In-network quality optimization for adaptive video streaming services
Summary (4 min read)
III. FORMAL PROBLEM DESCRIPTION
- Providers are exploring how they can offer VoD HAS services next to traditional TV services over their managed network environment.
- HAS services offer the same content at multiple qualities, each at their corresponding rate.
- This allows providers to perform QoE management by adjusting each sessions quality level, based on the current network utilization.
- At peak times, the consequences of an inadequate amount of resources in the network, can thus be anticipated by reducing the quality of individual streaming sessions, while still allowing admittance of all users.
- The set of clients in the service delivery tree for which the VoD traffic traverses node n ∈.
A. Definition of variables and assumptions
- Figure 1 gives an overview of the problem variables and assumptions.
- Note that typical access networks are using a logical tree for their delivery, although the underlying physical network is not a tree due to replication concerns.
- In summary, Table I lists the symbols introduced throughout this section.
B. Integer Linear Programming formulation
- The problem consists of maximizing the QoE over all clients c ∈ C, while adhering to the edge bandwidth constraints.
- The authors use this approximation of the maximum achievable throughput to limit the aggregated allocated rate of the different clients: EQUATION.
- Since this function is differentiable and strictly concave, it has only one maximum, which is therefore also the global maximum.
- The following weighted sum is used to model the impact on switching behavior, where µ represents the average quality, σ introduces a penalty for quality switching and α s represents a weighing factor used to emphasize either the impact of quality or the switching behavior: EQUATION.
- Since the decision variables a c,q are binary variables, the calculation of the objective function can be simplified by calculating µ c,q and σ c,q for each client c and it's associated quality range Q c .
B. Distributed ILP
- The number of constraints for the centralized ILP grows with an increasing depth of the service delivery topology tree.
- First, each node only needs to have local information on the properties of the upstream edge e n − and the video flows for clients C n traversing this node.
- First, since the distributed optimization is only performed when a client joins or leaves the service delivery network, only the proxies p ∈ P c on the delivery path for client c are required to perform local optimization.
- Second, the solutions determined by the predecessors of n are optimal and since these solutions are independent, their combination is optimal.
- If there are 1000000 clients in the network and k is equal to 10, then the total communication overhead is 16.5 MB per optimization.
C. Relaxed Distributed Linear Program (LP) Formulation
- Solving the distributed ILP optimally in a single node can however lead to execution times in the order of seconds when the number of VoD flows crossing that node becomes large.
- This relaxation can be solved in polynomial time but at the cost of optimality.
- The variables a c,q do not longer unambiguously define which quality each client is allowed to download, therefore a heuristic is required to transform the optimal floating point solution into an integer solution.
- First, the clients of the solution matrix A are sorted according to two criteria: first on the proximity of the floating point solution to the integer solution and subsequently on the contribution to the objective (line6).
- This assures that the limitations of the successors N n + are not violated which could lead to an infeasible solution further down the delivery tree.
A. Experiment Setup
- A VoD HAS scenario was implemented by using an NS3 based simulation framework, capable of the transmission of HAS video [25] .
- Furthermore, an additional client heuristic was implemented, which downloads each segment using the QoE management quality decision and checks if these decisions are feasible, given the measured throughput at the client.
- If the measurements indicate that the proposed quality is not achievable, the proposed client heuristic will select the highest sustainable quality based on the local throughput measurements.
- The configured congestion window allows transmitting segments at a rate that is two times bigger than the maximum bitrate of the stream.
- Table II gives an overview of the different quality layers, their associated bitrates, average Peak Signal-to-noise Ratio (PSNR) and Structural Similarity (SSIM) values.
B. Implementation details
- The IBM CPLEX 11 solver was used to implement and solve the proposed binary ILP-problems for both the centralized and distributed algorithm, as well as the relaxed distributed LPproblem.
- The authors executed the different experiments using two modes: Delayed ensuring that the configurations only become available when optimization is finished and Optimal which is agnostic to execution times and installs the configuration immediately.
- This also allows us to preempt a QoE optimization when additional requests lead to a changed environment and the optimal solution would be outdated.
- The heuristic optimization checks if the previous limitations in combination with the additional client are feasible for each edge e, if not the client qualities for c ∈ C e are lowered by one level until the solution is feasible again.
C. Evaluation Details
- The performance of the centralized ILP, distributed ILP and relaxed distributed LP was evaluated in terms of service assurance, quality delivery and oscillations.
- Also the impact of the different approaches on the decision time was quantified.
- The network size, the number of bottlenecks, the optimization objective, Round Trip Time (RTT) and number of servers were varied.
- The authors refer to the centralized and distributed ILP optimization as Centralized Exact and Distributed Exact respectively, while the relaxed optimization is indicated as Distributed Relaxed.
- Therefore these results are installed with a delay and are referred to as Delayed decisions.
D. Impact of Number of Clients
- The authors motivate the deployment of in-network quality adaptation algorithms for HAS delivery networks and quantify the impact of the delivery tree size on the in-network adaptation performance.
- The results show a significant improvement on the average played bitrate over traditional client-based heuristics ranging from 14% to 23%, while the number of switches can be reduced with a factor of 1.5 to 5.
- In-network quality adaptation is able to react more quickly to changing network environments and allows to fully utilize the available bandwidth.
- The average number of switches is slightly higher for the Centralized Exact Optimal optimization when compared to the Distributed Relaxed heuristic.
- Second, also the number of quality switches can be significantly reduced.
E. Impact of Number of Bottlenecks
- Figure 4 confirms a linear increase in execution time for both the Exact and Relaxed optimization with an increasing number of bottlenecks.
- The Centralized Exact optimization however, takes 300ms to execute, even in the absence of a bottleneck, while the Distributed optimization is only performed when the configuration assigning maximum quality to each client becomes infeasible, leading to an execution time of on average 20ms, consisting solely out of the delay introduced by forwarding the local solutions.
F. Impact of Optimization Objective
- An operator can optimize different policies when offering a HAS streaming service such as maximizing the total bitrate over all streams (Equation ( 8)), maximizing the proportional fairness across the streaming sessions (Equation ( 10)) or maximizing the QoE as a weighted sum of the total bitrate and bitrate variations (Equation ( 14)).
- Optimizing the total bitrate allocation is able to achieve a fairness index closer to 1 then AVC MSS, indicating a fairer distribution of the available throughput among the clients.
- When optimizing for proportional fairness, the innetwork optimization is able to increase the fairness index at the cost of slightly decreasing the average quality as indicated in Figure 5 (b) and 5(d) and increasing the average number of switches from 18 to 26.
- This indicates the trade-off between maximizing fairness and total bitrate allocation.
- As Figure 5 (c) shows, the innetwork adaptation is able to reduce the number of quality switches compared to AVC MSS for all optimization goals.
G. Impact of Delay
- The in-network optimization uses an approximation of the achievable throughput as an upper bound for each link.
- This slight performance decrease can be attributed to the fact that an approximation of the achievable throughput is used.
- As discussed in previous work [12] , HAS quality decreases quickly when RTT's increase due to the subsequent download-request cycles.
- Since the Distributed optimization requires a bottom up propagation of intermediary solutions, the decision times are also impacted by increasing delays.
- Figure 6(b) shows that for increasing delay, the performance of the innetwork decisions slightly decreases when compared to the optimal decision, due to the network delay increasing the decision time.
H. Impact of Multiple Servers
- The authors modified the Distributed optimization to also support this type of topologies by first performing two types of bottomup optimizations, one taking no limitations as input and a second optimization taking into account the limitations of performing a local optimization between Gateway and servers first.
- The content items were then assigned to the different server instances to evenly distribute the load among them.
- Adding rate shaping at the server, allows increasing the quality for AVC MSS when the number of servers increases.
- The Distributed approach is not able to achieve the same performance as the Centralized optimization.
- The Distributed approach even further decreases the average number of switches to 3.3, but at the cost of reduced quality compared to the Centralized optimization, as was mentioned before.
VI. CONCLUSION
- The authors proposed an in-network QoE management for VoD HTTP Adaptive Streaming in a managed network environment.
- The adaptation algorithms enable the network providers to control the quality selection at the client.
- This allows them to increase the average played quality with at least 14% compared to traditional client-based heuristics.
- The authors also discussed different variants of the in-network QoE management: an optimal Centralized ILP, a Distributed ILP and a relaxation of the Distributed algorithm.
- The impact of the number of clients, the Absolute Gap for the integer optimization and the number of bottlenecks were quantified.
Did you find this useful? Give us your feedback
Citations
96 citations
95 citations
Additional excerpts
...in terms of bitrate per client (20% higher on average), initial buffer delay (≈ 15%-20% smaller) and Jain’s fairness index [169]....
[...]
88 citations
Cites background from "In-network quality optimization for..."
...With regards to cross-layer, QoE-driven approaches, we note [13, 16, 37, 75, 103, 128]....
[...]
...Given the inherent bene ts when performing application-aware network management and network-aware application management [118, 151], studies have shown that further potential lies in integrated and cross-layer QoE management approaches [16, 69, 75], stemming from various forms of cooperative agreements and information exchange between involved stakeholders [4, 65]....
[...]
83 citations
Cites background or methods from "In-network quality optimization for..."
...[3] propose quality optimization considering in-network elements into their topology in order to adjust the client’s bitrate according to the available bandwidth on multiple bottleneck links....
[...]
...In the QoE-fairness problem formulation, we follow the discrete time slotted DASH scheduling operation [3] with fixed time duration for each time slot....
[...]
...We follow the discrete time slotted DASH scheduling [3] with total number of jT j time slots and the duration of each slot Dt seconds....
[...]
...We use the Jain’s fairness index [3] which is defined as JF 1⁄4 ðPi riÞ(2)=ðS Pi r(2)i Þ; where S is the total number of clients and ri denotes the average bitrate of client i during its streaming session....
[...]
75 citations
Cites background from "In-network quality optimization for..."
...Hence, according to [153] and [154], the support for coordinated management and global optimisation is imperative....
[...]
References
40,609 citations
5,566 citations
4,476 citations
"In-network quality optimization for..." refers methods in this paper
...To quantitatively evaluate the fairness degree of the different optimization schemes, the Jain’s Fairness Index is used [36]....
[...]
3,067 citations
2,145 citations
Related Papers (5)
Frequently Asked Questions (10)
Q2. How long does the Centralized Exact optimization take to execute?
The Centralized Exact optimization however, takes 300ms to execute, even in the absence of a bottleneck, while the Distributed optimization is only performed when the configuration assigning maximum quality to each client becomes infeasible, leading to an execution time of on average 20ms, consisting solely out of the delay introduced by forwarding the local solutions.
Q3. What is the impact of the delayed installation of the configurations?
The impact of the delayed installation of the configurations showed that, even though the Distributed Relaxed optimization yields suboptimal configurations, the immediate installation of these configurations allows them to yield higher average quality at a significantly lower number of switches compared to the Exact optimization algorithms.
Q4. How can the in-network management reduce the number of quality oscillations?
due to the in-network management, the number of quality oscillations can be reduced with a factor 5 and with a factor 2.5 when traditional client-based approaches are combined with server-based rate shaping.
Q5. How many servers can the Distributed Relaxed optimization outperform?
Up to 4 servers, the Distributed Relaxed optimization is able to outperform the Centralized Exact Delayed optimization due to the installation delay of the former approach, which was discussed earlier.
Q6. What is the heuristic for transforming the optimal floating point solution into an integer?
The variables ac,q do not longer unambiguously define which quality each client is allowed to download, therefore a heuristic is required to transform the optimal floating point solution into an integer solution.
Q7. What is the advantage of the Distributed Relaxed heuristic?
the Distributed Relaxed heuristic is able to calculate a suboptimal configuration at low execution cost, making the approach viable for real-time delivery systems.
Q8. How many switches can be reduced to 19 when applying server-based rate shaping?
For 8 servers, the average number of switches for AVC MSS amounts to 23, which can be reduced to 19 when applying server-based rate shaping.
Q9. How can the authors increase the execution speed of a distributed LP?
The authors can increase the execution speed at the expense of a suboptimal solution by moving from an Integer LP formulation to a Relaxed LP formulation by relaxing the boolean constraints on the variables ac,q in (1) by only requiring ac,q to belong to the interval [0, 1]:∀c ∈ C,∀q ∈ Qvc : 0 ≤ ac,q ≤ 1 (17) This relaxation can be solved in polynomial time but at the cost of optimality.
Q10. What is the average buffer starvation in seconds?
Figure 8(a) illustrates the average buffer starvation in seconds, showing how the in-network optimization is able to deliver the video stream without buffer starvations, whereas AVC MSS suffers some minor frame freezes due to competing behavior.