Traffic Contract. Fig. 1. QoS management of GPRS networks. This paper is organized as follows. Section II describes the. QoS profile of GPRS. The descriptions ...
Service Scheduling for General Packet Radio Service Classes ↑
Qixiang Pang↑, Amir Bigloo⇑, Victor C. M. Leung↑, Chris Scholefield⇑
Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, BC, Canada ⇑ Motorola Cellular Infrastructure Group, Richmond, BC, Canada
Abstract- While the quality of service (QoS) profiles for a number of general packet radio service (GPRS) classes has been specified [1] by ETSI, how QoS management is provided by means of traffic scheduling, traffic shaping, and connection admission control, in a GPRS network is an implementation issue that is attracting significant current research interest. This paper presents an evaluation of several traffic scheduling methods, including of FIFO, Static Priority Scheduling (SPS) and Earliest Deadline First (EDF) by simulations, with the objective of meeting the delay profiles defined for a number of GPRS classes. Traffic sources representative of GPRS applications, including email, fleet management and world wide web applications are employed. We focus on the forward link which represents the bottleneck of a typical GPRS data connection. Results show that EDF is able to meet the delay requirements at a much higher channel utilization compared to the other alternatives.
• • •
First In First Out (FIFO) Static Priority Scheduling (SPS) Earliest Deadline First (EDF)
The performance of these scheduling methods are compared and commented.
Traffic Characterization
New Traffic Type
Admission Control Traffic Contract
Traffic Source
Traffic Scheduling
Traffic Conditioning (Shaping ••• Policing)
B
I. INTRODUCTION
QoS Management of GPRS
General Packet Radio Services (GPRS), as a major component of GSM Phase 2+ [1, 2], aims at providing public packet-switched data services over the existing GSM network infrastructure. In order to meet the different requirements of a wide variety of user applications, a number of Quality of Service (QoS) profiles have been specified. There is significant current interest in the development of QoS management functions for GPRS networks that meet the requirements of these profiles while maintaining a high bandwidth efficiency. This is especially true since the radio spectrum is a limited and valuable resource. QoS management involves the characterization, shaping/policing and scheduling of user traffic, and admission control of users (Fig.1). In this paper, we focus on traffic scheduling methods for GPRS. Though there are a lot of scheduling methods or service disciplines [3-13] that have been studied, e.g. First-InFirst-Out (FIFO), Static Priority Scheduling (SPS), Virtual Clock, Weighted Fair Queuing (WFQ), Self-Clocked Fair Queuing, Start-Time Fair Queuing (STFQ), Worst-case Fair Weighted Fair Queuing (WF2Q), Earliest Deadline First (EDF), Delay Earliest-Due-Date (Delay-EDD), Jitter EarliestDue-Date (Jitter-EDD), Stop-and-Go, Weighted RoundRobin (WRR), Deficit Round Robin (DRR), Hierarchical Round Robin (HRR), Rate-Controlled Static Priority (RCSP), and Leave-in-time etc, we are not aware of any extensive evaluation of traffic scheduling relative to GPRS requirements (predictive and best effort) in the literature so far. In this paper, the results of the following scheduling methods are considered.
MS
BSS
SGSN
GGSN
PDNs or other networks
GPRS networks
Fig. 1. QoS management of GPRS networks. This paper is organized as follows. Section II describes the QoS profile of GPRS. The descriptions of the scheduling methods are given in Section III. Section IV includes the traffic source models, simulation scenarios and simulation results, and finally conclusions are given in Section V. II. QOS PROFILE OF GPRS There are five aspects in the QoS profile for GPRS, precedence, reliability, mean throughput, peak throughput, and delay classes [2]. There are four delay classes in the GPRS QoS profile: classes 1, 2 and 3 offer what we call predictive services and require QoS management, while class 4 provides a best-effort service. Table I gives the delay requirements for packets containing 128 and 1024 octets of information. To determine the delay requirements for different packet lengths in each predictive service class, we interpolate between the two packet lengths given in the specification. The delay requirement r(c,l) for a packet with any size l and delay class c as given in (1).
0-7803-5669-1/$10.00 (c) 1998 IEEE
Delay Class
1.(Predictive) 2.(Predictive) 3.(Predictive) 4.(Best Effort)
Packet Size 128 octets 1024 octets Mean 95 % Mean 95 % 0.5 1.5 2 7 5 25 15 75 50 250 75 375 Unspecified
0.5 × l c = 1; l ≤ 128 128 1.5 × l 0.5 + c = 1; l > 128 1024 − 128 5×l c = 2; l ≤ 128 r (c, l ) = 128 (1) 5 + 10 × l c = 2; l > 128 1024 − 128 50 × l c = 3; l ≤ 128 128 50 + 25 × l c = 3; l > 128 1024 − 128 where c stands for delay class, l stands for the size of the packet. A quantifiable performance measurement with respect to the different service classes and packet lengths is needed to facilitate comparisons between different scheduling methods. We develop the method of normalizing the measured delays to eliminate packet lengths as a factor in the comparisons. Let pk denote the packet arrived to the destination, class(pk) denote the class of the packet, and pk_len(pk) denote the packet length of the packet. The normalized delay for pk is given by
Experience d real delay of pk Required delay of pk
required delay
256
required delay curve real delay
512
768
1024
bytes
pk Size
Fig. 2. Normalized delay calculation. III. SCHEDULING METHODS A. First In First Out (FIFO) FIFO is the simplest scheduling and queuing method. First arrived packet is served first. Two buffers are used, one for predictive services(class 1, 2 and 3), the other for best-effort service(class 4). We always use a separate buffer for besteffort servces in all of the scheduling methods, which receives service only if the buffers for the predictive classes are empty. B. Static Priority Scheduling (SPS) With SPS, each delay service class has its own buffer and is assigned a fixed (static) service priority: highest for class 1 and lowest for class 4 (best-effort). When the next downlink time slot is available, a class i buffer will receive service only if all class j (j < i) buffers are empty.
(2)
N
i =1
sec 90 80 70 60 50 40 30 20 10 0 0
C. Earliest Deadline First (EDF)
The required delay of pk is calculated by the function r (class( pk ), pk _ len( pk )) indicated by (1). The mean normalized delay is given by
∑ Normalized delay of pk
Delay Requirement
TABLE I DELAY CLASSES AND REQUIREMENTS FOR DIFFERENT PACKET SIZES
i
(3) N Using the measurement of normalized delay, it is more convenient to evaluate the the delay performance of the queuing system with variable packet size and different delay classes. If the mean normalized delay is below 1, we can make conclusion that the delay performance requirement is met.
With Earliest Deadline First (EDF) or Earliest Due Date (EDD) method, each arrived packet has its own deadline (or due-date). The packets are served in the order of their deadlines. Assume the arrival time of a packet is a, and the length of the packet is l. Its delay class is c(1≤c≤3), the time-slot capability of its destination is s, and rate denote the data rate of one time-slot. Then we have l dead _ line = a + r(c, l) (4) s × rate where, the function r(c,l) denotes the delay requirement of the packet with delay class c and length l calculated by (1). The EDF mechanism needs to sort the packet queue using at least O(logN) insertion operation for each arrived packet. This may affect its application due to implementation difficulty.
0-7803-5669-1/$10.00 (c) 1998 IEEE
IV. PERFORMANCE EVALUATIONS A. Traffic Sources To generate the traffic presented to the simulator (by the sources), we used three types of traffic sources and their combinations including email, fleet management and WWW applications. FUNET model was used for Email and MOBITEX model for fleet management traffic generation [14]. For generating a realistic WWW traffic, which takes into account the self-similar nature of the traffic, we used available traces. 1) FUNET model This model is based on statistics collected on email usage from the Finish University and Research Network (FUNET). TABLE II shows the percentage of messages shorter than the indicated length. TABLE II DISTRIBUTION OF FUNET MODEL Percentage 10 36 54 67 79 87 91
The trace is downloaded from the web site http://ita.ee.lbl.gov/html/traces.html, where stores the socalled Internet Traffic Archive. "The Internet Traffic Archive is a moderated repository to support widespread access to traces of Internet network traffic, sponsored by ACM SIGCOMM. The traces can be used to study network dynamics, usage characteristics, and growth patterns, as well as providing the grist for trace-driven simulations. The archive is also open to programs for reducing raw trace data to more manageable forms, for generating synthetic traces, and for analyzing traces". Fig. 3 shows the data rate of the trace we used. The WWW document size distribution in the trace is given in TABLE III. TABLE III WWW DOCUMENT SIZE DISTRIBUTION Document size(bits) >10,000,000 >1,000,000 >100,000 >10,000
Number 28 224 3389 21267
Rate(105bps)
Message length (Kbytes)