Transmission Window Optimization for Caching-Based Transport ...

5 downloads 106 Views 423KB Size Report
May 21, 2015 - Transmission Window Optimization for Caching-Based Transport Protocols in Wireless Sensor Networks. Nestor Michael C. ... Reference tools ...
Transmission Window Optimization for Caching-Based Transport Protocols in Wireless Sensor Networks Nestor Michael C. Tiglao1(B) and Ant´ onio M. Grilo2 1

2

Ubiquitous Computing Laboratory EEE Institute, University of the Philippines Diliman, 1101 Quezon City, Philippines [email protected] INESC-ID/INOV/IST Rua Alves Redol, N◦ 9, 1000-029 Lisboa, Portugal [email protected]

Abstract. Traditional transport protocols have been designed to perform end-to-end transmission and retransmission. In terms of choosing the optimal transmission window, previous works suggest a value based on the bandwidth-delay product (BDP). For wireless networks, the BDP value is related to the round-trip hop length. However, there exists a new class of transport protocols that use intermediate caching which can drastically improve the performance even in the presence of high packet error rates. In this paper, we show that using a window size related to the BDP could lead to sub-optimal performance for caching-based protocols. Furthermore, we present a heuristic for choosing the optimal transmission window such that the optimal value is related to the average cache size in the intermediate nodes. Keywords: Wireless sensor networks Intermediate caching · Optimization

1

· DTSN · Transmission window ·

Introduction

Previous studies have shown that the transmission capacity of wireless ad hoc networks is related to the network size. Specifically, in [1] it was shown that the ideal capacity of a long chain of nodes is 1/4 of the raw channel capacity of the radio. In another work [2], the authors have shown that an upper bound on the bandwidthdelay product (BDP-UB) exists and is a function of the number of hops in the path. Furthermore, their study indicates that applying this BDP value to TCP’s congestion window limit effectively improves the performance of TCP. Datagram Transport Protocol for Ad Hoc Networks (DTPA), a reliable protocol designed for ad hoc networks, uses a fixed transmission window based on the bandwidth delay product for wireless networks (BDP-UB). However, such choice could lead to sub-optimal performance when applied to a caching-based protocols since intermediate caching improves packet delivery in the presence of packet loss compared with non-caching protocols [3]. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2015  S. Mumtaz et al. (Eds.): WICON 2014, LNICST 146, pp. 39–46, 2015. DOI: 10.1007/978-3-319-18802-7 6

40

N.M.C. Tiglao and A.M. Grilo

Reliable transport protocols require packets transmitted from the source to be acknowledged by the destination. Unacknowledged packets are deemed lost and are retransmitted by the source. The source may transmit multiple packets into the network before expecting any acknowledgment depending on its transmission window size. This value also corresponds to the maximum number of in-flight packets. The window size can be set dynamically (e.g., TCP [7]) or set to a fixed value (e.g., DTPA [5]). The window size must be optimized properly because it directly affects the end-to-end throughput, such that setting it too high can lead to congestion while setting it too low can result in low throughput. With TCP, the transmission window is controlled by the congestion control algorithm (i.e., AIMD). End-to-end performance depends not upon the transfer rate itself, but rather upon the product of the transfer rate and the round-trip delay. This bandwidth-delay product (BDP) measures the amount of data that would fill the network pipe. This BDP also dictates the buffer space required at sender and receiver to obtain maximum throughput on the TCP connection over the path. The traditional TCP has a maximum transmission window of 64 K bytes. For wired networks, this value works well in medium-sized pipes. For long fat pipes, the BDP exceeds this value and the 64 KB limit in the TCP header needs to be increased. However, for small-sized pipes where the BDP is well below 64 KB, such as in wireless networks, the BDP normally comprises a few packets only. Distributed Transport for Sensor Networks (DTSN) [6] belongs to a class of transport protocols that leverage intermediate caching. DTSN supports both full and differentiated reliability and employs selective repeat ARQ using ACK and NACK semantics. When the DTSN receiver detects a lost packet, it creates a NACK packet indicating missing packet numbers in the current DTSN window. As the NACK traverses the network, the intermediate nodes will examine the NACK to see if there are any copy of missing packets in their cache. If copy is found, that packet is sent to the DTSN receiver and the NACK packet is modified (removing that packet number) and forwarded on towards the DTSN sender. In this way, end-to-end retransmission is reduced. Paper contributions: This paper makes the following contributions: (1) We show that using a transmission window value based on the bandwidth-delay product leads to sub-optimal performance of caching-based transport protocols and (2) we develop a heuristic for choosing the optimal value for DTSN. To the best of our knowledge, this work is the first to consider the problem of choosing the optimal transmission window for caching-based protocols such as DTSN. The rest of the paper is structured as follows. Section 2 presents the related work. Section 3 explains the simulation environment and protocol parameters used in this work. Section 4 presents our results. Finally, Sect. 5 concludes the paper.

2

Related Work

The work of [2] has established that the upper bound on the BDP for 802.11based MANETs cannot exceed kN , where N is the round-trip hop length and

Transmission Window Optimization

41

1/8 < k < 1/4 is the reduction factor due to transmission interference at the MAC layer. A larger k would suggest that the interference is smaller and that the chain network can accommodate for in-flight packets. In the said work, the authors obtained the value of k empirically to be equal to 1/5. DTPA [5] proposes a fixed optimal transmission window size equal to BDP-UB + 3. In our study, we have established that using a BDP-based value leads to sub-optimal performance and that the optimal transmission window size is equal to the cache size allocated to the flow. While this is most evident in high packet error rates, this is also true even in lower packet error rates where packet collisions can occur. Table 1 summarizes the protocols we considered in this thesis. For the 9-hop linear chain topology we considered, the value of the transmission window were set to 3 for DTPA-BDP and 6 for DTPA. Table 1. Approaches to optimal transmission window. Approach

Protocol used Caching Strategy

Chen, et al. [4] TCP

N

Set congestion window limit to BDP-UB where BDP-UB=1/5 * round-trip hop length

Li, et al. [5]

DTPA

N

Set fixed window size to BDP-UB+3

Our Work

DTSN

Y

Set fixed window size to the cache size assigned to the flow

3

Simulation Environment

We implemented the DTSN protocol in ns-2 [8] and conducted extensive simulations. We consider a linear network topology consisting of 10 nodes with a single source (node 0) and destination (node 9). All the intermediate nodes have the same cache size. The source sends 500 packets of 500 bytes each in a realible stream-type transfer similar to TCP. We consider a network scenario considering uniform Frame Error Rate (FER) and we vary the FER from 0 to 0.70. The MAC retry limit is set to 3 (default value) unless specified otherwise. The DTSN EAR interval is set to 200 msec which is equal to the default minimum RTO setting in ns-2. In order to conduct a comparison and analysis of the effect of our mechanisms on protocol performance, we fix the FER and network topology for each case. For each experiment, we conducted 20 simulation runs and obtained the 95 % confidence intervals. Table 2 provides a summary of the simulation parameters.

4

Results

We performed a comparative analysis of the following protocols to determine the effect of the transmission window:

42

N.M.C. Tiglao and A.M. Grilo Table 2. Simulation parameters Parameter

Value

Network topology

Linear chain

Packet size

500 bytes

DTSN ACK window size

20 packets

DTSN cache size

20 packets

DTSN EAR interval

1 sec

Routing protocol

Static

MAC protocol

802.11 b

MAC retry limit (default) 3 PHY bandwidth

50 Kbps

PHY error model

Binary Symmetric Channel

– DTPA – The DTPA protocol as described in [5]. – DTPA-BDP – The DTPA protocol with the transmission window set to the value of the wireless bandwidth-delay product (BDP) [2]. – DTSN+ – The DTSN protocol with the enhanced NACK repair and adaptive MAC retry limit mechanisms [9]. – TCP− – The TCP protocol without the RTO exponential backoff. The goodput performance results (Fig. 1) show that the transmission window has a significant impact on the overall performance. Note that DTPA and DTPABDP only differs in the size of the transmission window (6 and 2, respectively). Both of these protocols use a fixed window size while TCP uses a dynamic window. However, all them perform end-to-end loss recovery. On the other hand, DTSN leverages on intermediate caching. Furthermore, we studied the effect of the transmission window on DTSN by varying the acknowledgment window (AW) size. In order to simplify our network 140 DTPA−BDP DTPA − TCP + DTSN

Goodput (in packets/sec)

120 100 80 60 40 20 0 0

0.10

0.30 Frame Error Rate

0.50

0.70

Fig. 1. Goodput performance comparison

Transmission Window Optimization

43

configuration, we assume that all intermediate nodes have the same cache size (CS) and we vary the AW. We obtained both goodput and transmission cost as performance metrics. The transmission cost is the average number of link-wise packet transmissions including control and MAC layer packets and computed as follows: Ndata + Nack + Nnack + Nmack (1) tx cost = pktno where Ndata is the total number of data packets transmitted, Nack is the total number of transport-layer ACKs, Nnack is the total number of transport-layer NACKs, Nmack is the total number of MAC-layer ACKs, and pktno is the total number of packets that need to be delivered end-to-end. Figure 2 compares the goodput for CS=10 and CS=20. For CS=10, the optimal (i.e., maximum) goodput is achieved at AWopt =[10,20] while for CS=20 optimal goodput is achieved at AWopt =[20,30] as shown in Figs. 2(a) and (b), respectively. In terms of energy efficiency, we see that the transmission cost is minimized in the same corresponding range of AWs as shown in Figs. 3(a) and (b). It can be seen that each figure shows a minimum value that corresponds to the optimal value of the transmission window. Figure 3 shows that for high framer error rates (i.e., FER ≥ 0.5), the transmission cost at AW=5 is higher than at AW=1. This can be explained by the fact that at such high error rates, the cache hits are so low resulting in fewer RNACKs. Another factor is that with only one in-flight packet, contention is greatly reduced. To quantify the optimality of the transmission window size, we calculate the overall gain of the goodput and transmission cost at a given AW relative to the worst case value (i.e., minimum goodput or maximum cost, respectively) using the following equations: GainGoodput =

FER=0.1

FER=0.3

FER=0.5

FER=0.7

FER=0

140

140

120

120 Goodput (in packets/sec)

Goodput (in packets/sec)

FER=0

GoodputAW − min(Goodput) min(Goodput)

100 80 60 40 20 0 1

FER=0.1

FER=0.3

(2)

FER=0.5

FER=0.7

100 80 60 40 20

5 10 20 30 40 Acknowledgment Window (AW) (in packets)

(a) CS=10

50

0 1

5 10 20 30 40 Acknowledgment Window (AW) (in packets)

(b) CS=20

Fig. 2. Goodput, as a function of AW

50

44

N.M.C. Tiglao and A.M. Grilo FER=0.1

FER=0.3

FER=0.5

FER=0.7

FER=0 200

150

150 Transmission Cost

Transmission Cost

FER=0 200

100

50

0 1

FER=0.1

FER=0.3

FER=0.5

FER=0.7

100

50

5

10

20

30

40

0 1

50

5

Acknowledgment Window (AW) (in packets)

10

20

30

40

50

Acknowledgment Window (AW) (in packets)

(a) CS=10

(b) CS=20

Fig. 3. Transmission cost, as a function of AW

where GainGoodput = goodput gain, GoodputAW = goodput at a specific AW, min(Goodput) = overall minimum goodput. GainCost =

max(Cost) − CostAW CostAW

(3)

where GainCost = transmission cost gain, CostAW = transmission cost at a specific AW, max(Cost) = overall maximum transmission cost. Figure 4 shows the gain for CS=10. The goodput gain for AW=10 is lower than for AW=20 but the transmission cost gain of the former is higher than the latter. This can be explained by the fact the energy efficiency is achieved when the intermediate caching mechanism is maximized (i.e., maximum cache hits). 0.35

2 1.8

AW=10 AW=20

0.3

AW=10 AW=20

0.25

1.4

Performance Gain

Performance Gain

1.6

1.2 1 0.8 0.6

0.2 0.15 0.1

0.4 0.05 0.2 0

0

0.1

0.3 FER

0.5

(a) Goodput

0.7

0

0

0.1

0.3 FER

0.5

(b) Transmission Cost

Fig. 4. Performance gain histograms, CS=10.

0.7

Transmission Window Optimization 3

0.35

AW=20 AW=30

0.3

1.5

1

0.2 0.15 0.1

0.5

0

AW=20 AW=30

0.25

2

Performance Gain

Performance Gain

2.5

45

0.05

0

0.1

0.3 FER

0.5

0.7

(a) Goodput

0

0

0.1

0.3 FER

0.5

0.7

(b) Transmission Cost

Fig. 5. Performance gain histograms, CS=20.

This occurs when the transmission window is equal to the cache size. However, in terms of goodput, a higher number of in-flight packets will tend to increase goodput since more packets can be delivered within the same period. This same pattern is observed for CS=20 as displayed in Fig. 5. In our study, we assume uniform allocation for cache size on all nodes in the path. We have seen that for a given cache size (CS), the optimal energy efficiency (minimum transmission cost) is obtained when the transmission window (AW ) when AW = CS while the optimal goodput (maximum goodput) is obtained when AW = CS + 10. Thus, in general, the optimal transmission window is in the range AWopt = [CS, CS + 10].

5

Conclusion and Future Work

Recent works have highlighted the tremendous benefit of caching-based transport protocols in improving network performance in the presence of high packet error rates such as in Wireless Sensor Network. These necessitates a study in the optimal transmission window for these new class of protocols. Basing it on wireless bandwidth-delay product values leads to sub-optimal performance. In this paper, we have determined a heuristic in choosing the optimal transmission window for DTSN which is a function of the cache size in the intermediate nodes. Our future work shall include testing this heuristic on more complex and dynamic network scenarios as well implementing a suitable congestion control mechanism.

References 1. Li, J., Blake, C., De Couto, D.S., Lee, H.I., Morris, R.: Capacity of Ad hoc wireless networks. In: Proceedings of the 7th Annual International Conference on Mobile Computing and Networking, MobiCom 2001, pp. 61–69. ACM, New York (2001)

46

N.M.C. Tiglao and A.M. Grilo

2. Chen, K., Xue, Y., Shah, S.H., Nahrstedt, K.: Understanding bandwidth-delay product in mobile ad hoc networks. Comput. Commun. 27(10), 923–934 (2004) 3. Tiglao, N.M.C., Grilo, A.M.: An analytical model for transport layer caching in wireless sensor networks. Perform. Eval. 69(5), 227–245 (2012). doi:10.1016/j.peva. 2011.12.002. ISSN 0166–5316 4. Chen, K., Xue, Y., Nahrstedt, K.: On setting TCP’s congestion window limit in mobile ad hoc networks. In: IEEE International Conference on Communications 2003, vol. 2, pp. 1080–1084 (2003) 5. Li, X., Kong, P.-Y., Chua, K.-C.: DTPA: A reliable datagram transport protocol over ad hoc networks. IEEE Trans. Mob. Comput. 7(10), 1285–1294 (2008) 6. Marchi, B., Grilo, A., Nunes, M.: DTSN: Distributed transport for sensor networks. In: 12th IEEE Symposium on Computers and Communications. ISCC 2007, pp. 165–172 (2007) 7. Postel, J.,: RFC 793. Transmission control protocol. RFC 793 (Internet Standard), 1981. Updated by RFCs 1122, 3168, 6093, 6528, IETF. http://www.ietf.org/rfc/ rfc793.txt 8. McCanne, S., Floyd, S.: The network simulator, 2012. http://www.isi.edu/nsnam/ ns/ 9. Tiglao, N.M.C., Grilo, A.M.: Caching based transport optimization for wireless multimedia sensor networks. IJARAS 5(1), 30–48 (2014). Web. 15 Aug. 2014

Suggest Documents