MoWIE for Network Aware Application

2.1. Cloud Gaming

As mentioned above, cloud gaming is widely used, and this kind of games requires low latency and highly reliable transmission of motion tracking data from the user to the gaming server in the cloud, as well as low latency and high data rate transmission of processed visual content from gaming server cloud to the user devices. Cloud gaming is regarded as one major killer application as well as traffic contributor to wireless and cellular networks including 5G. The major advantages of cloud gaming are easy & quick starting (no/less need to download and install a high-volume software in the user device), less cost and process load in the user device and it is also regarded as an anti-cheating measure. Thus, cloud gaming has become a competitive replacement for console gaming using cheaper PC or laptop. To support high quality cloud gaming services, the application needs to get the information from the network layer, e.g., the data rate value or range which lower layer can provide in order to perform rendering and encoding, during which the application in the cloud can adopt different parameters to adjust the size of produced visual content within a time period.¶

2.2. Low Delay Live Show

In 2019, over 500 million active users were using online personal live show services in China and there are 4 million simultaneous online audience watching a celebrity's show. Low delay live show requires close interaction between the application and the network.¶

Compared with conventional broadcast services, this service is interactive, which means that audience can be involved and they are able to provide feedback to the anchor. For example, a gaming show broadcasts the gaming playing to all audiences, and it also requires playing game interaction between the anchor and audiences. A delay lower than 100 ms is desired. If the delay is too large, there will be undesirable degradation on user experiences especially in a large-scale show. To lower the latency and provide size-adjustable show content, the application also requires real-time lower layer information.¶

2.3. Cloud VR

Cloud VR data volume is large which is related to different parameter settings like DoF (Degree of Freedom), resolution and adopted rendering and compression algorithm. The rendering can be performed at the cloud/network side or a mix of the cloud and the user device side. Because the latency in cloud VR is even as low as 20 ms, the application may need to interact with network to get the information about the segmentation or transport block information, and these lower layers information may be dependent on different layer 2 and layer 3 wireless protocol designs.¶

2.4. Performance Requirements of these Use Cases

There are different bandwidth, latency and lagging requirements for the above services which are characterized as parameter range. The reason of using a range is because such requirements are related to a group of parameter settings including resolution, frame rate (FPS, frame per second) and the compression mechanism. We consider 1080p~4K as the resolution range, 60-120 FPS (Frames per second) as the frame rate and H.265 as an example compression algorithm. The end-to-end latency requirement is not only related to FPS but also the property of the service, i.e., for weak interactive and strong interactive services.¶

With the typical parameters setting, cloud gaming generally needs a bandwidth of 20~60 Mbps; we also consider that significant lagging happens when the latency is larger than 200 ms, depending on the types of games (e.g. 40 ms for First Person Shoot games, 80 ms for Action games, and 200 ms for Puzzle games). In order to avoid bad user experiences, lagging is better when it is lower, and can be as low as zero (in an optimal QoE). For low latency live shows, 20~50 Mbps bandwidth may be needed and the end-to-end latency requirement is less than 100 ms. Cloud VR service generally requires 100~500 Mbps bandwidth and 20~50 ms end-to-end latency. It is noted that these values are dependent on the parameter settings and they are provided to illustrate the order of magnitude of these parameters for the aforementioned use cases. These value range may be updated according to specific scenarios and requirements.¶

3.1. Video Compression Based on ROI (Region of Interest)

A foveated mechanism [Saccadic] in the Human Visual System indicates that only small fovea region captures most visual attention at high resolution, while other peripheral regions receive little attention at low resolution. And we call those regions which attract users most, the regions of interest (ROI)[Fahad].¶

To predict human attention or ROI, saliency detection has been widely studied in recent years, with applications in object recognition, object segmentation, action recognition, image caption, image/video compression, etc.¶

Since there exists the region of interest in a video, the cloud server can give the ROI region higher rate while making other regions a lower rate. As a result, the whole rate of the video is reduced while the watching experience will not be harmed.¶

This method means to detect the ROI and re-allocate the coding scheme for interested and non-interested regions in order to save the bandwidth without sacrificing user's QoE. In recent years, the ever- increasing video size has become a big problem to applications. The data rate of a cloud gaming video in 1080P can reach 25 Mbps, which brings huge burden to the network, even for 5G network. Those ROI- based video compression methods are mainly applied to high concurrency networks to relive the burden of networks and then keep QoE in an acceptable range.¶

However, current methods utilize application information like application rate and application buffer size as the indicators to roughly adjust the algorithm in interactive video services. That information is hard to reflect the real-time network status precisely. Therefore, it is hard to balance the QoE and bandwidth saving in real-time scenario. More direct information is helpful for those ROI methods to improve the performance.¶

3.2. AI-based Adaptive Bitrate

This method intends to reduce lagging and ensure acceptable picture quality.¶

Applications such as video live streaming and cloud gaming employ adaptive bitrate (ABR) algorithms to optimize user QoE [MPC][CS2P].¶

Despite the abundance of recently proposed schemes, state-of-the-art AI based ABR algorithms suffer from a key limitation. They use fixed control rules based on simplified or inaccurate models of the deployment environment. As a result, existing schemes inevitably fail to achieve optimal performance across a broad set of network conditions and QoE objectives.¶

A reinforcement learning based ABR algorithm named Pensieve was proposed [Hongzi] recently. Unlike traditional ABR algorithms that use fixed heuristics or inaccurate system models, Pensieve's ABR algorithms are generated using observations of the resulting performance of past decisions across a large number of video streaming experiments. This allows Pensieve to optimize its policy for different network characteristics and QoE metrics directly from experience. Over a broad set of network conditions and QoE metrics, it has been proven that Pensieve outperformed existing ABR algorithms by 12%~25%.¶

For this method and those methods built upon this, it has been proven that all information, including rate, download time, buffer size or network level information which can reflect the performance, are useful to reinforcement learning. Since those data can reflect the network dynamics, they have been used to help the applications to know how to change the rate and improve users' QoE.¶

However, all these data are obtained from the client side or the server side. In reality, it is not easy to obtain such data in an effective and efficient way. The lack of standardized approach to acquire these data makes it difficult to make this usable for different applications for large scale deployment. Meanwhile, since these data reflect the real-time network status, they may change rapidly and randomly, and hence can be hard to use a theoretical model to characterize.¶

To summarize, current practices can make some improvements by indirectly measuring network status and react in the application.¶

However, the network status data are not rich, direct, real-time; they also lack predictability, especially when in the mobile and wireless network scenarios, which results in long reaction delay or high QoE fluctuations.¶

4.1. MoWIE Architecture and Network Information Exposure

The fundamental idea of MoWIE is to achieve on demand and periodic network information from network to applications, helping network service providers to realize a better policy control and to improve users' experience.¶

A possible MoWIE architecture includes three core components: the Client Application, the Mobile Network and the Application Server.¶

The raw data are collected firstly from the radio network and core network; further processing on these collected data and the exposure of Network information are provided to the application Server.¶

An application server can send network information request about UE/Cell level information and obtain the NIS response on network information from the mobile network. After user data pre-processing, the application server will make best use of the network information to perform analytics and directly enhance the application functions e.g. bit rate, latency, and jitter.¶

Typically, the network information provided by MoWIE includes two types of information as below:¶

Cell level Information:¶

• The number of Downlink PRBs (Physical Resource Blocks) occupied during sampling period;¶
• the cell load;¶
• the downklink (DL) MAC data rate per cell;¶
• the channel status (e.g. RSRP (Reference Signal Received Power) and CQI (Channel Quality Indicator));¶
• the DL data rate;¶
• the PDCP (Packet Data Convergence Protocol) buffer status;¶

UE level information (without privacy information):¶

• The Downklink Signal to Inference plus Noise Ratio (SINR);¶
• MCS: The index of Modulation and Coding Scheme (MCS);¶
• The number of packets occupied in PDCP buffer;¶
• The number of downlink PDCP Service Data Unit (SDU) packets;¶
• The number of lost PDCP SDU packets;¶
• The per UE downlink MAC data rate;¶
• the per-UE channel status (e.g. RSRP (Reference Signal Received Power) and CQI (Channel Quality Indicator));¶
• the per-UE DL data rate;¶
• the per-UE PDCP (Packet Data Convergence Protocol) buffer status;¶

The network information listed here can also be found in 3GPP (PRB [TS38.211], cell load [TS38.300] PDCP for 5G [TS38.323] RSRP, RSRQ, RSSI [TS38.331], MCS, CQI [TS38.214], The number of packets occupied in PDCP buffer, the number of Downlink PDCP SDU packets, the number of PDCP SDU packets lost, the per-UE PDCP buffer status [TS38.323]), to demonstrate the potential benefits of MoWIE for network-application integration over cellular network. Figure 3-1 and Figure 3-2 list the data types correponding to the cell-level information and UE-level information, respectively.¶

            +----------------------+--------------------------+
            |Cell-level Information|     Data type/Range      |
            +----------------------+--------------------------+
            |         PRB          |          Uint16          |
            +----------------------+--------------------------+
            |         CQI          |          Uint8           |
            +----------------------+--------------------------+
            |        RSRP          |          Uint8           |
            +----------------------+--------------------------+
            |        RSRQ          |          Uint8           |
            +----------------------+--------------------------+
            |     Cell load        |          [0,1]           |
            +----------------------+--------------------------+
                        Figure 4-1: Cell level data type

            +------------------------------------+---------------+
            |         UE-level Information       |Data type/Range|
            +------------------------------------+---------------+
            |            Downlink SINR           |     Uint16    |
            +------------------------------------+---------------+
            |                MCS                 |     Uint8     |
            +------------------------------------+---------------+
            |      Downlink PDCP SDU packets     |     Uint8     |
            +------------------------------------+---------------+
            |    PDCP SDU packets lost           |     Uint8     |
            +------------------------------------+---------------+
            |    Packets occupied in PDCP buffer |     [0,1]     |
            +------------------------------------+---------------+
            |                CQI                 |     Uint8     |
            +------------------------------------+---------------+
            |                RSRP                |     Uint8     |
            +------------------------------------+---------------+
            |                RSRQ                |     Uint8     |
            +------------------------------------+---------------+
                        Figure 4-2: UE level data type

4.2. RAN-assisted TCP optimization based on MoWIE

The RAN information is used to assist TCP sending window adjustment rather than traditional transport layer measurement and acknowledgement. The RAN proactively predicts available radio bandwidth and the buffer status per UE in a time granularity of RTT level (e.g. 100 ms) and then piggybacks such information in TCP ACK.¶

We have conducted trial in real mobile networks. It is observed that for the UE with good SINR, the throughput is significantly improved by nearly 100%, and the UE with medium SINR can achieve approximately 50% gain.¶

4.3. NAA QoE Test based on MoWIE

Different from traditional video streaming, cloud gaming has no buffer to accommodate and re-arrange the received data. It must display the stream once the stream is received. Any late stream is of no use for the player. Cloud gaming performs not well in the existing public 4G network according to our actual measurements. The end to end delay is often greater than 100 ms for a gaming client in Shenzhen to a gaming server in Shanghai, coupled with the codec delay. Here the delay is defined as the total delay from the user's operation instruction to show the response picture on user's screen.¶

Once the network fluctuates, users will experience a longer delay.¶

The poor user experience is not only because of the relative low network throughput, but also because that the server cannot adapt the application logical policies (e.g., codec scheme and data bitrate).¶

The popularity of 4K and even higher resolution and increasing FPS for cloud gaming and AR/VR services require both high bandwidth and low latency in wireless and cellular networks. The increasing resolution would incur a higher encoding and decoding delay. However, users' tolerance to delay will not increase with the resolution, which means the application needs to adapt to the network dynamics in a more efficient way. The higher resolution, the larger range of the rate adaptation can be used.¶

In this section, we make experiments based on the methods described in section 3 to improve the QoE of cloud gaming. The performance between network-aware and native non-network-aware mechanisms are compared.¶

4.4. ROI Detection with Network Information

The first experiment is based on the ROI detection. We will investigate the impact of network perception.¶

Saliency detection method has successfully reduced the size of videos and improve the QoE of users in video downloading [Saliency].¶

However, it is not effective when applied to real-time interactive streaming such as cloud gaming.¶

As we know, more accurate saliency region detection algorithm needs more time to obtain the result. However, when the users are suffering a bad performance network in cloud gaming, this precise detection may incur more delay to the system. As a result, it will harm the final QoE.¶

If the application can learn the network well in a real-time manner, it can choose the algorithm based on how much delay the system can tolerate. If the network condition is good enough, it can adopt an algorithm which has deeper learning network and the added delay will not be perceived by the end users. Thus, it can save huge bandwidth without harming the QoE. On the other side, in a network with bad condition, the server can use the fastest method to avoid extra delay.¶

We make the experiments to show how the network information will influence the total QoE and bandwidth saving in ROI detection.¶

The following 4 methods are compared:¶

1) The original video, without using ROI method. This acts as a baseline.¶

2) Quick saliency detection and encoding method, which is not accuracy in some cases. It only brings 10ms delay.¶

3) A relative accuracy saliency detection method. In general, if an algorithm is more precise, it will take more time to get the results.¶

And the complexity of the picture will also influence the detection time and accuracy. Based on our test video, we adopt the method which brings delay about 40~70ms.¶

4) The application server in the cloud has the current bandwidth information which derived from the wireless LAN NIC. Here it is a simulation that all the collected bandwidth traces are already known by the server. Thus, it can use the bandwidth traces to compute transmission delay. Then the server can change the saliency detection algorithm based on this information and then encode the video.¶

Although the result of future bandwidth prediction is not always accurate in real environment, the assumption here will not influence the final results much. Since in cloud gaming the server encodes the stream based on ROI information frame by frame instead of in a grain of chunks, the future bandwidth prediction window size doesn't have to be long. Therefore, even the server can only get the bandwidth or delay prediction for a short time window, the server can still use this method with network information.¶

Test environment:¶

A 720P game video segment with a rate of 6.8Mbps. This is not a very high bandwidth requirement example in cloud gaming. We just show how it will benefit from MoWIE. High bandwidth requirement case will benefit more if the bandwidth fluctuates much.¶

The three different networks are all wireless networks and the available bandwidth is varied frequently, where Network 1: The overall network condition is not very good, the average network bandwidth is 7.1Mbps, but it continues to fluctuate, and the minimum is only 3.9Mbps.¶

Network 2: The overall network condition is good, with an average network bandwidth of 12Mbps and a minimum of 6.4Mbps.¶

Network 3: The network fluctuates dramatically, with an average network bandwidth of 8.4Mbps and a minimum network bandwidth of 3.7Mbps¶

Test content:¶

The four methods are conducted on the original video under each three networks. After re-encoding based on the saliency detection, we calculate the new QoE and the saved bandwidth. The results are shown in the Figure 4-1 (The QoE value is the MOS as standardized in the ITU):¶

    +---+-----------------+-----------------+-----------------+
    |   |    Network 1    |    Network 2    |    Network 3    |
    +---+---+-------------+---+-------------+---+-------------+
    |   |QoE|  BW Saving  |QoE|  BW Saving  |QoE|  BW Saving  |
    +---+---+-------------+---+-------------+---+-------------+
    | 1 |3.8|      0      |4.8|      0      |4.3|     0       |
    +---+---+-------------+---+-------------+---+-------------+
    | 2 |3.8|     5%      |4.8|     9%      |4.3|    7%       |
    +---+---+-------------+---+-------------+---+-------------+
    | 3 |2.2|    2.1%     |4.6|     38%     |3.1|    34%      |
    +---+---+-------------+---+-------------+---+-------------+
    | 4 |3.6|     9%      |4.7|     33%     |4.3|    25%      |
    +---+---+-------------+---+-------------+---+-------------+
               Figure 4-3: QoE and Bandwidth Saving

Conclusion:

It can be seen that the methods such as method 2 and method 3 that do not rely on the network information directly, have certain limitations.¶

Though the method 2 is simple and time-consuming, it can only detect a small part of region of interest accurately. Thus, even if the network condition is very good, it can only save a small amount of bandwidth, and sometimes there are some incorrect ROI detection. The QoE will be reduced without hitting the ROI region.¶

For Method 3, the algorithm is complicated, and it can correctly detect the user's area of interest, so that it can re-allocate encoding scheme and save a lot of bandwidth. However, its algorithm will introduce higher delay. When the user network condition is poor, the extra delay will cause even worst user's QoE. Although the bandwidth is saved, it affects the user experience seriously.¶

Method 4 is based on the application's awareness of the network. If the application can know certain network information, it can balance the complexity of the algorithm (introducing delay) and the accuracy of the algorithm (saving bandwidth) according to the actual network conditions. As can be seen from the experiment, method 4 can ensure the user's QoE and save the bandwidth greatly at the same time.¶

4.5. Adaptive Bitrate with Network Capability Exposure

This experiment is AI-based rate adaption by utilizing the network information provided by the cellular base station (gNB) in cellular network.¶

Tencent has launched real network testing of NAA-enabled cloud gaming in China Mobile LTE network, with the enhancement in gNB supporting base station information exposure.¶

To enable the NAA mechanism, some cellular network information from gNBs are collected in an adaptive interval based on the change rate of network status. There information is categorized in two levels, i.e., cell level and UE level. Cell level information are common for all the UEs under a serving LTE cell and UE level information is specific for different UEs. 3GPP LTE specifications have specified how the PDCP, RLC (Radio Link Control), MAC (Medium Access Control) and PHY (Physical) protocols operate and this information are very essential statistics from these protocol layers.¶

It is noted that in NAA mechanism, as the network information is from gNB, and the gNB has the real-time information of radio link quality statistics and layer 1 and layer 2 operation information, NAA mechanism can expose rich information to upper layer, e.g., it is capable to differentiate packet loss and congestion [MengZ], which is very helpful to the applications in practice.¶

In order to compare the cases with and without NAA, the cloud gaming test environment is setup with 1080p resolution and around 20Mbps bitrate.¶

Test scenarios 1~5 are as follows.¶

Test scenarios 1: Weak network. This scenario is the case where radio link quality is low, e.g., in cell edge area and the bandwidth is not able to serve cloud gaming.¶

Test scenario 2: User competition scenario. This scenario is defined as the case when user amount is large thus the cellular network bandwidth cannot serve all the cloud gaming users.¶

Test scenario 3-5: Other scenarios with random user movement trace and user distribution.¶

Test method: To simplify to comparison, we just use the MCS (MCS index) information derived from the gNB [TS38.214]. The information is provided directly to the application, and the application then adjusts the bit rate according to this information. Here, MCS index shows the modulation (e.g. QPSK, 16QAM,...) and the coding rate used during physical layer transmission, which is relevant to the real data rate per UE. The benchmark method is adopting a constant bit rate without any information to help it predicting the network condition. We compare these scenarios and observe the reduction of delay when those gNB data are utilized.¶

For different scenarios, the lagging rate is defined as the performance indicator. In our experiments, we assume lagging happens when transmission delay is greater than 200ms and lagging rate is defined as the ratio between the number of frames greater than 200ms and the total number of frames.¶

            +-------------+--------------------------+
            |Test Scenario| Reduction of Lagging Rate|
            +-------------+--------------------------+
            |     1       |           46%            |
            +-------------+--------------------------+
            |     2       |           21%            |
            +-------------+--------------------------+
            |     3       |           37%            |
            +-------------+--------------------------+
            |     4       |           56%            |
            +-------------+--------------------------+
            |     5       |           32%            |
            +-------------+--------------------------+
              Figure 4-4: Reduction of Lagging Rate

It can be clearly seen that with the MCS information, the application can adjust the bit rate to decrease the lagging rate and then significantly improve the user QoE. In weak network scenario, 46% lagging can be avoided by NAA.¶

4.6. Analysis of the Experiments

The above-mentioned technologies demonstrate the performance gain of NAA with MoWIE.¶

Although application information can also help to predict the network and have already been used in adaptive bit rate methods, the application information is not as sensitive as gNB information at the very beginning in a lot of cases. For example, when more users enter the cell, the PRB information will first reflect that each user may get less bandwidth. However, the application information needs to react after there is a trend that the bitrate is decreasing. That is to say, the lower layer network information is more directly.¶

Without MoWIE, the application cannot get the lower layer network information directly and then try to detect "blindly" to adapt to the dynamics of the lower layer network, which cannot meet the requirements of cloud interactive applications like cloud gaming, low delay live show and Cloud VR.¶

It is noted that the more real-time network resource status the application can learn, the better it can predict how much network resource it can use within a prediction time window. However, there is tradeoff between network information collection frequency and its load and feasibility to the network devices. In principle, the total network resource consumed for such network status reporting is also designed in light-weight manner, e.g., by properly controlling the interval of report and also the number of bits needed to convey the reported information elements. In our experiments, the network status information can be obtained in an adaptive interval based on the change rate of network status, in order to provide good prediction with less load introduced in the network. In fact, not all scenarios need a very frequent information collection. If some information only changes in a very small range and won't influence the final decision, it is unnecessary to report such information all the time. However, if its value varies over the preset threshold, it will inform the application immediately.¶

The distribution and impact of the exposed data to the performance gain for different algorithm needs to be further studied. This draft is to give a guidance to figure out what kind of data needs to be exposed during initial deployment of these mechanisms.¶

In our current cloud gaming, the application information can help to reduce about 50% the lagging rate. The left 50% improvement room can be achieved by network information exposure with MoWIE. Actually, the effect of the two-layer information can be accumulated. However, due to current deployment limitation, we cannot collect the application information with the gNB information at the same time. Thus, in this version of the draft we compare the performance with and without MoWIE. We don't compare between application information assisted mode and network information assisted mode in this draft. This is our on- going work. Since both application and gNB information can reflect the network variation, we will compare the performance among application information assisted mode, network information assisted mode and the mode of utilizing both layer information.¶