TCP. Evaluation of TCP Throughput Estimation Method for Cellular Networks. Takayuki Goto, 1 Atsushi Tagami, 1 Teruyuki Hasegawa 1 and Shigehiro Ano 1

Transcription

1 TCP TCP TCP TCP CDMA2000 1xEV-DO TCP Evaluation of TCP Throughput Estimation Method for Cellular Networks Takayuki Goto, 1 Atsushi Tagami, 1 Teruyuki Hasegawa 1 and Shigehiro Ano 1 It is important for carriers to grasp and improve communication quality in cellular networks since the amount of data passing through the networks has been increasing. In cellular networks, it is necessary to grasp communication quality extensively. Hence, it is required to develop a technique that can grasp communication quality efficiently. Therefore, we proposed a lightweight method to estimate TCP throughput which is one of the major communication quality metrics in the past. This method first generates an equation for estimation using TCP throughput measurements and delay distributions which are obtained from lightweight probing. Then, the method estimates TCP throughput using delay distributions and the generated equation. In this paper, we evaluate that our proposed method can reduce measurement load and improve estimation accuracy of TCP throughput compared with conventional methods by the experiment in a commercial CDMA2000 1xEV-DO network. 1. 1) TCP TCP 2) TCP TCP TCP TCP TCP TCP TCP 3) 6) Mirza 7) TCP CDMA2000 1xEV-DO TCP TCP 1 KDDI KDDI R&D Laboratories, Inc. 1 c 2010 Information Processing Society of Japan

2 TCP TCP TCP ˆB F TCP TCP 2. TCP 3. TCP TCP 2.1 TCP TCP FTP Iperf 8) Netperf 9) TCP BTC (Bulk Transfer Capacity) cap 10) TCP ),12) TCP Mathis 11) TCP RTTPadhye 12) ACK TCP TCP TCP 4) TCP TCP TCP 3) 7) TCP TCP TCP TCP Swany 3) He 4) TCP Kharat 5) Shah 6) TCP RTT Mirza 7) TCP TCP TCP PathPerf TCP TCP TCP 3. TCP 3.1 TCP CDMA2000 1xEV-DO 13) EV-DO wired network radio access network learning estimation TCP 2 c 2010 Information Processing Society of Japan

3 Fig. 1 1 Framework of proposed estimation method TCP 3.2 TCP TCP RTT / TCP TCP EV-DO 1% Hybrid ARQ DRC (Data Rate Control) EV-DO Rel [kbps] [kbps] 13) DRC TCP TCP TCP DRC DRC ˆB F TCP Throughput = f( ˆB, F ) (1) 3.3 ˆB F N L i [bytes] (i = 1, 2,, N) T [ms] 1 1 P 3.4 ˆB DRC ˆB 14) ) M H M 14) M ( ) L H M (L) = + l i + q i B i i=1 L B i l i q i i q i( 0) (2) 3 c 2010 Information Processing Society of Japan

4 Fig. 2 1 min + ˆ { D( L) } = L b B 2 Principle of bandwidth estimation by one-packet method M ( ) L min{h M (L)} = + l i (3) B i i=1 min{h M (L)} min{h M 1 (L)} = L + l M (4) B M H M (L) H M (L) 4 L j H M (L j ) H M 1 (L j ) EV-DO pathchar 14) EV-DO 100[Mbps] Rel.0 2.4[Mbps] j = r B j B r 3 L min{d(l)} = l i + L + E B i B r i i D(L) L M D(L) = H M (L) min{d(l)} D(L) B r L E min{d(l)} L 2 B r ˆB L i (i = 1, 2,.., N) min{d(l i )} N j=1 ˆB = (L j L) 2 N {(Lj L) (min{d(lj)} min{d})} j=1 L min{d} min{d(l j )} 3.5 F F F 2 3 q i F P P F (L) = P D j=1 j(l) min{d(l)} F L 1) f SVR (Support Vector Regression) y = f(x) = w T x + b (5) y TCP x ˆB F (L) w, b Vapnik ϵ-svr 16) C 1/σ 2 C 1/σ 2 C 2 5, 2 4,..., 2 19, /σ , 2 14,..., 2 2, OS Linux kernel CPU 2.13[GHz] 1[GB] 4 c 2010 Information Processing Society of Japan

5 Fig. 3 3 System structure of experiment OS Windows XP EV-DO PC 17) CPU NTP Windows XP SNTP ( 1 ) TCP ( 2 ) ( 3 ) TCP 3.3 TCP 2[MB] ESN SSN TCPthroughput ET ST SSN ESN ST ET TCP EV-DO TCP 1 3 TCP TCP N=5 L 1=50, L 2=400, L 3=800, L 4=1200, L 5=1400, P =90 RLP Radio Link Protocol RLP T =500[ms] [kbps] 225 [sec] DRC 38.4[kbps] 1/ CC (Correlation Coefficient) ( 1 ) ( 2 ) 3.6 ( 3 ) 2 TCP ( 4 ) TCP TCP CC ( 5 ) CC i CC = (x i x)(y i y) i (xi x)2 (yi y)2 i x i x y i y TCP TCP TCP TCP CC CC TCP TCP TCP ˆB (BE) ˆB F (BE+F) 1000 CC CDF 4 BE BE+F BE CC BE+F TCP BE BE+F TCP DRC 5 c 2010 Information Processing Society of Japan

6 CDF BE+F BE PathPerf-1 PathPerf Fig. 4 Correlation coefficient 4 CC CDF Comparison to conventional methods by CDF of CC 5 BE Fig. 5 Estimation result by BE F 6 BE+F Fig. 6 Estimation result by BE+F PathPerf 7) PathPerf PathPerf 864[kbps] 30[sec] 3.24[MB] TCP 3.24[MB] PathPerf [kB] PathPerf-2 CC CDF 4 BE+F PathPerf-1 PathPerf-1 BE+F PathPerf-2 BE+F PathPerf PathPerf-1 PathPerf-2 CC PathPerf F ˆB BE ˆB F BE+F 5 6 ideal = BE BE+F (N = 5) 90 (P = 90) (L 1, L 2 ) = (50, 400) (L 1, L 2 ) = (1200, 1400) (L 1, L 2) = (50, 1400) 3 CC CDF c 2010 Information Processing Society of Japan

7 types(50,1400) 2 types(50,400) 2 types(1200,1400) 5 types packets per size 20 packets per size 50 packets per size 90 packets per size CDF CDF Correlation coefficient Correlation coefficient 7 8 Fig. 7 Comparison in terms of various packet sizes Fig. 8 Comparison in terms of number of packets per size CC Table 1 Average CC of 2 types and 5 types Packet size Method Average CC 2 types(50,1400) BE BE+F types(50,400) BE BE+F types(1200,1400) BE BE+F types BE BE+F BE BE+F CC 1 BE BE BE+F (P = 10, 20, 50, 90) 90 n = 1, 11, 21 n P P = 10, 20, 50, P CC CDF 8 2 (L 1, L 2) = (50, 1400) 5 (P = 10, 20, 50, 90) 9 CC [kB] [kB] 6 7 c 2010 Information Processing Society of Japan

8 Correlation coefficient Fig. 9 2 types(50,1400) 5 types Amount of probe packets [kb] 9 Comparison in terms of the amount of probe packets TCP TCP TCP TCP TCP DRC EV-DO 6 WiMAX 1) Goto, T., Tagami, A., Hasegawa, T. and Ano, S.: TCP Throughput Estimation by Lightweight Variable Packet Size Probing in CDMA2000 1xEV-DO Network, International Symposium on Applications and the Internet, pp.1 8 (2009). 2) Prasad, R. S., Murray, M., Dovrolis, C. and Claffy, K.: Bandwidth Estimation: Metrics, Measurement Techniques, and Tools, IEEE Network, Vol. 17, pp (2003). 3) Swany, M. and Wolski, R.: Multivariate Resource Performance Forecasting in the Network Weather Service, ACM/IEEE conference on Supercomputing, pp (2002). 4) He, Q., Dovrolis, C. and Ammer, M.: On the predictability of large transfer TCP throughput, Computer Networks, Vol.51, No.14, pp (2007). 5) Khayat, I.E., Geurts, P. and Leduc, G.: Machine-larnt versus analytical models of TCP throughput, Computer Networks, Vol.51, No.10, pp (2007). 6) Shah, S. M.H., Rehman, A., Khan, A.N. and Shah, M.A.: TCP throughput estimation: A new neural networks models, Emerging Technologies, ICET 2007, pp (2007). 7) Mirza, M., Sommers, J., Barford, P. and Zhu, X.: A Machine Learning Approach to TCP Throughput Prediction, SIGMETRICS, pp (2007). 8) NLANR/DAST: Iperf. 9) Jones, R.: Netperf. 10) Allman, M.: Measuring End-to-End Bulk Transfer Capacity, ACM SIGCOMM Internet Measurement Workshop, pp (2001). 11) Mathis, M., Semke, J., Mahdavi, J. and Ott, T.: The Macroscopic Behavior of the TCP Congestion Avoidance Algorithm, Computer Communication Review, Vol.27, No.3, pp (1997). 12) Padhye, J., Firoiu, V., Towsley, D.F. and Kurose, J.F.: Modeling TCP Reno Performance: A Simple Model and Its Empirical Validation, IEEE/ACM Transactions on Networking, Vol.8, pp (2000). 13) 3GPP2: CDMA2000 High Rate Packet Data Air Interface Specification C.0024-a v3.0, 3GPP2 (2006). 14) Jacobson, V.: pathchar a tool to infer characteristics of Internet paths (1997). 15) Carter, R. L. and Crovella, M. E.: Measuring Bottleneck Link Speed in Packet- Switched Networks, Performance Evaluation, Vol.27/28, No.4, pp (1996). 16) Vapnik, V.: Statistical Learning Theory, John Wiley (1998). 17) Pasztor, A. and Veitch, D.: PC Based Precision Timing Without GPS, SIGMET- RICS, pp.1 10 (2002). 8 c 2010 Information Processing Society of Japan