-ML performance evaluation
18. Segment Prefetching and Caching at the Edge
● Segment prefetching transmits the future video
segments in advance closer to the client to serve
content with lower latency.
● Different policies with different metrics and
resource utilization are proposed and evaluated.
● Architecture:
○ Layer 1: CDN, storing the video segments encoded
in different representations.
○ Layer 2: Control edge computing nodes
coordinating layer 3 nodes with a global indexing
table.
○ Layer 3: Edge computing nodes co-located with a
base station serving the segment to the clients.
○ Layer 4: Clients’ devices.
18
Proposed architecture
19. Segment Prefetching and Caching at the Edge
● Prefetching Based on Last Segment Quality:
○ Last segment quality (LSQ).
○ Last segment quality plus (LSQ+).
○ All segment qualities (ASQ).
● Prefetching Based on a Markov Model:
○ Each client has its own Markov state matrix
continuously updated with each request.
○ Prefetches the quality with the highest
probability of being requested.
○ Matrix example with four qualities:
M =
19
Markov finite state machine example with four qualities
69 11 3 3
27 73 6 1
3 14 81 6
1 2 10 90
20. Segment Prefetching and Caching at the Edge
● Prefetching Based on Transrating:
○ A high-quality representation is sent to the
edge.
○ Then, the segment is converted to the desired
quality.
○ The extra delay caused by the live transrating
can potentially be compensated by:
■ The delay saved for delivering from the
edge node
■ The player buffer
○ This policy assures a 100% hit rate, i.e., the
requested quality from the client will always be
served directly from the edge node.
20
Measured transrating times for 4-sec. segments (ms)
21. Segment Prefetching and Caching at the Edge
● Prefetching Based on Machine Learning
○ Predict future segment requests of ABR
algorithms.
○ Four different input metrics to make
predictions are chosen:
■ Buffer size
■ Link bitrate
■ Previous quality
■ Previous link bitrate
○ Custom dataset with different
combinations of bitrate ladders and ABR
algorithms (TBA, BBA, SARA).
○ 5 different Machine Learning models:
Random Forest, Gradient Boost,
AdaBoost, Decision Trees and Extremely
Randomized Trees.
21
Prediction evaluation of different ML models
Bitrate ladders utilized to conduct the experiments
22. Segment Prefetching and Caching at the Edge
● Prefetching Based on Super-resolution:
○ Send a low-quality representation to the edge, then perform super-resolution to the requested quality.
○ Clients can consume a high-quality representation with reduced backhaul link bitrate utilization.
○ Trade-off between link bitrate consumption and the visual quality of the representation.
○ Three operations:
(1) Decode the segment
(2) Perform the super-resolution technique to upgrade the resolution
(3) Encode the segment to be sent to the client
22
Super-resolution computation times for 4-seconds segment
(1) (2) (3)
23. Segment Prefetching and Caching at the Edge
23
Performance of different segment prefetching policies
24. Segment Prefetching and Caching at the Edge
● The best segment prefetching policy depends on the service provider’s preferences and
resources:
24
Straightforward
implementation with low
resource utilization
LSQ and Markov-based
Prioritize QoE at the expense
of costs
ASQ, LSQ+ and
Transrating-based
Balance between
performance and costs
ML-based
● Prefetching policies can adapt dynamically to the current resource capabilities.
● Different policies can coexist within the same edge node:
○ Better policies for premium clients
25. Contributions and Concluding Remarks
● This thesis investigates and develops novel approaches and tools to improve the content delivery
for adaptive video streaming services using multi-access edge computing
● Several contributions have been done during the course of this thesis:
25
Dynamic Segment Repackaging at the Edge for HTTP Adaptive
Streaming
IEEE International Symposium on Multimedia (ISM), 2020
EADAS: Edge Assisted Adaptation Scheme for HTTP Adaptive
Streaming
Proceedings of the IEEE Conference on Local Computer Networks
(LCN), 2021
ANGELA: HTTP Adaptive Streaming and Edge Computing
Simulator
Proceedings of the IFIP International Conference on Performance
Evaluation and Modeling in Wireless and Wired Networks
(PEMWN), 2021
ECAS-ML: Edge Computing Assisted Adaptation Scheme with
Machine Learning for HTTP Adaptive Streaming
Proceedings of the International Conference on Multimedia
Modeling (MMM), 2022
Segment Prefetching at the Edge for Adaptive Video Streaming
Proceedings of the International Workshop on Selected Topics in
Wireless and Mobile computing (STWiMob), 2022
SPACE: Segment Prefetching and Caching at the Edge for Adaptive
Video Streaming
IEEE Access, 2023
26. Contributions and Concluding Remarks
● ANGELA: HTTP Adaptive Streaming and Edge Computing Simulator
○ Research question #1
○ Testing of edge computing mechanisms to assist HAS.
○ Accurate radio layer simulation while providing lower simulation time.
■ Time decrease of 99.76% in ANGELA compared with ns-3.
○ This simulator was used in several publications to test edge mechanisms.
● Dynamic Segment Repackaging at the Edge for HTTP Adaptive Streaming
○ Research question #1
○ CMAF in the backhaul and performs an on-the-fly conversion to the requested media content at
the edge.
○ It improves cache efficiency and increases bandwidth savings.
○ It reduces delivery latency if the edge has more than 1.64 times the compute power per segment
than the origin server.
26
27. Contributions and Concluding Remarks
● Edge Assisted Adaptation Schemes
○ Research question #2 (EADAS and ECAS-ML) and #3 (ECAS-ML)
○ Leverage edge computing resources and available radio, player and network metrics to perform
better adaptation decisions.
○ EADAS and ECAS-ML work on-the-fly, are lightweight, and incur very low additional latency
○ EADAS:
■ Support client-based ABR algorithms.
■ Improves both QoE and fairness among clients.
■ EADAS improved performance of buffer-based, throughput-based and hybrid ABR algorithms.
○ ECAS-ML:
■ Manage the tradeoff among bitrate, segment switches and stalls to achieve high QoE.
■ ML techniques are used to adapt the algorithm parameters to the current radio network
conditions.
■ ECAS-ML outperforms in terms of QoE popular ABR algorithms like BBA by 13.8% or other
edge-based schemes such as GBBA by 19.29%.
27
28. Contributions and Concluding Remarks
● Segment Prefetching and Caching at the Edge
○ Research question #3 and #4
○ Different segment prefetching policies with different resource utilization are proposed, analyzed and
evaluated.
○ These policies use different techniques such as a Markov prediction model, transrating,
super-resolution and machine learning.
○ The ML-based policy, regarding the baseline simulations without prefetching:
■ Increases the average bitrate by ≈46%.
■ Reduces the average number of stalls by ≈20%.
■ Increasing the extra bandwidth consumption only by ≈6%.
○ Other prefetching policies offer a different combination of performance enhancement and resource
usage that can adapt to the service provider’s needs.
28
31. EADAS: Edge Assisted Adaptation Scheme for HTTP Adaptive Streaming
31
● The quality index varies from 0 (the lowest) to 19 (the highest).
● EADAS improves the mean quality index and the mean segment bitrate.
● It prevents many stalls and drops in segment quality when the radio throughput decreases
suddenly, e.g., between seconds 200 and 400.
32. Segment Prefetching and Caching at the Edge
32
● Least Recently Used (LRU)
● Least Frequently Used (LFU)
● Least Popular Used (LPU)
33. Future Work
● Security and DRM
○ Encryption and decryption at the edge can be used to increase network bandwidth savings
○ For example, it is possible to reduce the packet header in a trusted link (e.g., the networks’ backhaul)
by encrypting it just when the packet arrives at the edge node before sending it through the radio
link, which is less trustworthy.
● Per-Segment QoE Assessment
○ The QoE is usually assessed at the end of the video streaming session with QoE models such as
VMAF or ITU-T P.1203.
○ However, computing per segment the future QoE that a specific quality representation will provide is
a worthwhile task that can lead to better adaptation decisions, i.e., better final QoE.
● ML Techniques to Predict Future Network Conditions
○ The cellular networks’ workload follows some patterns over time. Cells located in an office area serve
more clients during working hours, i.e., in the morning and afternoon. The opposite happens for
cells located in residential areas.
○ ML techniques can leverage these trends to predict future network conditions. Specifically,
predicting the future bandwidth for each client is targeted to perform better adaptation decisions in
video streaming
33
34. Simulation configuration (i)
● Dynamic Segment Repackaging at the Edge for HTTP
Adaptive Streaming
○ The encoding, segmentation, transmuxing,
encryption and decryption operations were
measured
○ Tools used: ffmpeg (encoding, segmentation and
transmuxing) and Bento4 (encryption and decryption)
○ The platform used for running these experiments is a
regular PC with Intel Core i7-9750H 2.60GHz, 16GB
RAM, 64-bit Windows 10.
○ Big Buck Bunny
○ libx265 to perform the HEVC encoding into four
different resolutions: 3840x2160, 1920x1080, 1280x720
and 640x360 pixels.
○ The bitrates are 8295, 2456, 815 and 354 kbps,
respectively. The video has 24 frames per second and
a total duration of 596 seconds.
34
35. Simulation configuration (ii)
● Edge Assisted Adaptation Schemes
○ ANGELA simulator
○ Its architecture consist of (i) a video streaming server, (ii) an edge computing node
co-located with the base station, and (iii) multiple clients.
○ Real traces from 4G radio dataset*
○ 25 clients
○ For the backhaul link that connects the video server and the edge computing node, we
assume 1 Gbps bandwidth and 50 ms of latency
○ Big Buck Bunny, Elephants Dream, and Of Forest and Men videos, with a segment length of
two seconds
○ Bitrates levels of [50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 900, 1200, 1500, 2000, 2500,
3000, 4000, 5000, 6000, 8000] kbps in resolutions ranging from 320x240 to 1920x1080.
35
* D. Raca, J. J. Quinlan, A. H. Zahran, and C. J. Sreenan, “Beyond throughput: a 4G LTE dataset with channel and context metrics,” in Proc. 9th ACM Multimedia Systems
Conference, 2018, pp. 460–465.
36. Simulation configuration (iii)
● Segment Prefetching and Caching at the Edge
○ ANGELA simulator
○ Real traces from 4G radio dataset*
○ We simulate a CDN connected to a control edge computer of layer 2 serving four edge computers of
layer 3, each serving 25 clients, i.e., 100 clients in total.
○ The control edge computing node has a caching capacity of 2 GB, while the edge computing nodes of
layer 3 have a caching capacity of 1 GB each.
○ The CPRI link connecting edge computing nodes has a link bitrate of 1000 Mbps.
○ We transmit 225 segments with a four-second duration each. We utilize the representations of a
multi-codec DASH dataset**
○ We utilize three different bitrate ladders:
36
* D. Raca, J. J. Quinlan, A. H. Zahran, and C. J. Sreenan, “Beyond throughput: a 4G LTE dataset with channel and context metrics,” in Proc. 9th ACM Multimedia Systems
Conference, 2018, pp. 460–465.
** A. Zabrovskiy, C. Feldmann, and C. Timmerer, “Multi-codec DASH dataset,” in Proceedings of the 9th ACM Multimedia Systems Conference, 2018, pp. 438–443