Call Admission Control

In traditional algorithm [4], if the current total traffic power consumed by active fundamental channels (FCHs) plus the power needed to support the incoming voice user exceeds the maximum power budget assigned for voice traffic, a voice call will be blocked to enter the system. If the current total traffic power consumed by active FCHs plus the FCH power that would be needed to support data users with non-empty queues but currently are not assigned data bursts plus the power needed to support an FCH for the incoming data user exceeds the upper power limit, a data user will not be assigned an FCH. Apparently, the data connection admission is more crucial than the voice call admission under power overload condition because voice service always has the highest priority. In practical implementation, once a data user is denied a FCH assignment to prevent system overload, this user will continuously retry to request resources through competition until success.

However, such admission control evidently sacrifices interests of data traffic and leaves voice traffic the highest priority, even needs to deprive the existing data users of the reserved resources. The QoS based call admission control, depending on the predefined QoS criteria, should positively reserve resources for existing users, and should release tolerable flexibilities to improve whole system performance. Table 1 shows the characteristics of three mainstream media services and is regarded as predefined QoS criteria. In QoS based call admission control flow, as shown in figure 3-9, QoS criteria, BW allocation, resource reservation, and resource estimation are four major components to produce three key factors—required resources, preservative resources, and reserved resources. These factors are the basis to make final admission decision. QoS criteria are stringent entry barriers that pessimistically estimate the required resources for a new incoming user or a handoff user. To protect the interests of existing users, system must reserve sufficient amount of resources.

Thus bandwidth allocation is used to estimate these resource budgets. Resource reservation tries to preserve reasonable amount of resources for the predicted handoff traffic. The preservative resources are critical to control the risk of call dropping, though it may limit the system capacity and the resource efficiency. Resource estimation integrates reserved resources and preservative resources, and announces the available resources by subtracting them from the total resource budgets. Only when the required resources are fewer than available resources, this user can be served.

Figure 3-9 call admission control flow

In the real implementation, the call admission control algorithm, as shown in figure 3-10, tries to achieve two goals—guaranteeing sufficient resources for existing users and saving as more resources to incoming users as possible. Depending on the QoS (transmission rate) guarantee that the system have predefined, the corresponding amount of resource (power) must be reserved. Since the basic quality of MPEG4 scalable video streaming needs roughly 30kbps bandwidth, this algorithm treats the estimated power budget “PVQ,existing” as a reference, and the system designer can modify the QoS guarantee by resetting the parameter

“rreserved”. The higher “rreserved”, the better visual quality the existing users enjoy, but the more difficult the incoming users be accepted. Hence the available power budget for incoming users can be calculated as the total power budget (Ptotal) minus the sum of the overhead of control channels (Poverhead), the preservative power (Phandoff), and the reserved power (Pexisting).

If the basic power requirement of an incoming user is less than the available power, system accepts this call, or it will be blocked.

Additionally, the video traffic arrival interval is modeled as a log normal distribution, with a density parameter “video occupancy”. “video occupancy” decides the average busy interval caused by video session in an hour. Thus “video occupancy” is a kind of data Erlang and can be modified by “k_video” in our simulation. Changing “rreserved” and “k_video” can we see the impact to the system performance under various QoS guarantee and traffic density.

traffic

Figure 3-10 call admission control algorithm

Table 1 QoS criteria Voice traffic

Delay Real-time, circuit-switched design makes tiny delay between voice packets unacceptable.

Bandwidth Fixed 9.6kbps is enough to supply voice codec adopted in 3G system.

Walsh code Single, monopolized FCH is assigned.

Power The power requirement depends on fading channel, mobile speed, cochannel interference, and media codec; thus it can’t be set as expected mean value.

Video traffic

Delay It’s hard to precisely quantize each end-to-end, application-based, codec-related, subjective delay constraint, especially in unstable wireless environment. To avoid quality degradation, several buffers are equipped in both BS and MS, and hundred milliseconds of end-to-end packet delay should be kept. For example, video conference has 100ms delay bound, and real time video depending on frame rate (40fps~10fps) may have 25ms~100ms delay constraints.

Bandwidth Minimum average throughput should be 28.8kbps when MPEG4 scalable video coding techniques is adopted since at least base layer packets must be transmitted to play the lowest quality of image. Instantaneous BW variation is allowable to improve system performance but delay constraint must be obeyed.

Walsh code Converting the radio BW into Walsh code can we easily know that the least necessity of Walsh code is three, and the variation of code number is proportional to the assigned BW. Though the BW of a SCH can be arbitrarily relocated slot by slot as a common traffic channel, basic FCH is a dedicated traffic channel and can’t be reused unless current video session is terminated. This is the severe difference between BW and Walsh code and makes them not a direct mapping.

Power The power requirement depends on fading channel, mobile speed, cochannel interference, and media codec; thus it can’t be set as expected mean value.

Data traffic

Delay Besides upper layer (TCP/IP or above) timing constraints, non-real time data services have no stringent delay bounds between contiguous packets. To keep stable service quality, however, long term average transmission rate should be kept.

Bandwidth Burst delivery of data packet is suitable to be transmitted at a short interval with large bandwidth, or to be consumed through a long period by the lowest transmission rate.

Adequately alternating the transmission policies can help increasing resource utilization rate.

Walsh code Once the transmission bandwidth is assigned, the number of Walsh code can be known by direct mapping.

Power The power requirement depends on fading channel, mobile speed, cochannel interference, and media codec; thus it can’t be set as expected mean value.

Chapter 4 Simulation Results

In this chapter, we explore the superiorities of the proposed CFA solution, and its performance is compared with typical solutions—fairness based RRM combined with client based adaptation mechanism and fairness based RRM combined with RNF adaptation mechanism. Since the objectives of the proposed CFA solution are to enhance system performance and to simultaneously guarantee QoS, performance metrics are naturally categorized into two types—system perspective metrics and user perspective metrics. In section 4.1 and section 4.2, the system perspective metrics and the user perspective metrics are respectively defined and the relative performance study with detailed analysis follows.

Section 4.3 is the summary of this chapter.

4.1. System Perspective Metrics and Performance Analysis

u System capacity:

The system capacity is defined as the maximum number of active video users that system can simultaneously serve under predefined QoS guarantee. Thus we will observe the degradation of capacity following the increasing reserved_ratio. It depends on the design the RRM, power control, and resource reservation for handoff prediction. If the power control and handoff prediction are precise enough, only the RRM can affect system capacity because it decides the resource efficiency.

u Blocking rate:

The blocking rate is a probability that a new incoming user is refused to enter the system. Call blocking can avoid the system overload and protect existing users’ interests. It is the resultant of the call admission control, which is related to the QoS guarantee (reserved_resource), the handoff prediction (preservative_resource), the entry barrier (required_resource), and the total resource budgets (total_resource). The handoff prediction is a fixed mechanism in our study. The entry barrier is set as the predefined QoS criteria. The total resource budget is an equipment-related constant. Thus we will observe the blocking rate by following the increasing video traffic arrival rate under the specific reserved_ratio that is

used to reserve resources for existing users.

u Channel efficiency:

In our simulation, we set maximum available bandwidth is fixed to 614.4kbps in a cdma2000 carrier and quantize it as 64 9.6kbps channels. Thus removing 4 overhead control channels, how the system exploits the remaining 60 channels to maximize its allocation efficiency is quantified as channel efficiency. The design of RRM and the number of active video users are two major factors that can affect the channel efficiency. Thus the performance comparison among three different solutions following the increasing number of active video users is analyzed.

u Power efficiency:

Like channel efficiency, the maximum cell site transmit power is assumed to be 20 Watts, and 16 Watts can be used to support traffic channels while 4 Watts are reserved for Pilot, Paging and Synchronization channels. Thus the proportion of power usage of the remaining 16 Watts is defined as the power efficiency. Power control, power allocation of the RRM, and power reservation for handoff request are three major factors that can affect the power efficiency. Since the power control and handoff prediction are assumed to be precise enough, the power efficiency can be regarded as a performance metric to compare the proposed solution with typical solutions.

u System throughput:

Basically, if every user is always active, like real time video streaming (bandwidth requirement due to continuous transmission of new arrived video packets in BSQ is inevitable), the channel efficiency can be directly transfer to the system throughput. However, different application has different traffic characteristic. For example, a voice call has 40% that people really speak some words (active), ant the remaining 60% is kept silent (inactive). Data applications, like web browsing and E-mail, request resources bursty (short interval but higher bandwidth). Actually, the channel efficiency is not equivalent to the system throughput in most cases.

Figure 4-1 shows the capacity (maximum number of active video users) that cdma2000 system can support. Evidently, capacity is decreasing following the increasing reserved_ratio because more and more radio resources are reserved to existing users to

guarantee their QoS. The issue we should discuss here is how to choose a reasonable reserved_ratio. If a low reserved_ratio is chosen, system easily enters into overload congestion, and each user will suffer buffer underflow frequently. If a high reserved_ratio is chosen, existing user’s resources are over-guaranteed. Each existing user can request higher visual quality (encoding bandwidth), but this limits the system capacity which is directly related to vendors’ profit. Figure 4-2 is an example showing system blocking rate under reserved_ratio fixed to 1, and k_video represents the density of video arrival rate.

Figure 4-1 system capacity Figure 4-2 system blocking rate

Figure 4-3~4-5 represent three traditional metrics—channel efficiency, power efficiency, and system throughput. These metrics can prove that our proposed CFA solution is superior to the traditional solution from system point of view. Since different system loading (simultaneously active users) exploits different advantage of the proposed solution, the following discussion separates the metrics into three regions—light loading (1~4 users), medium loading (5~10 users), and heavy loading (11~15 users).

In the light loading scenario, system resources are sufficient to support arbitrary request from each user, and video users can always enjoy the optimal quality without suffering the risk of contention or scheduling. Hence the performance of three metrics in this region is a nearly linear growth, following the increasing number of video user. Since the power budget in this scenario is sufficient, all users would be served no matter their RF conditions. This is intuitively inefficient because the system must waste unnecessary power, and the slope of the power consumption becomes steep.

When system enters medium loading scenario, it can no more support every user in the optimal quality, but the resource budgets are still sufficient to handle the lowest quality.

Though the long-term trend keeps on growing due to increasing number of existing user, each solution may vibrate its resultant channel efficiency under different video user existence. Such vibration mainly stems from the exponential growth of bandwidth up switch (9.6à19.2à38.4à76.8à153.6kbps). For example, if we suppose power budget to be infinite, six users maximize system throughput as 163.2+163.2+86.4+86.4+48+28.8=576kbps;

however, seven users can exploit only 163.2+163.2+48+48+48+48+48=566.4kbps as the maximal system throughput. Consequently, channel efficiency and system throughput changes due to the number of video user, finite stages of the assigned bandwidth, and different combinations of the RRM and encoding adaptation mechanism. The proposed CFA solution achieves 81% channel efficiency, but two typical solutions can achieve only 65% and 35%

channel efficiency in average, respectively. Power efficiency is still growing up but the slope tends to be flatter because system now can efficiently schedule users in good RF conditions first. The CFA solution is evidently superior to others because it can utilize analogous power budget to produce highest throughput.

If heavy loading scenario happens, half or above power budget is consumed to support each user’s lowest visual quality, and the remaining budget limits the flexibility of quality adaptation. These result in stably high power utilization rate (80%). Simulation results prove that the CFA solution (85% channel efficiency) is superior to the typical solutions (the RNF solution with 35% channel efficiency and the client based solution with 80% channel efficiency). The 5% difference between the CFA solution and the client based solution reveals that the adaptation mechanism based on feedback information is better than blindly trial and error even rare flexibility is available.

Figure 4-3 channel efficiency Figure 4-4 power efficiency

Figure 4-5 system throughput

4.2. User Perspective Metrics and Performance Analysis

u Number of rebuffer event and rebuffer interval:

From the experience of serving a real time video streaming on the internet, the most unacceptable event that severely degrades the user perceived quality is the frequent suspension due to network congestion or client buffer underflow. Besides, how fast a buffer underflow event can be compensated and return to a normal clip playout is also critical.

Number of rebuffer event shows the average number of rebuffer event during a 120 seconds video session, representing the frequency of buffer underflow. Rebuffer interval means the average suspend time to manipulate rebuffer events. Frequent rebuffer events and longer rebuffer interval result in the deterioration of user perceived QoS. The most effective solution is to reserve more resources for existing video users as the QoS guarantee. Thus we will observe these two metrics following the increasing reserved_ratio. Only when a reasonable reserved_ratio is chosen, the avoidance of buffer underflow and the maximization of resource efficiency can be optimally balanced.

u Personal throughput:

The mean of personal throughput shows how much bandwidth a user can share following the increasing number of simultaneously active users. The standard deviation of personal throughput distinguishes the deviation of bandwidth allocation (QoS) among video users. This is a metric that shows the tradeoff between the fairness (everyone has the consistent visual quality no matter the RF condition) and the efficiency (a user in better RF condition can enjoy higher visual quality).

u Visual quality:

Visual quality can be defined as the combination of stability and resolution. Because of the lack of radio resources, to support a steadily high resolution video stream is impossible.

High flexibility of the RRM and fast response of the video encoding adaptation mechanism intuitively results in relatively frequent switch of the resolution. Hence what we seek is the tradeoff between stability and resolution under the maximization of resource efficiency. Since we assume the media server can encode video frame in finite stages of scalability, corresponding to Walsh assignment in SCH (19.2 /38.4 /76.8 /153.6 kbps), the bandwidth adaptation ratio is twice per switch.

Figure 4-6 draws average number of rebuffer event and figure 4-7 draws average suspend interval of a rebuffer event during 120 seconds continuous video session. To analyze the performance of each solution, we separate performance curves into three regions—rarely reserved (reserved_ratio=0~0.7), proper reserved (reserved_ratio=0.7~1.8), and over reserved (reserved_ratio=1.8~4).

Figure 4-6 average number of buffer

underflow during a 100sec clip

Figure 4-7 proportion of video suspension during a video session

Rarely reserved scenario, mapping to the capacity shown in figure 4-1, implies that 16~36 video users must simultaneously be served. Since the system bandwidth has been quantified into 64 channels (4 channels are control overhead, and other 60 channels are traffic channels), corresponding to 64 Walsh codes, the MPEG4 scalable video streaming, which requires roughly 30kbps to transmit base layer packets, needs 3 channels as the QoS lower bound. Thus to support 16~36 video sessions, at least 48~108 channels are necessary. We

have discussed that the maximum channel efficiencies of each solution are 81% (CFA), 65%

(client based), and 35% (RNF), and the largest amount of simultaneously available channels is 48 under power budget limitation. Apparently, it is impossible for a system to sustain rarely reserved scenario, and the frequent rebuffer events are inevitable. Multiplying “average number of rebuffer event” by “average suspend interval of a rebuffer event “can we obtain the proportion of the suspend interval during a video session. Simulation results show that the CFA solution suffers 82.5%~ 2.5% suspensions, the client based solution suffers 70.2%~14%

suspensions, and the RNF solution suffers 70%~12.5% suspensions. In this scenario, performances of three solutions are unsatisfied, and it is meaningless to compare the superiority. However, sharper degradation of the rebuffer event in figure 4-6 and lower proportion of the suspend interval in figure 4-7 can still help observing faster convergence of rebuffering elimination using the CFA solution.

In the proper reserved scenario, the suspend ratios of three solutions decrease from 2.5% to nearly 0% (the CFA solution), 14% to 1.6% (the client based solution), and 12.5% to 1.2% (the RNF solution), respectively. Though the performances of two traditional solutions may still not be acceptable, the CFA solution has successfully solved rebuffer problem and at the same time maintained system capacity without waste of over-reservation. Thus we call this scenario “proper reserved”. If a system designer adopts the CFA solution, it’s strongly recommended setting reserved_ratio among this region (0.7~1.8). A vendor would tradeoff its QoS guarantee and capacity which affect the profit margin.

There are three phenomena can be discovered in over reserved scenario. First, reserved ratio higher than upper bound of proper region CFA is worthless unless system promises supplying advanced visual quality. Second, if the RNF solution must be used, reserved_ratio above three is suggested and this implies that the RNF solution is adequate to handle light loading. Third, even a high reserved_ratio can’t completely avoid rebuffer events when the client based solution is used unless the capacity shrinks to less than four users. This is because the maximum power budget (1/4 total traffic power per user) is absolutely sufficient to support a user transmitting its video clip in the lowest quality.

In our simulation database, bandwidth assigned to a user on every time slot (100ms) creates a row vector, and the records of multiple users are combined as a two-dimensional

deviation are calculated first, and these two statistic factors among multiple users also result in probability distributions. Thus four final metrics--user expectation (mean) of timing

deviation are calculated first, and these two statistic factors among multiple users also result in probability distributions. Thus four final metrics--user expectation (mean) of timing