L-EDD Algorithm for Handling Streaming Traffic in Tactical Networks

0 downloads 0 Views 263KB Size Report
IN TACTICAL NETWORKS. Damian Duda and Piotr Pyda. Military Communication Institute. 05-130 Zegrze, Poland phone: +48 22 688 55 92, fax: +48 22 688 55 ...
L-EDD ALGORITHM FOR HANDLING STREAMING TRAFFIC IN TACTICAL NETWORKS Damian Duda and Piotr Pyda Military Communication Institute 05-130 Zegrze, Poland phone: +48 22 688 55 92, fax: +48 22 688 55 44 e-mail: [email protected]; [email protected] ABSTRACT The well known scheduling algorithm, called EDD (Earliest Due Date) has some useful abilities in servicing real time data. They are:(1) providing hard delay bound for serviced traffic and (2) relative differentiation of queuing delay. The main disadvantages of EDD algorithm are inability to differentiate loss ratio in presence of congestion and necessity to implement a sorted queue to support the mechanism of deadline handling. In the paper we propose improvement of EDD service discipline by introducing similar mechanism as in Round Robin algorithm. To avoid implementation of sorted queue the emulation of EDD by multiple FIFO system was used. The simulation results show that there is possibility to differentiate loss ratio using such modified algorithm.

INTRODUCTION In activities of headquarters, radar posts and other units there are terminals, multimedia applications, etc. that generate traffic with different characteristics [1]. Mainly, voice and data applications are used, which produce traffic divided into two classes - streaming and elastic. Each class has different requirements according to loss ratio, delay and jitter. Class generating elastic traffic requires low losses, but it is not directly sensitive to delay characteristics. In the contrary, sources that produce streaming type data are sensitive to transfer delay and delay variation. In addition, such kind of data often have their expiry time. It means that the data are useless when transfer time through the network exceeds predefined threshold. On the other hand, voice and video transmission has ability to accept relatively high losses. The measurement results obtained during tests in the Military Communication Institute show, that in the real ATM tactical network cell loss ratio at 10-3 level allows

proper handling of voice calls and data exchange. Exemplary characteristics of data from applications and their requirements for bandwidth, delay and losses are presented in the Table 1. Table

1. Characteristics of data from applications in ATM tactical network

Application Telephony Video telephony Data from radar Remote control

Type of streaming traffic Constant or Variable bit rate Constant or Variable bit rate CBR/VBR CBR

Bandwidth [kbps]

Requirements Delay [ms]

64

Low ≤200

n*64

Low ≤200

a few a few

Low Medium

Losses High CLR up to 10-3 High CLR up to 10-3 Low Low

In a tactical network the network nodes are interconnected by links with limited capacity, usually 2 to 8 Mbps. Because of increasing needs of bandwidth from applications in headquarters, the proper allocation of link capacity is always an important issue. Additionally, network should guarantee quality of data transfer at specified level. This one is achieved by implementation of suitable mechanisms of resources allocation. One of such mechanisms is the scheduling algorithm used in network nodes, which plays essential role in the process of resources allocation. The simplest and widely used in today’s, especially civilian networks is FIFO (First In First Out) algorithm. It is easy to implement in hardware as well as in software platforms. The most significant disadvantage of FIFO is that it treats traffic from different applications in the same way. So, particular requirements, which are necessary to guarantee demanded quality of service, may not be met. Obligation to provide guaranteed quality of service, which are specific to military networks, constrain applying more sophisticated scheduling mechanisms. Some of feasible mechanisms are Static Priority, Fair

Queuing, Earliest Due Date (EDD) and their modifications. Because of its properties, the scheduling algorithm EDD, also known as EDF (Earliest Deadline First), seems quite attractive for modern military applications. The useful feature of EDD is that it guarantees specified delay bound, which allows dropping “out of date” packets/cells. It is crucial for application demanding real-time data, e.g. recognised air picture. The EDD in its basic form, known as Classic EDD (CEDD), has ability to relative differentiation of queuing delay between served flows. The inconvenience of CEDD is that it does not differentiate losses between flows in the case of servicing homogenous traffic [2, 4]. This simply results in a lack of isolation between flows [5]. It should also be noted, that the algorithm needs maintaining of queue sorted according to assigned deadlines to cells/packets. This requirement generates serious implementation issues in high speed networks [3]. These mentioned above, two negative characteristics were the reason for researching work in finding more functional and less complicated enhancements of CEDD. The first problem, that is inability to differentiate losses between served flows has interesting solution in [2]. The presented Weighted Earliest Due Date (WEDD) algorithm can explicitly provide different deadline violation probabilities, proportional to given weights. This benefit is obtained at the cost of computational complexity of algorithm, which is relatively high – WEDD uses sorted queue, real-time computation of loss ratio and, in overload conditions, changes handling procedures of queue. According to the problem of the computational complexity, much work has been done in decreasing of complexity of EDD algorithm. One of the way is described in [3]. The proposed Rotating Priority Queues+ (RPQ+) algorithm takes advantage of using multiple FIFO system and traffic classification based on deadlines from limited set of values. In this paper the Limit based Earliest Due Date (LEDD) algorithm is proposed. Our objective was to develop algorithm with functionality like WEDD, i.e. differentiation of

delay and losses but with less complicated implementation. In the next sections of the paper we will describe functional architecture of LEDD algorithm, simulation model and scenarios, numerical results and conclusions. LEDD ALGORITHM DESCRIPTION The LEDD algorithm is an enhancement of EDD algorithm to support differentiation of delay and losses. It uses similar mechanism as in Round Robin algorithm to choose next packet/cell to service. We expect that such modified EDD scheduler will better treat good behaving flows in present of overload (e.g. caused by misbehaving sources). To decrease complexity of algorithm we did not apply the sorted queue. LEDD algorithm uses multiple FIFO system to serve flows with different deadlines TD. Using of such a system allows aggregation of flows into classes according to their deadlines. We assume that the range of deadline values is limited. Deadlines assigned to cells are not continuous set of values, but belong to defined, finite subset D={d1; d2; …; dn ), where n is the number of classes. Next, such classified traffic is served by multiple FIFO system, where a single FIFO queue is assigned for each of class. The main assumption for LEDD is that for each class (FIFO buffer) a limit counter LC is assigned. The selection to service of particular packet/cell is based on the following rule: − a cell with smallest deadline time TD from buffers is chosen for service, but only if limit LC of buffer storing the cell is greater than zero; − if the cell was selected to service from a buffer, then limit LC assigned to this buffer is decreased by 1; − the buffer which limit is exhausted does not participate in process of searching cell for service, with exception described in the next point; − if the buffer/buffers with positive limit are empty, then searching cell in another buffer (with limit equal to zero) is allowed; − if all buffers’ limits reach zero, then limits are set to their initial values.

When the limit of a particular flow reaches zero, then appropriate buffer is temporary unavailable for the server. Traffic waiting in such a buffer suffers large delay and losses due to its increased deadline violations ratio. But it is profit for other classes, especially to those with larger limit. Limit Counter assigned to each buffer counts successful entering of a cell into service. Dropped cells do not change value of Limit Counter. Functional architecture of proposed LEDD algorithm is depicted in the Figure 1. For the purpose of simulation the discrete event simulator was implemented using C++ programming language. Flows

F1

F2

Fi

Deadline assignment

FIFO queues

Limit Counters

Selection for service

Q1 TA+D1

LC1 LEDD Scheduler

Q2 TA+D2

C

LC2

Qi TA+Di

LCi

deadline time TD. Time of expiration of a particular cell (TD) is computed as a sum of the cell’s arrival time and constant, defined for each class value of deadline parameter D. In addition, dropped cells do not modify values of the limit counters. There were three tests provided to investigate behavior of proposed algorithm. The first and second test compared abilities of Classic EDD and L-EDD algorithms for providing delay and losses differentiation respectively. The second test examined handling of CBR traffic in the presence of congestion caused by traffic with Poisson characteristics. Incoming traffic was Poisson distributed in the first two tests. The third test was performed using CBR sources with Poisson traffic in the background. Results of simulation were obtained during simulated time interval, which provided at least several millions of events. The functional diagram of simulated system is depicted in the Fig. 2.

Discarding (optional)

LEDD – Limits based EDD

Traffic of class 1

Deadline assignment

Limit Counters FIFO 1

TAi+D1

Fig.1. The proposed LEDD algorithm – deadline and limit based EDD scheduling

SIMULATION RESULTS In this section the model of system with LEDD scheduling discipline and simulation scenarios are described. Three simulation tests with exemplary numerical results are presented. The simulation model was created based on the following assumptions. First of all, as transfer of data in an ATM tactical network is cell based, we fixed the time of service to duration of a single slot (transfer time of ATM cell through the outgoing link). Capacity of simulated outgoing link corresponds to the tactical E1/ATM link. The system has two simulated FIFO buffers, so there are two classes available. To avoid distortion of simulation results by buffer overflows each FIFO buffer has infinite size (in fact, they are limited by available memory of computer). Service discipline is based on EDD algorithm, which means that cells are dropped when their waiting time exceeds defined threshold, called

Traffic of class 2

LC1

LEDD Scheduler

C

FIFO 2 TAj+D2

TA – arrival time

LC2

Discarding of “out of date” cells

Fig.2. Functional diagram of implemented LEDD simulator

In the first test the comparison of Classic EDD and L-EDD algorithm was performed. The objective was to present specific behavior of CEDD and LEDD in providing relative differentiation of queuing delay and losses. The test included two simulation scenarios, aimed to obtain delay and losses statistics for the first and second scenario respectively. There were two classes defined. The parameters of simulation were set as follows: − offered traffic arrives according to Poisson process; − total load of system equals to ρ=0.8 (scenario 1) or ρ=0.8 ÷ 1.2 (scenario 2); − two classes were set up: Class 1 with deadline equal to 10 slots, Class 2 with deadline equal to 40 slots;

− Limit Counters (LC) for LEDD were set to equal values: LC1 = LC2 = 50 cells; − Load shares for each class were set to 50%;

0

t [slot]

10

20

30

40

1.E+00

1.E-01

Class 1 Class 2

1.E-03

1.E-04

1.E-05

1.E-06

Fig.4. CCDF of queuing delay of serviced cells by L-EDD algorithm – load shares: Class 1 (Poisson)50% with D1=10 slots, LC1=50 cells; Class 2 (Poisson)- 50% with D2=40 slots, LC2=50 cells Total Load 0.7

0.8

0.9

1

1.1

1.2

1.3

1.E+00

1.E-01

CLR

The results of simulation are depicted in the Figures 3 to 6. Figures 3 and 4 contain graphs of complementary cumulative distribution function (CCDF) of queuing delay of successfully serviced cells. Figures 5 and 6 present cell loss ratio (CLR) dependence on offered load. As we expected, the C-EDD and L-EDD have abilities to relative differentiation of queuing delay. The conditional CCDF graphs of delay for both of algorithms are similar, i.e. they have distinct knees at the values of delay equal to deadline for particular class. Introduction of limit counters in L-EDD changed the CCDF curve for both of classes. In particular, the CCDF function of delay has greater values for class with smaller deadline and less values for classes with greater deadline. This means that for the same conditions as in C-EDD, the mean waiting time is greater for Class 1 (with smaller deadline) and smaller for Class 2 (with greater deadline). Increasing of waiting time for a cell with assigned deadline means that there is a greater probability of dropping that cell from the queue. This is clearly visible in the Figures 5 and 6, presenting CLR for both of algorithms.

P{W>t}

1.E-02

1.E-02

Class 1 Class 2

1.E-03

1.E-04

1.E-05

1.E-06

Fig.5. No differentiation of losses by Classic EDD algorithm – load shares: Class 1 (Poisson)- 50% with D1=10 slots; Class 2 (Poisson)- 50% with D2=40 slots Total Load 0.7

t [slot] 0

10

20

30

1

1.1

1.2

1.3

40

1.E-01

Class 1 Class 2

1.E-03

1.E-04

CLR

P{W>t}

1.E-02

0.9

1.E+00

1.E+00

1.E-01

0.8

1.E-02 1.E-03

Class 1 Class 2

1.E-04 1.E-05 1.E-06

1.E-05

1.E-06

Fig.3. CCDF of queuing delay of serviced cells by Classic EDD algorithm – load shares: Class 1 (Poisson)- 50% with D1=10 slots; Class 2 (Poisson)- 50% with D2=40 slots

Fig.6. Differentiation of losses by L-EDD algorithm – load shares: Class 1 (Poisson)- 50% with D1=10 slots, LC1=50 cells; Class 2 (Poisson)50% with D2=40 slots, LC2=50 cells

Total Load 0.7

0.8

0.9

1

1.1

1.2

1.3

1.E+00

CLR

1.E-01 1.E-02 1.E-03

Class 1 Class 2

1.E-04 1.E-05 1.E-06

Fig.7. Differentiation of losses by L-EDD algorithm – load shares: Class 1 (Poisson)- 50% with D1=10 slots, LC1=50 cells; Class 2 (Poisson)50% with D2=40 slots, LC2=25 cells

Total Load 0.7

0.8

0.9

1

1.1

1.2

1.3

1.E+00

1.E-01

CLR

C-EDD algorithm does not provide differentiation of losses, which is also stated in [2, 4]. In the L-EDD losses are differentiated according to changes made to CCDF of queuing delay. In the Figure 6, it is noticeable that losses for Class 1 were increased, although limits for particular classes were equal. Functioning of Limit Counter mechanism leads to a situation, when cells with smaller deadline are not served because limit for appropriate queue is exhausted. Such cells suffer larger delay and then it is more likely to drop them. In the second test the capabilities of losses differentiation of LEDD were evaluated. The test included two scenarios, which were performed with Limit Counters set do different values. First, scenario was performed for the case where LC of Class 1 had greater value than LC of Class 2. They were LC1=50 cells, LC2=25 cells. In the second case limits were inversed, and had values: LC1=25 cells, LC2=50 cells. The rest of parameters had the same values as in the first test. The results of simulation are shown in the Figures 7 to 8. It is clearly visible, that the class with smaller LC suffers larger losses. The differentiation of CLR for classes is done by changing values of limits assigned to particular queues. Notice, that losses not only depends on limits but on deadline values too. Results presented in the Figure 7 are similar to ones obtained for WEDD algorithm [2].

1.E-02

1.E-03

Class 1 Class 2

1.E-04

1.E-05

1.E-06

Fig.8. Differentiation of losses by L-EDD algorithm – load shares: Class 1 (Poisson)- 50% with D1=10 slots, LC1=25 cells; Class 2 (Poisson)50% with D2=40 slots, LC2=50 cells

The third test included three simulation scenarios with CBR traffic handled by Class 1 and Poisson traffic handled by Class 2. The scenarios were performed to obtain losses statistics with Limit Counters set do different values. The parameters of simulation were set as follows: − Class 1 – 5 CBR sources 64 kbps each, deadline D1=10 slots, shares 18 % of outgoing link capacity; − Class 2 - Poisson distributed traffic, deadline D2=40 slots, with varying mean arrival rate; − total load of system varied: ρ=0.8 ÷ 1.2; Scenario 1 was performed with limit counters LC set to equal values LC1=LC2=50 cells. Scenario 2 was run for LC of class 1 greater than LC of class 2. They were LC1=50 cells, LC2=25 cells. In the scenario 3 limits were inversed, and had values: LC1=25 cells, LC2=50 cells. The results of simulation for particular scenario are shown in the Figures 9 to 11 respectively. Like in the first test, the ability of L-EDD to differentiate losses was observed. Notice, that on the Fig. 10, where the limit counter for CBR had greater value than counter for Class 2, there were no losses observed during simulations.

CONCLUSIONS

Total Load 0.8

0.9

1

1.1

1.2

1.3

1.E+00 1.E-01

CLR

1.E-02 Class 1 Class 2

1.E-03 1.E-04 1.E-05 1.E-06

Fig.9. Differentiation of losses by L-EDD algorithm – Class 1 (CBR)- with D1=10 slots LC1=50 cells; Class 2 (Poisson) with D2=40 slots, LC2=50 cells Total Load 0.8

0.9

1

1.1

1.2

1.3

1.E+00

CLR

1.E-01 1.E-02 Class 1 Class 2

1.E-03 1.E-04 1.E-05 1.E-06

In the paper the L-EDD algorithm was proposed. The algorithm has properties which make it suitable for handling streaming traffic in tactical network. The simulation results for two serviced classes show that L-EDD algorithm allows differentiation of loss ratio among classes. The differentiation is relative which means that improvement of performance for one class implies degradation of another one. Simulation tests

were performed for two types of traffic, Poisson and CBR. Additionally, simulation results proved that there is a possibility to create a privileged class, with stringent requirements concerning delay and losses. Especially, CBR traffic can be handled with practically no losses (see Fig.10). Additional advantage of L-EDD is that it has simpler implementation than Classic EDD. By using multiple FIFO system maintaining of sorted queue is no longer needed. Future work should include finding the method of limits assignment for handling different types of traffic. Bibliography

Fig.10. Differentiation of losses by L-EDD algorithm – Class 1 (CBR)- with D1=10 slots LC1=50 cells; Class 2 (Poisson) with D2=40 slots, LC2=25 cells. Note: For Class 1 no cell losses were observed Total Load 0.8

0.9

1

1.1

1.2

1.3

[1] A.Bęben, W.Burakowski, P.Pyda: „A Comparison of ATM and IP QoS Network Capabilities for Handling LAN Traffic with QoS Differentiation“, Proc. of the 4th NATO Regional Conference on Military Communications and Information Systems (RCMCIS 2002), Oct. 2002, pp.157-163

1.E+00

CLR

1.E-01 1.E-02 Class 1 1.E-03

Class 2

1.E-04 1.E-05 1.E-06

Fig.11. Differentiation of losses by L-EDD algorithm – Class 1 (CBR)- with D1=10 slots LC1=25 cells; Class 2 (Poisson) with D2=40 slots, LC2=50 cells

[2] S. Bodamer: „A New Scheduling Mechanism to Provide Relative Differentiation for RealTime IP Traffic”, Proc. of the IEEE Global Telecommunications Conference (Globecom ‘00), San Francisco, Nov. 2000, pp. 646-650 [3] Dallas E.Wrege, Jörg Liebeherr: “A NearOptimal Packet Scheduler for QoS Networks”, Proceedings of the IEEE Infocom (1997) [4] V.Sivaraman, F.M.Chiussi: “Statistical Analysis of Delay Bound Violations at an Earliest Deadline First (EDF) Scheduler”, 1999 [5] Pawan Goyal, Harrick M. Vin and Haichen Cheng: „Start-Time Fair Queueing: A Scheduling Algorithm for Integrated Services Packet Switching Networks”, IEEE/ACM Transactions on Networking, Vol. 5, No. 5, October 1997

Suggest Documents