SLA Framework for Effective Telemedicine using DEA S. Rajeev1, K.V. Sreenaath2, A.S.Bharathi Manivannan3, S.Anusha4 1 Department of Electronics & Communication Engineering 2 Department of Information Technology 3 Department of Electrical & Electronics Engineering 4 Department of Mathematics and Computer Applications PSG College of Technology, Coimbatore, India Phone: (0422) 2572177 e-mail: [email protected], [email protected], [email protected], [email protected]

Abstract: Telemedicine is the exchange of medical information from one site to another using electronic communications for the health and education of patients or providers and to improve patient care [1]. Internet is the widely used communication medium for telemedicine [2]. With the use of a public domain (internet) for real time telemedicine, Quality of Service (QoS) becomes a critical factor. Telemedicine through Internet lacks efficient service delivery to the patients [3]. To overcome this, effective service level agreement (SLA) [4] should be traded between the patient node (or patient domain) and service providing end (either doctor’s node or a node containing junk of remedial datas that is automated to give suitable reply to the ailing patient). Performance metrics such as bandwidth, service delay, throughput, queuing delay, jitter, congestion, packet loss play an important role for effective telemedicine through internet. SLA is a service contract between a customer and a service provider that specifies the forwarding service a customer should receive. Earlier works on SLA trading algorithm such as “Provisioning algorithm and Profitability Analysis algorithm” [5], does not suit effective telemedicine through internet as it takes into consideration very few performance metrics for trading. So provisioning algorithm and Profitability Analysis algorithm with appropriate enhancements is proposed to suit for real time telemedicine through internet. In this paper a mathematical model and a SLA trading algorithm is given to support decisions on effective SLA between the patient end and the service providing end. The Framework uses a Policy Server and Data Envelopment Analysis (DEA) [6] to take decisions on the service trading issues and to enhance them. The decision on what to choose and when to choose is written as policies. The policy server takes the decisions by taking appropriate policies from the directory server using Lightweight Directory Access protocol (LDAP) [7]. Keywords- Telemedicine, Service Level Agreement, Quality of Service, Policy Based Networks Directory Enabled Networks.

1. Introduction Telemedicine is the process of exchanging the medical information through an electronic medium such as internet [1]. Through telemedicine patients query the doctors through an electronic medium such as internet and get the response back from the doctor. Quality of Service becomes an important and critical factor when internet is used for telemedicine. Telemedicine is a real time communication process for which the best possible service should be granted. But when telemedicine is done through a public domain such as internet guarantying the best possible service becomes a difficult task. In this paper we have modeled a “SLA Framework for Effective Telemedicine using DEA” for providing the best possible service for telemedicine through internet. Performance metrics that affect the QoS such as Bandwidth, Throughput, Jitter, etc., are considered in this model. A mathematical model is developed for Service Level Agreement (SLA) between the client (patient) and the service provider (doctor). An effective SLA trading algorithm is also developed for trading of SLAs between the client and the service provider. Policy Speicifcation for SLA trading is developed and tested along with Data Envelopment Analysis (DEA). The rest of the paper is organized as follows: A Mathematical Model and a Policy Based Framework for Effective Telemedicine are given in section 2 and section 3. Section 4 describes the SLA Trading algorithm. Policy specification, Analysis using DEA and Results & Discussion are given in section 5, 6 and 7. Conclusions are drawn in section 8. 2. Mathematical Model A mathematical model considered for our framework uses Linear Programming and Simplex Method [8] to solve it. The following performance metrics that are crucial for effective SLA trading are considered in our model.  Bandwidth  Delay  Demand  Packet Loss  Congestion  Queuing Delay  Throughput  Buffer Capacity  Jitter Let, Tij -Total (maximum) Bandwidth (channel capacity) available from host i to host j (a host may be either the client (patient) or service provider (doctor)). U ij -Bandwidth being used for traffic flow between host i to host j at instant ’ t ’ Rij -Reserved bandwidth from host i to host j .

Hence the bandwidth that can be leased to other hosts Gij is given by Gij  Tij  U ij  Rij Let the required bandwidth i.e. the bandwidth consumed by the host k to reach host j through host i be RBij .

And, Dij -Delay from host i to host j C ij -Cost of reaching host j through host i . Fij -Fraction of bandwidth bought from host i to reach host j .

The objective here is to minimize the cost of reaching host j through other hosts. Minimise  Fij C ij

(1)

i, j

As stated earlier, C ij in the above equation represents the cost host i charges to reach host j through host i . 2.1 Constraints There are a set of constraints that define the model. The first constraint is that the demand for bandwidth to reach host j through host i , DEij should be less than or equal to the amount of bandwidth host i is ready to offer for cost to reach host j , Gij , DEij   Gij

(2)

i, j

The following constraints check if the Service performance metrics in the service offered by the host i to reach host j fall within the predetermined and pre-calculated boundaries as expected by the host k which needs the service. These boundary constants for the performance metrics can also be set dynamically and SLA negotiated accordingly. Buffer Capacity Bij should not be less than a bearable value given by the constant N=Number of packets that can be buffered Bij  N (3) The time delay D should be set to a limit expressed by a constant ‘ p1 ’ as expected by the ‘ISP k’ which needs the service. The constant ‘ p1 ’ is arrived as derived below [9] p1 = Propagation Time + Transmission Time + Queuing Delay (+ Setup Time) Propagation Time: Time for signal to travel length of network = Distance/Speed of light Transmission Time = Size/Bandwidth Therefore, we have Dij  p1 (4) Queuing Delay Qij should not exceed an allowable limit ‘ p 2 ’expressed as [10] p2 

D   N  1 2

where, D -the time delay, N is the Buffer Capacity Qij  p 2

(5)

The Packet Loss Pij for the service provided should not exceed a maximum limit set as constant‘ p3 ’ and Congestion in the channel offered for service Coij should also be

within the acceptable limits represented by the constant ‘ p 4 ’ both of which are arrived as shown below [11] Tmin = Minimum Inter-Arrival Time observed by the receiver P0 : Out of order packet Pi : Last in-sequence packet received before P0 Tg : Time between arrival of packets P0 and Pi .

n : Packets missing between Pi and P0 If (n  1)Tmin  Tg  (n  2)Tmin then n missing packets are lost due to transmission errors and hence ‘ p3 ’=’ n ’ and Pij  p3

(6)

Else n missing packets are assumed to be lost due to congestion and hence ‘ p 4 ’=’ n ’ and Coij  p 4

(7)

Throughput TH ij should be greater than or equal to ‘ p5 ’ which is given by [12]

p5  {MSS / RTT }  C /( p ) where MSS - Maximum Segment size in bytes. Typically 1460 bytes RTT -Round Trip Time in seconds, measured by TCP. p - Packet loss C - Constant assumed to be 1. TH ij  p5

(8)

The jitter J ij should be within the acceptable limit ‘ p 6 ’ given by [13]

p6  p6  ( D(i  1, i )  p 6) / 16 given D(i, j )  ( R j  S j )  ( Ri  S i )

where S i , S j are sender timestamps for packets i, j and Ri , R j are receiver timestamps for packets i, j . Therefore J ij  p 6

2.2 Non-Negativity Constraints The following are the non-negativity constraints applied in the model:

(9)

Cost C ij should always be positive, C ij  0

(10)

Fraction of bandwidth bought from host i to reach host j , Fij should also be positive, Fij  0

(11)

The bandwidth that can be offered for cost to other hosts by host i should be positive, (12) Gij  0 Given the objective i.e. to minimize the agreement cost along with the performance metrics constraints, the proposed linear programming model solved using simplex method suffices for arriving at a suitable agreement for service with other hosts. There are always cases that the above model will fetch more than one solution if there exists. Hence in such cases the decision of choosing the most appropriate of the available solutions should be taken using policies stored in directory server. 3. Policy Based Framework SLA for Effective Telemedicine The growing interest in the field of Policy-Based Networking [14] to monitor and control the access rights of resources in large distributed systems, and in areas like Quality of Service (QoS), Wireless Networks [15, 16], Network Security, SLA and IP address allocation etc., have identified us to use policy based approaches to SLAs, as SLAs are normally setup manually [17]. Internet is a very large-scale network which is also very dynamic. Thus the SLAs cannot be set up manually. Using the mathematical model given above the effective Service Level Agreement can be traded between the Service provider (doctor) and the clients (patients) satisfying the constraints. The architecture for the Policy Based SLAs for effective telemedicine is given in the following section. 3.1 Architectural Framework of the Policy Based SLA for Effective Telemedicine Policy Based approach is used in this architecture to enforce relevant policies for each client (patient) after effective service level agreement is traded between the client and the service provider and also to choose a better agreement when more options end up in trading for SLA. In the architecture shown in Fig.1 policies are stored in the Directory server. In this case, the client (patient) requests the service provider (doctor) for service. The service provider (doctor) then queries the policy server which takes appropriate policies from the directory server through LDAP. The policy server also communicates with other servers such as Time Servers, Certificate Servers and Authentication, Authorization and Accounting (AAA) servers and validates the clients’ requests by means of certificates and AAA. The policy server takes the decision on whether the client is an authenticated one and his services are authorized with accountability and certificates. After validating the clients’ request and enforcing other policies the clients’ request is responded by the service provider.

Fig.1 SLA Framework for Effective Telemedicine

Fig 2: Simulation Environment

4. SLA Trading Algorithm

Table 1: Parameters for SLA Algorithm UPDATE_PERIOD Time for updation volume() volume function for an SLA object(time  bandwidth) send bid() sends an offered SLA to peer accept bid() sends an accept message known dests reachability list MIN_REQUIREDBW Minimum required bandwidth expressed in Mbps or in Kbps MAX_ALLOWED_DELAY Maximum allowed delay expressed in unit of time such as seconds or in milliseconds MAX_ ALLOWED_PL Maximum allowed packet loss expressed as numeral representing the no. of packets MIN_ REQUIRED_BC Minimum required buffer capacity expressed as numeral repressing the number of packets MAX_ ALLOWED_QD Maximum allowed queuing delay expressed in unit of time such as seconds or in milliseconds Minimum required throughput expressed as MIN_ REQUIRED_TP numeral with each numeral representing throughput in the network Maximum allowed jitter expressed as numeral MAX_ ALLOWED_JI with each numeral representing jitter severity in the network Maximum allowed congestion factor expressed MAX_ ALLOWED_C as numeral with each numeral representing congestion severity in the network Maximum allowed cost expressed in unit of MAX_ ALLOWED_COST currency such as rupees or in $ Table 2: SLA Trading Algorithm struct bid { Host_dest, // Destination Host bw, // bandwidth delay, packet_loss, buffer_capacity, queuing_delay, throughput, jitter, congestion, cost } process trading () { while (true) { for each d in known_Host_dests { /* buy bids */

if (bw_to_Host(d)>MIN_REQUIREDBW) and (delay_to_Host(d)MIN_ REQUIRED_BC) and (queuing_delay_to_Host(d)MIN_ REQUIRED_TP) and (jitter_to_Host(d)
enhancements is proposed and given in Table 2. The constants used in Table 2 are given in Table 1 the values for which are derived as explained in section 2.We call the algorithm responsible for the determination of what resources are needed the provisioning algorithm. A passive provisioning algorithm does wait for requests from its customers to select which resources to buy. An active provisioning algorithm tries to forecast future needs. It will then buy resources in advance, before they become scarce. Buying in advanced may be based on statistical information (e.g. previous weeks usage by time of day) or on trend analysis. Once an SLA trader knows it needs to buy some resource from one of its peers, it will have to select one of the bids and buy it. The selection of the bid is made based on the bid’s value for the SLA trader and its price. For bids of equal value, if no special policy is applied, the bid with the lower price will be selected. The SLA trader will also have to evaluate if the selected bid is worth buying using a profitability analysis algorithm. This algorithm does evaluate if by buying that bid, money will be made through the selling of derived services. It is this algorithm which will also ensure that SLA traders won’t build service loops. Trading is done by the method ‘process trading()’ which finds the bid with the highest volume/cost ratio and finds out if that bid is profitable using the ‘is_profitable(bid)’ method and if found profitable accepts the bid using the ‘accept_bid(bid)’ method. Then for each neighbour (Hosts) if bid is not already sent then bid is sent for every UPDATE_PERIOD. The profitability is tested by comparing the bid price with the expected income. Bidding with the neighbouring hosts is done using the ‘make_bid()’ method. 5. Policy Specification for SLA Trading The policy Specification for the SLA trading is given below. The policy specification takes into account the various performance metrics. The policy specification can be run either in the policy server in the wired network or as a module in the policy service inside the host itself. // On a trading event the action trade sends a pop-up displaying that the trading is being analyzed to check if all the performance metrics constraints are satisfied with minimal cost based on the mathematical model proposed in section 2.

Table 3: Policy Specification for SLA Trading inst oblig/Policies/TradingPolicy { on trading (host_name,required_bandwidth,delay,,congestion, queuing_delay,buffer_capacity,throughput); subject /PMAs/TradePMA; do trade(host_name,required_bandwidth,delay,,congestion, queuing_delay,buffer_capacity, throughput); }

The rule /Policies/Trading Policy will invoke the action trade within the /PMAs/TradePMA’s engine, when the event trading is dispatched to the PMA from the trading event service. The corresponding java code which enforces this policy is given in Table 4.

Table 4: Java code for the policy Specification for SLA Trading public void execute(LinkedList params) throws Exception { // parameter0: The string representing the trade that will be considered // For debugging: if (DEBUG) { System.out.println("*******"); System.out.println("Trading the offer of : "+ (String) params.get(0)); } // Pop up the action window JFrame window = new JFrame(); window.setTitle("Trading Console"); JLabel traders = new JLabel("Trader: "+ (String) params.get(0)); JPanel mainPanel = new JPanel(); mainPanel.add("Center",Trader); window.setContentPane(mainPanel); window.pack(); window.setSize(200,100); window.setLocation(100,50); window.setVisible(true); //the code for the simplex method to solve the Linear Programming model as given in section 2 should be added here }// end of method execute

6. Analysis Using DEA Data Envelopment Analysis (DEA) is a technique used to examine the relative performance of organizational units, which carry out similar functions. The 100% efficient unit is identified as ‘best practice’ and a peer based comparison is made, to identify potential improvements of inefficient units. DEA is used in this architecture to improve the Quality of Service being provided to the patients. The service provided to the client is an imporatant factor in “Telemedicine in Internet” as the patients should always get the best possible service. In DEA, the efficiency is measured by the ratio of the aggregated outputs to aggregated inputs. Successful application of DEA requires a number of units (Decision Making Units or DMU’s) which are performing a similar process. An input is any resource used by a unit to produce its outputs. An output is any product or service produced by a unit or a measure of how effectively a unit has achieved its goals. After the required data inputs and outputs are identified, the next step is to

make decisions about the type of model that you want to use to analyze your data. The first step is to determine whether the analysis is for minimizing inputs or maximizing outputs for the process which is under consideration. The second is whether to assume constant or variable returns to scale. The two models used are the BCC model [19], for variable returns to scale and the CCR model [20] for constant returns to scale. CCR model is used for analysis of “Policy Based SLA Architecture for EEN” to measure the total efficiency. The CCR model is as follows:

s   rYrk 1 Maximize E  rm k vX i1 i ik

(13)

s   rYrj 1  1; j  1,..., n Subject rm vX i 1 i ij

(14)

 r    0; r  1,...., s

(15)

vi    0; i  1,...., m Where Ek : the relative efficiency of DMU ‘k’;

 r : the weight given to output r; vi : the weight given to input I; Yrk : the amount of output r from DMU k; Xik : the amount of input I from DMU k; Yrj: the amount of output r from DMU j; Xij : the amount of input I from DMU j;

 : a small ‘non-Archimedean’ constant; n : the number of DMU; s : the number of outputs; m : the number of inputs 7. Results and Discussion The CCR model is constructed with the following performance metrics (inputs): 

ReB- Remaining Bandwidth(MBps)



DB- Demanded Bandwidth of the customers(MBps)



GB: Given Bandwidth(MBps)

(16)



QD: Queuing Delay(x 10-4 s)



BC: Buffer Capacity – number of packets



TH: Throughput (x103Bits/sec)



Delay (x10-3sec)



PLF: Packet loss factor



Jitter( x 10-4 /Sec)



Congestion

Efficiency of performance is measured using Quality of Experience (QoE) [21]. QoE is defined as the level of satisfaction with a service from the perspective of that customer, based on their needs, wants and expectations. The output QoE factor is expressed as a number between 0 and 1 which is obtained based on client (patient) satisfaction levels with the performance metrics obtained using the Qualnet network simulator.

P1 D P2 D P3 D P4 D P5 D P6 D P7 D

ReB DB 3.0 3.0 2.0 4.0 1.0 3.0 4.0 5.0 5.0 2.0 2.5 3.0 2.0 1.0

GB 3.0 2.0 1.0 4.0 1.0 2.0 1.0

Table 5: Analyses of Performance Metrics Performance Metrics Delay PLF Congestion QD BC TH Jitter QoE 3.00 0.30 0.30 0.10 120 4.30 4.0 2.90 0.20 0.30 0.20 100 4.20 3.8 0.625 3.00 0.36 0.23 0.15 110 6.00 2.5 3.10 0.30 0.35 0.20 110 3.90 3.4 0.375 2.80 0.50 0.25 0.60 105 5.00 1.3 2.60 0.55 0.14 0.25 115 4.55 5.1 0.625 2.50 0.35 0.19 0.10 136 2.22 3.2 2.50 0.32 0.15 0.32 140 2.41 3.6 0.75 1.50 0.15 0.35 0.35 100 2.20 4.2 2.00 0.51 0.24 0.22 140 4.10 4.9 0.5 2.40 0.32 0.26 0.11 130 2.36 2.5 2.30 0.18 0.25 0.20 170 5.10 1.6 0.875 2.50 0.31 0.14 0.2 150 5.00 1.6 2.65 0.24 0.15 0.13 140 4.10 5.3 0.25

The DMU’s taken here are the services provided for various clients (patients). From these varied data, using DEA the reference frequency graph is obtained (see Fig.3), which shows that the higher the frequency with which an efficient unit appears in reference sets the more likely that it is an example of good performance. Efficient units which appear in few reference sets are likely to have an unusual combination of inputs and outputs and as such are not likely to offer the best operating practices for inefficient units to emulate. P6 is found to be the highly referenced unit with a count of 3 and good QoE of 0.875. Hence in other instances the Service provider (doctor) should enhance the services to achieve better QoE- close to achieving the local maxima.

Fig.3: Reference Frequency graph In Fig 4, the pie chart shows the totals of input/output potential improvements for all input/outputs over the whole set of units. This provides a quick view of the possible improvements and how much that improvement is. A segment that is relatively large indicates that there is a lot of room for improvement in that input/output variable. A segment that is very small indicates that there is little room for improvement.

Fig 4: Total potential improvements graph In Fig 5 to Fig 11 we could find that in each case the reduction of the performance metrics like Queuing delay, congestion, jitter, packet loss factor and delay would result in an increase in the QoE factor for the service provider. With the help of these graphs the

service provider (Doctor) could understand the need for better utilization of resources and helps in target setting. The relative efficiency is provided for each of the considered 7 cases.

Fig 5.(P1,D)Efficiency-100%

Fig 6.(P2,D)Efficiency- 60.9%

Fig 7.(P3,D)Efficiency – 100%

Fig 8.(P4,D)Efficiency – 100%

Fig 9.(P5,D)Efficiency – 80.9%

Fig 10.(P6,D)Efficiency – 100%

Fig 11.(P7,D)Efficiency – 85.7% 8. Conclusion A Policy SLA Framework for Effective Telemedicine is modeled using Linear Programming, SLA Trading algorithm and DEA. The performance metrics that are considered crucial such as Bandwidth delay, demand, packet loss, congestion, queuing delay, throughput, buffer capacity are included in this model. These models were simulated and tested using Data Envelopment Analysis (DEA). The results show significant understanding on the inclusion of Policy SLAs using DEA for Effective Telemedicine using Internet. References [1] Available at:http://www. cms.hhs.gov/glossary/default.asp [2] Available at:http:// http://www.coiera.com/glossary.htm [3] Available at:http://www.omge.org/publications/archive/2002_2/sci/sci6.htm [4] Fankhauser .G, Schweikert .D, Plattner .B, “Service Level Agreement Trading for the Differentiated Services Architecture”. Swiss Federal Institute of Technology, Computer Engineering and Networks Lab, Technical Report No. 59. Nov. 1999. [5] Stefan Savage, Tom Anderson, Amit Aggarwal, David Becker, Neal Cardwell, Andy Collins, Eric Hoffman, John Snell, Amin Vahdat, Geoff Voelker, John Zahorjan,1999. [6] Steering Committee for the Review of Commonwealth/ State Service Provision, Data Envelopment Analysis: A technique for measuring the efficiency of government service delivery, AGPS, Canberra. [7] Wahl, M., Howes, T., and S. Kille, “Lightweight Directory Access Protocol (v3)”, RFC 2251, Dec. 1997. [8] Simplex Method available at: URL:http://mathworld.wolfram.com/SimplexMethod. html [9] Time Delay available at http://www.inst.eecs.berkeley.edu/~ee122/disc/disc02_ling. doc [10] Queuing Delay Available at http://encyclopedia.thefreedictionary.com/ Queueing%20delay [11] Saad Biaz, Nitin H.Vaidya (1999), “Discriminating Congestion Losses from Wireless Losses using Inter-Arrival Times at the Receiver”, Proc. IEEE Symposium ASSET’99, USA [12] Throughput Available at: http://www.cw.com/our_network/packet_loss_calculator/ plc_model.html [13] H. Schulzrinne, S. Casner, R. Frederick, V. Jacobson (1996), “RTP: A Transport Protocol for RealTime Applications”, RFC 1889, January. [14] L. Lewis, “Implementing Policy in Enterprise Networks,” IEEE Communications Magazine, vol. 34, no. 1, pp. 50-55, January 1996.

[15] S. Rajeev, S.N. Sivanandam, K. Duraivel, Santosh G. Rao, P. Pradeep “Policy Based Provisioning For Wireless Differentiated Services”, Proc. IEEE 12th Annual Symposium on Mobile Computing and Applications, Bangalore,Nov. 2003 [16] S. Rajeev , S. N. Sivanandam , Mothi V. Sabaresan, B. Anand, “Frequency Allocation and Priority Handling in Multi-Service Wireless Differentiated Networks”, International Journal of System Modeling and Simulation, Vol. 2, pp.26-29, Jan.2004. [17] Appan Ponnappan, Lingjia Yang, Radhakrishna Pillai.R Peter Braun, “A Policy Based QoS Management System for the IntServ/DiffServ Based Internet”, Proc. Third International Workshop on Policies for Distributed Systems and Networks, 2002. [18] Stefan Savage, Tom Anderson, Amit Aggarwal, David Becker, Neal Cardwell, Andy Collins, Eric Hoffman, John Snell, Amin Vahdat, Geoff Voelker, John Zahorjan, “Detour: A Case for Informed Internet Routing and Transport”, IEEE Micro, Vol.19, No 1 Jan. 1999. [19] Banker, R. D., A. Charnes, and W. W. Cooper, 1984, “Some Models for Estimating Technical and Scale Efficiencies in Data Envelopment Analysis,” Management Science, 30, (9), 1078-1092. [20] Charnes, A., W. W. Cooper, and E. L. Rhodes, 1978, “Measuring The Efficiency Of Decision Making Units,” European Journal of Operational Research, 2, (6), 429-444. [21] Nigel Sheridan-Smith, “A Distributed Policy-based Network Management (PBNM) system for Enriched experience Networks (EENs)”, Ph.D, University of Technology, Sydney.