Reservation Based Admission Control of a Multimedia Server

Transcription

1 Reservation Based Admission Control of a Multimedia Server Md. Mujibur Rahman Dr. Mostofa Akbar Rukhsana Afroz Ruby Dept. of Computer Science Dept. of Computer Science & Engg, Millinium Solutions Ltd. American International University Bangladesh Bangladesh Univ. of Engg. & Technology ruby@msl.co mujib@aiub.edu mostofa@cse.buet.ac.bd Abstract: In this paper we present a future reservation based admission controller for a single server media service provider. The admission controller determines the future schedule of a multimedia session to maximize the earned revenue from the sessions for the media server. This is an NP Hard problem and we present a heuristic for the admission controller. We only consider I/O bandwidth in this research as VoD service requires retrieval of huge volume of media data from the hard disk of the server and the use of memory and CPU are much less compared to other applications. Keyword: Admission Controller, Quality of Service (QoS), VoD, Utility, Rready time, Deadline, Start time, Eend time. 1. Introduction Multimedia applications such as video conferencing, WebTV, E-Commerce, VoD for watching movies require guaranteed Quality of Service (QoS) to satisfy the users who are willing pay certain amount of money for a particular service. It is difficult to ensure guaranteed QoS of these applications as we have limited resource in different levels of service providers (server, network and client). More precisely: The user s machine must have enough CPU cycles, I/O bandwidth, memory and hard disk space to play the multimedia stream. A fixed path from server to client of sufficiently high bandwidth with low enough latency & jitter is required from server to client. The server serving the multimedia stream must have the resources (CPU cycles, I/O BW and memory) to deliver multimedia streams to all admitted users with guaranteed absolute QoS. Media server is a high-end server that stores voluminous MPEG movies and offers Video on Demand (VoD) service to the users connected through communication network. The users can get movie streams from the server through the network. This service is attractive to both the users and the owners of the servers and networks. A busy user can enjoy a recently released movie or any of a very large number of classic movies while sitting at home, and the owners of the servers and networks might be able to earn significant amounts of revenue by providing this service. The following points depict the plausibility of this service over the IP datagram network: The PCs available in the market are fully capable to play multimedia streams. The servers can be located in the telephone exchanges and the users may be connected to the exchange through ADSL lines. These lines are capable to carry 1.5 Mbps MPEG-1 streams with considerable delay and jitter. If the links between the telephone exchanges are fiber then we do not find any chance congestion in the network. Thus admission control is not required for the network resources. IP offers best effort data-gram service over the network and TCP retransmits the lost packets. Thus TCP/IP protocol is not suitable for video data transmission. MPLS protocol ensures a predefined path from the source to the destination during a multimedia session and UDP protocol in the transport layer avoids unnecessary retransmission. Thus UDP in the transport layer and IP over MPLS in the network layer is a perfect solution for receiving video streams from the server to the users. An admission controller is necessary to decide whether a user will get service from the server or the user will be rejected due to resource contention. The objective of this admission controlling is to maximize the revenue from the guaranteed service to the users. Our research addresses the problem of admission controlling when the users want to get admitted in advance before the starting of movies sessions. If the resources of server are over-subscribed especially in the peak load hours, there is a chance of rejection if a user submits a request to start her movie session right away. In that case some users may be interested to make reservations days or even weeks before the starting time of the session. The session might start in future. Users can reserve the resource in order to admit in the system at a particular time in future. Section 2 of this paper presents the related works on this topic. Section 3 presents the mathematical formulation of the problem. The heuristic to do admission control will be presented in Section 4. Section 5 presents the simulation of the proposed admission controller with experimental results. Section 6 concludes the paper with future research plan. 2. Previous Works & Problem Definition The Utility Model, proposed by Khan [1][2][3] demonstrates the admission control and QoS adaptation principles of Adaptive Multimedia System (AMS) where the utility (revenue earned) is maximized for a single server multimedia service provider. This problem can be mapped to an Multi-dimensional Multiple-choice Knapsack Problem (MMKP), [4] which is an NP-hard problem. I-HEU, an

2 incremental heuristic to solve the MMKP, is used to do real time admission control. Domjan [6] proposed an admission controller using the Simplex Linear Programming method for resource reservation of adaptive applications. There is a starting and an end time for each application and the admission control for new requests is done to maximize the utilization of resources. In [7] a novel observation-based admission control algorithm was proposed, in which a client is admitted for service by a multimedia server only if the predicted extrapolation from the status quo measurements of the storage server utilization indicating that the services requirements of all the clients can be met satisfactorily. The performance of admission control algorithm and hence, the number of clients admitted and serviced simultaneously are maximized by employing a disk scheduling algorithm that minimizes both the seek time and rotational latency incurred while accessing a sequence of media blocks from disk. Chen [8] proposed an admission control algorithms that make acceptance/rejection decisions not only to satisfy the hardware requirements of clients requests but also to optimize the reward of the system based on a performance criterion as it services clients of different priority classes. The resources of a multimedia server may be oversubscribed especially in the peak load hours. Hence there is a chance of rejection of some requests if a customer submits a request to start his (her) multimedia session right away. In that case busy and valued customers may be interested to make reservations days or even weeks before the starting time of the session in order to enjoy the session (e.g. movie) in future with guaranteed QoS. The above mentioned [1][2][3][7][8][9] admission controlling strategy does not address reservation issues. Our research addresses the resources reservation scheme and profit (revenue) maximization by the system. The customers submit the session requests for consideration. The request might for starting the session right away or later at a particular time. The requests are batched over an epoch. Batches of requests are submitted to the admission controller for consideration. The Admission controller keeps the record of available resources and it runs the admission controlling algorithms for determining the profitable sessions that will be admitted. It also keeps the track of resources that will be available future and revenue earned from the accepted sessions. 3. Mathematical Formulation of the Problems. Let us consider a multimedia server system of Video-on-demand services provider where n session requests s 1, s 2, s 3 s n from n users are willing to book their movie sessions. Each session requires b i resources (Bandwidth) from the server. A multimedia session request can be specified by Ready time (r i ), the earliest time when a session can be started; Deadline (d), the latest starting time of the session, Duration of the session (l i ), length of the session. Required resource (b i ) for a multimedia session may be I/O Bandwidth, processor cycle, main memory, and network Bandwidth. Here we consider Bandwidth (b i ) as required resource only. Utility (u i ) is defined as the revenue earned by the server from a user enjoying a session. Mathematically a multimedia session reservation request can be described by s i s i = (r i, d, l i, b i, u i ) Where, r i = Ready time l i = Duration of the session u i = Utility d = Deadline b i = Required resource 3.1 Inputs of the problem Customers place their session request to the server of MSS (i.e. Video on Demand) through the network in order to book their sessions, so that they can enjoy multimedia streams in future at their desired time. Hence the input of the problem is the session requests which are defined by ready time, deadline, and duration of the session, required resources and utility of the session. 3.2 Constraints of the problem At any moment, the system state has to obey the following constraints i.e. boundary of session constraints and resource constraints. 3.3 Boundary constraints The session must be started on between ready time and (ready time + deadline). Boundary of starting time: from r i to r i +d The session may be started between ready time (r i ) and ready time (r i ) + deadline (d). Adding the duration of session (I i ) with the starting time, we will get the ending time. Boundary of ending time: from (r i + I i ) to (r i + I i + d) The difference between ending time and starting time must be equal to the length of the session (I i ). Duration of session: I i = E i - S i 3.4 Resource constraints The required resource for the session is assumed as Bandwidth (b i ) and the resource is additive. In a particular time, the sum of quantities of the resources allocated to all the sessions cannot exceed the total available quantity of the resource. For example,

3 the total resource required for n sessions in a particular time n are b i (t). If the available system resources of the multimedia i=1 server are B, the resource constraints are expressed by the following inequality. n B > b i (t) i=1 Where, b i (t) = 0, when t < s i and t > s i + l i b i, when t >= s i and t <= s i + l i 3.5 Objectives of the problem Commercial multimedia server system must have an objective which optimizes the use of resources and maximizes the revenue. Hence the offered price (utility) for a session can be expressed in mathematical notations as follows - Objective: U = u i w i is maximized (i.e. maximum revenue will be earned by the MSS). Where, u i = Utility (price) offered for a session w i = 0, if request is rejected 1, if request is accepted. S i = The starting time of session if w i = 1 E i = The starting time of session if w i = Output of the problem If a multimedia session request is submitted to the server, then the following output will be found- Depending upon the availability of resources, the multimedia session request may be accepted or rejected by the admission controller. If multimedia session request is accepted than the starting time (S i ) and ending time (E i ) of the session will be computed and it will be informed to the user by the admission controller. Starting time : The time when a user starts to get multimedia stream from the server is called the starting time (S i ). Ending time: The multimedia session will not be stopped on or before the ending time (E i ) after successful enjoyment. The user must start the session on the starting time as decided by the admission controller. 4. Algorithms for Reservation based Admission Control The algorithms of Resrvation based Admission Controller are described below- 1. All the requested sessions are ranked according to utility. 2. Preliminary accepted and preliminary rejected sessions are calculated and stored in two lists as stated above. If rejected session list is not empty then one rejected session is taken 3. Find out in which segments resources is not available, there is any intersection in that segment with another preliminary accepted session s segment, preliminary accepted session s starting time can be adjusted between ready time and (ready time + deadline ). If all these are possible starting time of preliminary accepted session is adjusted and resource of preliminary rejected session is accommodated in that segment. 4. Another Segment is found out where resource is not available for that preliminary rejected session and previous calculation is applied. 5. If above two calculations are successful then one or more segments may be infeasible and same calculation is applied for making infeasible segments feasible. 6. If accommodation of rejected session is not possible with the above calculations then preliminary accepted sessions will be stored in less utility list whose utility(ies) are less than preliminary rejected session s utility. 7. Less utility list will be sorted in ascending order in terms of utility. 8. One or two sessions from less utility list will be rejected and step 3 and 4 are performed. 9. If resource accommodation is possible starting point of rejected session is determined between (ready time + deadline) and ready time. Then for rest of the time to makeup the rejected session s duration steps 3 to 8 are performed. If accommodation is possible rejected session is stored in the accepted session list. And take another rejected session from preliminary rejected session list and step 3 is continued. 10. If step 9 is not successful steps 6, 7, 8 are performed and then step 9 is performed. 11. If step 5 is successful then step 9 is continued. 12. If after step 8 preliminary rejected session cannot be accommodated then it is assumed finally rejected session and if preliminary rejected session is not empty, take another rejected session. 13. The above calculations are applied for all the rejected sessions until the rejected session list is empty and finally admitted sessions are found.

4 5. Experiment Details and Simulations In our simulation, we consider the scenario of Video-ondemand (VoD) services. We setup our experiments in accordance with the VoD services. An environment of single server VoD is created so that users can book their multimedia sessions that they will enjoy in future. Multimedia session requests are submitted on batches in a regular time interval which we called an epoch. In each epoch there are some session requests that will be considered by the admission controller. The no. of users in each epoch is a random variable. In the simulation we advance the clock in a regular time (i.e. epoch). We used fixed increment time advance mechanism in our reservation based admission control simulation and next time advance mechanism in case of online admission control simulation Ready time is generated by using exponential random number and duration, bandwidth and utility are generated by uniform random number. The VoD server will maximize its earned revenue by admission controlling. 5.1 Arrival process Let us consider an arrival process where the ith epoch of sessions arrives at time t i and the number of users in this epoch is a random variable. 1. Generate the next arrival time t Generate exponential random variates r i independently of any previous r i s and that will be treated as ready time of the sessions. Ready time will be current time + 3 hour ± exponential random number (1.5 hours). 3. Generate the uniform random variates l i independently of any previous l i s that will be treated as duration. Similarly utility and bandwidth will be generated using uniform random number. No. of session requests in an epoch is also a uniform random number. 4. Returning the information in the following format that some session requests are arrived in ith epoch for admission control. s i = (Ready time, duration, bandwidth, utility) After the arrival of the batches of sessions the admission controller determines which session will be admitted in an epoch. 5.2 Working Environment We use Microsoft Visual C++ in order to simulate our proposed system as because this compiler is available. We took the data in a Microsoft WindowsXP environment. Our hardware was Pentium 4/2.00GHZ and 256 MB RAM with 32 MB AGP. 5.3 Validation of Simulation We validated our simulation by observing the effect of (i) system size vs. average sessions per epoch and (ii) current time vs. current revenue. 5.4 Results We computed revenue earned by (i) our proposed reservation schemes and (ii) the traditional (we called it online) scheme. The relation between system size (Bandwidth) and Revenue is proportional that is shown in the figure 2. The Nature of the curve depends absolutely on sessions starting point and ending point. In comparison of the three schemes Online shows better than others. This is also relative. Revenue (Millions) Revenue vs. Bandwidth Bandwidth (MB) Reservation 15 Without adjustments Online We also computed the average admission controlling time of each epoch for our proposed reservation based admission controller. We found that in the primary stage when more resource is available each epoch calculation does not require much time. But when average numbers of sessions in an epoch are increased then the number of preliminary rejected sessions also will be increased. As there are more preliminary rejected sessions, the admission controller will try to accommodate these preliminary rejected sessions. Hence more time is needed for epoch calculation. The effect of admission controlling time on average sessions per epoch is shown in the figure 3. Time (Milliseconds) Admission controlling time Average sessions per Epoch Best case can be obtained if all the sessions of an epoch are accepted at first Bandwidth availability check. Worst case can be obtained when some sessions are accepted and some are rejected and some sessions need to be removed and all the accepted session s utility are less than the rejected session that have to be inserted. If average sessions per epoch are increased in a given system then there will be some worst case. In that case more time will be required for admission controller. We also observed the change of earned revenue with respect to (i) arrival rate, (iii) deadline and (iii) epoch etc. Arrival rate is inverse of inter arrival time. If arrival rate increases then the revenue will also be increased. From the 4500

5 curve it is seen that revenue is proportional to arrival rate. But after a certain arrival rate the revenue is found to be constant. This is because after a certain arrival rate the distribution of session requests will be constant. Revenue (Million) Arrival rate In our simulation we increase current time by certain amount. This certain amount of time is called epoch. In each epoch we take random number of sessions. Starting time of sessions is calculated in terms of current time as previously shown. It is seen that if epoch is increased the sessions starting time are elaborately distributed in larger time line. So more resources are available in future and more revenues are earned. This is shown in the figure-6.1 for a variable system size. Revenue (Millions) Revenue vs. Bandwidth Bandwidth (MB) In our simulation we consider deadline for each session is fixed amount added to the ready time. If deadline of the sessions is increased then more sessions can be accommodated by the admission controller and earned revenue will be increased. This is because the admission controller has wide choice of adjusting the starting time. This is shown in the curve. Revenue (Million) Deadline vs. Revenue Deadline (Minute) Conclusion Admission controller s calculation for selecting sessions in each epoch for admission is suboptimal. Because for accommodation of any preliminary rejected session one or two preliminary accepted sessions are rejected whose utility(ies) are less than preliminary rejected session. But this rejection is not optimal. We do not check whether these preliminary accepted sessions utility(ies) is (are) minimum. Optimal solution was possible. But it requires more time. In admission controller s calculation rejection of preliminary accepted session s are rarely shown. This solution can be assumed about optimal. In simulation it is seen that in the primary stage when more resource is available each epoch calculation does not require much time. But as preliminary rejected sessions are increased more time is needed for epoch calculation. After observation of the simulation result we find that there is no fixed rule or regulation in which case revenue is better reservation scheme or online scheme. In which case revenue will be better depends absolutely on the session s starting time and ending time and deadline of the reservation scheme. In our simulation, we observe that earned revenue of reservation based system and online system are almost the same. Here the rate of the service is considered same, but practically it is not a regular practice. Users will pay more for reservation because users will be certain that they will be able to enjoy the session at their desired time. The Quality of Service (QoS) will be guaranteed. Reservation scheme will be much more profitable for service providers. References: [1] S. Khan and K. F. Li and E. G. Manning. Padma: An Architecture for Adaptive Multimedia Systems. IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, pp , Aug 20-22, [2] S. Khan and K. F. Li and E. G. Manning. The Utility Model for Adaptive Multimedia Systems. International Conference on Multimedia Modeling, Singapore, pp , Nov 17-20, [3] L. Chen, S. Khan, K. F. Li and E. Manning. Building an Adaptive Multimedia System using the Utility Model. International Workshop on Parallel and Distributed Real-time Systems, San Juan, Puerto Rico, April, [4] S. Khan, K. F. Li, E. G. Manning, M. M. Akbar. Solving the Knapsack Problem for Adaptive Multimedia System, Studia Informatica, [5] M. M. Akbar. The Distributed Utility Model Applied to Optimal Admission Control & QoS Adaptation in Multimedia Systems & Enterprise Networks, PhD

6 Dissertation. Department of Computer Science, University of Victoria, [6] H. Domjan and T. R. Gross. Managing Resource Reservations and Admission Control for adaptive Applications, Proceedings of IEEE International Conference, Zurich, [7] Harrick M. Vin. Alok Goyal, Anshuman Goyal, and Pawan Goyal. An Observation-Based Admission Control Algorithms for Multimedia Servers. In http//citeseer.ist.psu.edu/ vin94observationbased.html [8] Ing-Ray Chen and Chi-Ming Chen. Threshold-Based Admission Control Policies for Multimedia Servers. The Computer Journal, Vol. 39, No. 9, pp , February, [9] M. M. Akbar, E. G. Manning, G. C. Shoja. Admission Control and QoS Adaptation in Distributed Multimedia Server System. ITCOM 2001, Denver, USA, August 2001.