âQuality of Serviceâ metrics of throughput, delay, delay variation, probability of error, ... Once the data is on the network, it is best to schedule the routing information using ..... [3] I. Kleinrock, "On the Modeling and Analysis of Computer Networks.
Intelligent Communication Systems Sensor Scheduling Based on Network Information Erik. P. Blasch and Andrew Hoff Department of Electrical Engineering, Wright State University, Dayton, Ohio
ABSTRACT Increased reliance on data communications has driven the need for a higher capacity solution to transmit data. Many decision making problems include various types of data (i.e. video, text), which require an intelligent way to transfer the data across a communication system. Data fusion can be performed at the transmitting end to reduce the dimensionality of the data transferred; however if there are errors, the receiving end would have no way to search through the pedigree of information to correct the problem. Thus, a desired analysis is to be able to transfer all the data types, while achieving the “Quality of Service” metrics of throughput, delay, delay variation, probability of error, and cost. One way to solve this problem is by using the Asynchronous Transfer Mode network data scheduling. An ATM network allows multiple types of data to be sent over the same system with dynamic bandwidth allocation. The following paper provides a description of an intelligent scheduling model to enhance the capability to transmit data for fusion analysis. Key Words: Communications, Fusion, queuing, Intelligent Systems
1. INTRODUCTION Increased reliance on data communications has driven the need for a higher capacity solution. One way to solve this problem is by using the Asynchronous Transfer Mode (ATM) network. An ATM network allows multiple types of data to be sent over the same communication system assuming bandwidth allocation scheduling. The intelligent design of a communication system is based on sensor scheduling where sensor scheduling is a function of the bandwidth available. Typically, the limited bandwidth is optimized over different routers scheduling the passing of data based on the network congestion. To determine whether or not a system is functioning properly, the user would determine the credibility of the communication system to meet operational needs. The needs are based on the fusion of various sets of information which are called “Quality of Service” metrics (1) probability of error, (2) throughput, (3) delay, (4) delay variation, and (5) cost. Probability of error relates to the packet lost. Throughput is a result of the information that is reliably transmitted from the source to the destination. Delay and delay variation are a function of time. Finally, cost is related to the equipment that is needed to meet the QOS objectives. More equipment means a higher cost. Figure 1 shows the network control strategies that can be used to control congestion and hence maximize throughput.
Figure 1. Network Control Strategies
Many sensor scheduling systems strategies have been designed to optimize the communication criteria. Typical strategies are based on the network congestion control techniques. The implicit techniques are within the network, while the explicit techniques are outside the network. For optimum scheduling of sensor use, one would have to control the data source before it reaches the network. Once the data is on the network, it is best to schedule the routing information using an forward explicit control network (FECN) strategy that looks at all the routers and determines where next to send the data. Using the forward looking strategy is similar to a data pull as opposed to a data push technique. To control the sensors, it is desirable to understand the communication medium of the data type. Using an ATM system, the packet size for the data is fixed. However, different mediums determine the different numbers of packets. If there is a large file size, such as an MPEG video, then more packets would be attributed to the video as opposed to small text sizes. The third detail for intelligent sensor scheduling is priorities. Priorities essentially weight the importance of the data in the communication system. The service rate, based on different classes, are as follows: Constant Bit Rate (CBR) - The CBR service is used by applications that require a fixed data rate that is continuously available during the connection lifetime and a relatively tight upper bound on transfer delay. It can be processed with a small amount of jitter. Real-time Variable Bit Rate (RT-VBR) - is intended for time-sensitive applications, somewhat bursty, that transmit at a rate that varies with time that the standard approach to video compression results in a sequence of image frames of varying sizes. Non-real-time Variable Bit Rate - (NRT-VBR) - - the end system specifies a peak cell rate, a sustainable or average cell rate, and a measure of burstiness. With this information, the network can allocate resources to provide relatively low delay and minimal cell loss. Unspecified Bit Rate (UBR) - All of this unused capacity could be made available for the UBR service. This service is suitable for applications that can tolerate variable delays and some cell losses, which is typically true of Transfer Control Protocol (TCP)-based traffic. Available Bit Rate (ABR) - Specifies a peak cell rate (PCR) that it will use and a minimum cell rate (MCR) that it requires. The network allocates resources so that all ABR applications receive at least their MCR capacity. Any unused capacity is then shared in a fair and controlled fashion among all ABR sources. ABR is local area network (LAN) interconnection. In this case, the end systems attached to the ATM network are routers. To optimize the service rate to meet the quality of service constraints, intelligent computation strategies are needed. Sensor scheduling, as per the connection to the network, are then based on the network control strategy, because if the sensor can not put its data on the network, it can not transmit the information to the destination. The following paper will provide a description of the intelligent strategies to design a communications network. Section 2 briefly discusses intelligent communication and Section 3 details the problem formulation. Section 4 shows the implementation and section 5 shows the results. Section 6 provides a discussion.
2. INTELLIGENT COMMUNICATION The task for intelligent communication system requires designing an ATM network to transfer many different types of data simultaneously, including voice, video, and text for bursty TCP real-time and non-real time services. ATM is a networking scheme designed to support both traditional TCP and User Datagram Protocol (UDP). For mobile communication systems, it is important to allow the user the flexibility to select the service type as well as the send a variety of data sizes with control of the service metrics. To perform an analysis of the sensor scheduling, we need to address the capabilities of the network to provide communication over the various service information. Two techniques that are common are time division multiplexing and statistical division multiplexing. Synchronous time-division multiplexing A method of TDM in which time slots on a shared transmission line are assigned to I/O channels on a fixed, predetermined basis. Time-division multiplexing (TDM) The division of a transmission facility into two or more channels by allotting the facility to several different information channels, one at a time.
For the high priority sensors, we chose to use the SDM, while the other traffic that is integrated on the line would use the TDM model. The integration of information on a single communication network is similar to fusion systems as the quality, timeliness, and accuracy of the data needs to be optimized. Data that arrives late, is not as useful as data that models the most immediate environmental information. For example, if you desire an accurate tracker, you would want timely and accurate updates to the Kalman filter. Delay and delay variation model expected update characteristics. If delay variation is small, the fusion system could anticipate exactly when the data was to arrive; however, if the data has a high delay variation it is unsure when it will arrive and causes speculation. Finally, for reliable communication, the data has to be time-tagged, so that if it arrives out of order, it can be processed and reorder (all of which adds to the delay).
3. PROBLEM FORMULATION The modeling set up includes modeling the sensors data arriving on the network and departing the network to the destination. The network map used is shown in Figure 2. In the network, there are 12 routers, where each is modeled as a M/M/1 queue. The design is based on the fact that most data communication systems have some intermediary processing before the results reaches the destination. R02
R01
R03
R04 R08
Initiation
R12
R07 R10 R05
R06 R09
R11 Destination
Figure 2. Communication Network with multiple routers
The services types were introduced in order of priority of transmission. The system was modeled as a “pipe” of available bandwidth utilized to 80% capacity. CBR, being the highest priority, was first allocated a specific section of bandwidth and then the RT-VBR and so forth. By modeling these services as a pipe, the ATM system simulation can keep these services separate and allocate whatever section of the bandwidth needed to accomplish appropriate transmission. Figure 3 outlines this concept.
CBR Figure 3. Bandwidth Allocation for CBR in ATM “pipe” Given parameters to model the network, we can analyze the delay of data transmission through the network. The percentage of use, variability of delay, peak, average, and minimum cell rate for each type of data, and the burstiness of the traffic was to be studied. Each ATM cell (packet) contains 53 octets--48 information and 5 header with these data sizes o Capacity of the network is 56Kbits / sec o Videoconferencing audio/ video (CBR) – 1 Megabits o Video-compression (RT-VBR) – 200 pictures at 56K a piece o Banking transactions (NRT-VBR) – 10 Megabits
o o o
Home users - Remote terminal (UBR) – 7 Meg (FIFO processing) LAN interconnection – (NRT-ABR) – 25 Meg Buffer size = 1 Meg (later reducing to 0.5 Meg for comparison)
Using these given parameters, the network simulation was initiated and data was analyzed for the performance of the network. The Poisson arrival rates were given as follows: Videoconferencing audio/ video (CBR) 8 Megabits l = 10 kb/sec Video-compression (RT-VBR) 200 pictures at 56K a piece l = 56 kb/sec Banking transactions (NRT-VBR) 10 Megabits l = 10 kb/sec Home users - Remote terminal (UBR) 7 Meg (FIFO processing) l = 05 kb/sec LAN interconnection – (NRT-ABR) 25 Meg l = 100 kb/sec The processing rates for the queues Ts = 0.001 sec / item The specific information about how much data is being sent through the network allows us to do a network analysis. For this paper, the following objectives are to be addressed: Compare the single server case (project 1 - M/M/1) versus multiple servers Address the issue of priority of the queues Measure the performance based on the variance. Find the different buffer sizes needed to hold the data Address cost issues from the server rate and the number of servers (e.g. routers) The overall goal was to address the issue of priority while implementing actual values for the size of each different data type. By choosing QOS design metrics, one could determine the buffer sizes needed. To simulate flow and error control different schemes: we used a Go Back N strategy. We were to analyze the throughput of these different schemes and do a least cost analysis at the link level to pick a fair buffer scheme. The propagation time is equal to the distance d of the link divided by the velocity of propagation V. The transmission time is equal to the length of the frame in bits L divided by the data rate R. Therefore, a=
Propagation time Transmission time
⇒
a=
d/V Rd = L/R VL
(1)
This parameter is very useful in assessing the performance of various link control schemes and provides an insight into the factors affecting performance. The following variables: d = Distance of the link between two stations. V = Velocity of propagation of the signal along the link. For unguided transmission through air or space, V is the speed of light, 3 x 108 m/s. For guided transmission, V is approximately the speed of light for optical fiber and about 0.67 times the speed of light for copper media. L = Length of a link control frame in bits; for now, we assume a constant length. R = Data rate on the link, in bits per second. Thus, for fixed-length frames and a fixed distance between stations, a is proportional to the data rate times the length of the medium. A useful way of looking at a is that it represents the length of the medium in bits [R x (d / V) ] compared to the frame length (L). If the buffer occupancy exceeds the threshold, then the next new incoming packet on is dropped if the window size W(i) exceeds a parameter Z. Experiments show good results for a value of Z of a little less than one. The following conclusions were reached: • Selective drop provides improved fairness. The dropping of a packet forces the corresponding TCP connection to back off and reduce window size. At the same time, ATM buffer resources are freed up and other TCP connections can increase their window size and throughput. Thus, selective drop works with TCP congestion control to balance loads. • Fairness and total throughput increase with increased ATM switch buffer size. • Fairness decreases with increasing number of sources.
Selective drop begin to drop cells when a fixed threshold is reached. Fair buffer allocation (FBA) adopts a policy of more aggressive dropping of packets as congestion increases. The rule for FBA is based on the following condition: B-R (N > R) AND W(i) > Z x N - R
(2)
where R is the data rate, W the window size, and B the queue buffer length The overall normalized throughput is calculated as V
∑ xi
i=1
Throughput = V x M where
(3)
xi = throughput of the ith TCP source V = number of TCP sources (= number of VCs) M = maximum possible TCP throughput The FBA is only slightly better than selective drop. The following as a measure of fairness: V
( ∑ xi)2 Fairness =
i=1 V
V x ∑ (xi)2
(4)
i=1
This is a normalized measure of the dispersion of the values of x,. The generalized process sharing (GPS) is attractive because it allows a router to assign the appropriate weight to each flow to guarantee service and because it is possible to guarantee an upper bound on delay. These qualities carry over to the weighted fair queuing (WFQ) approximation to GPS. In this case, Bi (Ki - 1) Li Ki Lmax Di ≤ R + ∑ Cm Ri i m=1 where
(5)
Di = maximum delay experienced by flow Bi = token bucket size for flow i Ri = token rate for flow i Ki = number of nodes in the path flow i through the internet Li = maximum packet size for flow i Lmax = maximum packet length for all flows through all nodes on the path of flow i Cm = outgoing link capacity at node cit
The first term carries over from the GPS case and accounts for delay due to bucket size, which is the same as delay due to burstiness. The second term is proportional to the delay experienced at each node for each packet by this flow. The final term reflects the consequence of packet-by-packet rather than bit-by-bit transmission. In WFQ, a packet may leave some time later than it would have under bit-by-bit processing. The reason is that the node can only choose for transmission among all the packets that have already arrived. If the next packet chosen is relatively large and a small packet arrives during this transmission, it may be that under GPS, this small packet has the earliest finish time and therefore should have been transmitted first under WFQ. Because it did not arrive, it was not transmitted and has to be delayed up to at most the full length of the longest packet that moves through this node.
4. RESULTS The following plots are output from the simulation. The following figures (1-10) are indicative of a particular simulation in which juncture routers/nodes have 5 servers while the routers/nodes in a path have only one. The implementation was conducted in MATLAB.
Figure 4. Results of the Service types for the M/M/1 Queue
Figure 5. Results of the Cell Rate and Jitter for the M/M/1 Queue
Figure 6 shows the amount of data (bits) in each buffer at each time step for the entire network. Note how routers 2 and 5 never reach their maximum buffer size, which in this case is 1 Meg. From the network design, we would expect that the congestion is maximum at 2 and 5 since they are responsible for distributing the data to the rest of the network where router 5 has the maximum control decision logic and needs a larger buffer. Router 1 could also be a case for congestion, but it was the start of the simulation and all data wash pushed onto the router.
Figure 6. Buffer Activity Also plotted is the total data resident in the network for the entire simulation, shown in Figure 7. Figure 8 shows the peak, minimum and average cell rate for each service type.
Figure 7. Total Bites in Network as a Function of Time
Figure 8. Cell Rates: Peak, Minimum, Average
Lastly, the parameter “a” is plotted for the network in Figure 9. This value is different for each connection between the 12 routers (16 total paths). The parameter “a” affects the throughput and with the probability of error, determines the window size calculations. [See Stallings for further details].
Figure 9. The Parameter “a”
Figure 10: Throughput after Buffer Size is reduced to 500 k (vice 1 Meg). As expected, there was an increase in the variance through router 12. Unexpectedly, routers 5 and 2 were different. They never did fill up their buffer, no matter what are simulation variables were. It makes sense that router 2 would be less likely than the others to fill up, since it has two outputs and only one input. Given that router 5 is larger in peak capacity, router 5 would be the last router to reach saturation with an effective control strategy. Also implemented into the simulation, is the calculation of least cost path calculations (Djisktra), probability, entropy, as well as a and window size, W, (the later two for each time, t). The values for probability and entropy are displayed in the same manner as the adjacency matrix. These calculations are echoed to the Matlab workspace. The values at each time, t, for a and W are assessed, but not shown.
5. DISCUSSION The purpose of comparing the sensor control scheduling is dependent on meeting the various metrics: maximize throughput, minimize delay and packet loss. All can not be achieved separately. Implementation of the scheduling based on the criteria selected and further efforts include adapting the queuing simulation to include different types of queues. As given, it is assumed that each node is capable of transmitting 56 kbps. We assumed 80% utilization, which equates to 44.8 kbps. Buffer status is checked in a backward direction, that is the last buffer sends buffer status back to the preceding buffers/switches. Ts for the server, is given to be 1 msec per item, which was converted to time per bit. It indicates how fast it can send the data to the next queue/buffer. The source is shoving 181 kbps (sum of the arrival rates) into a network with a capacity of 56 kbps. As a result, we get a backlog in the network and the network has to use some logic to process and deliver the information through the network to the end user. Different values might have different results. This simulation was extremely complex. This type of simulation has rarely been implemented purely in software. Based on our analysis of some websites, none of these implementations have been successful in purely software implementations. Many have given up and not completed their goals. The NIST implementation appears to be a possible successful implementation which we could not verify.
6. CONCLUSIONS Overall, the network scheduling shows promise for sensor management control strategies for fusion systems. Without queuing, the network was able to throughput the data in a little over 1000 seconds (we assumed 80% utilization). Once queuing was implemented, this number was slightly more. This slight difference shows how our 80% estimation was fairly accurate and also gives evidence that that data we are handling is not self-similar (which we know that it isn’t). There was a difference in the time of calculation depending on the number of servers implemented. The one server case for each node was the slowest. The best case that we could find was to put 5 servers at the decision points and one server at the pass-through points. Surprisingly, it was faster to but single server in the nodes that were in the paths. We assume that putting more than one server simply increased the decision time for the router and overall had the effect of reducing the throughput of that router. Future explorations of queuing analysis will be developed to optimize the fusion scheduling analysis.
7. REFERENCES
[1] W. Stalling, High-Speed Networks and Internets, Performance and Quality of Service, 2nd Edition, 2001 Prentice Hall, Upper Saddle River New Jersey 07458 [2] I. Kleinrock, " The latency / Bandwidth tradeoffs in Gigabit Networks." IEEE Communications Magazine, April 1992. [3] I. Kleinrock, "On the Modeling and Analysis of Computer Networks." Proceedings of the IEEE, August 1993.