An Investigation into Transport Protocols and Data Transport ...

More documents

Recommendations

Info

9.3. Internet Transfers 232 1000 900 800 Goodput (mbit/sec) 700 600 500 400 300 200 100 0 Standard TCP HSTCP ScalableTCP HTCP FAST BicTCP New−TCP Figure 9.27: Goodput from CERN to Stanford. that results could be gathered. This was achieved by setting the initial value of ssthresh upon a TCP connection to 2 segments rather than max(INT). The effect of disabling slow-start would be similar to the effects of large bursts of transient cross traffic which would similarly cause slow-start to exit early. 9.3.2 Results: CERN to Stanford Goodput Figure 9.27 shows the goodput performance distributions of the various New- TCP algorithms from CERN to Stanford, California, U.S.. It clearly shows the consistent inability of Standard TCP to achieve reasonable goodput over such long distance paths. For the other New-TCP algorithms, the results show a large range of measured goodput performances with symmetric distribution of goodputs. Most notably all algorithms have a minimum which is near zero, suggesting that that the network load can be high, thus preventing high performance network transfers. Whilst it is clear that FAST is able to achieve the highest median goodput performance, it also has the widest distribution. The high maximum also
9.3. Internet Transfers 233 10 1 Stability 10 0 Standard TCP HSTCP ScalableTCP HTCP FAST BicTCP New−TCP Figure 9.28: Stability from CERN to Stanford. suggests that FAST is very capable of reaching higher speeds. Out of all the New-TCP algorithms, HSTCP has the lowest median goodput, and generally was unable to achieve goodputs much greater than 200Mbit/sec, with typical results around 100Mbit/sec (which is still about twice that of Standard TCP). ScalableTCP and BicTCP, although they performed very well in artificial tests, do not achieve as high goodputs across this real-life environment. They both provide a similar distribution, with ScalableTCP having a slightly higher deviation towards higher goodputs. H-TCP exhibited goodput performance higher than ScalableTCP, but not much more in terms of absolute maximum range. Stability Figure 9.28 shows the stability (See Section 7.1.2) distributions of the various New-TCP algorithms to Stanford for the goodput performance results shown in Figure 9.27. The large number of outlier points suggests that all New-TCP algorithms have problems maintaining a low variance and high goodput. Most notably,
Page 1 and 2:
An Investigation into Transport Pro
Page 3 and 4:
Contents List of Figures . . . . .
Page 5 and 6:
Contents 5 4.6 Summary . . . . . .
Page 7 and 8:
Contents 7 8.2 Test Calibration . .
Page 9 and 10:
Contents 9 10.3.5 Further Tests . .
Page 11 and 12:
List of Figures 11 5.7 Effect on cw
Page 13 and 14:
List of Figures 13 8.16 Unfairness
Page 15 and 16:
List of Figures 15 8.32 Friendlines
Page 17 and 18:
List of Figures 17 9.19 Dynamic swi
Page 19 and 20:
List of Figures 19 D.1 Internet pat
Page 21 and 22:
List of Tables 21 A.1 Hardware and
Page 23 and 24:
Publications And Workshops Some of
Page 25 and 26:
1.2. Research Scope 25 • Identifi
Page 27 and 28:
1.4. Dissertation Outline 27 • Id
Page 29 and 30:
Chapter 2 High Energy Particle Phys
Page 31 and 32:
2.2. Analysis Methods 31 the data i
Page 33 and 34:
2.3. Data Volumes and Regional Cent
Page 35 and 36:
2.3. Data Volumes and Regional Cent
Page 37 and 38:
2.4. Data Transfer Requirements 37
Page 39 and 40:
2.5. Summary 39 data, the most impo
Page 41 and 42:
3.1. Data Intercommunication 41 mov
Page 43 and 44:
3.2. Network Monitoring 43 Whilst l
Page 45 and 46:
3.2. Network Monitoring 45 Internet
Page 47 and 48:
3.2. Network Monitoring 47 1 0.8 Bo
Page 49 and 50:
Chapter 4 Transmission Control Prot
Page 51 and 52:
4.1. Overview 51 Throughput Delay L
Page 53 and 54:
4.2. Protocol Description 53 is dep
Page 55 and 56:
4.2. Protocol Description 55 As a r
Page 57 and 58:
4.3. Congestion Control 57 excessiv
Page 59 and 60:
4.3. Congestion Control 59 probe fo
Page 61 and 62:
4.4. Retransmission Timeout 61 tion
Page 63 and 64:
4.5. TCP Variants 63 mechanisms are
Page 65 and 66:
4.5. TCP Variants 65 connection ini
Page 67 and 68:
4.5. TCP Variants 67 20 15 cwnd sst
Page 69 and 70:
4.5. TCP Variants 69 no more acks t
Page 71 and 72:
4.5. TCP Variants 71 [Hoe96]. In bo
Page 73 and 74:
4.5. TCP Variants 73 TCP Timestamps
Page 75 and 76:
4.6. Summary 75 traffic for a parti
Page 77 and 78:
Chapter 5 TCP For High Performance
Page 79 and 80:
5.1. TCP Hardware Requirements 79 G
Page 81 and 82:
5.1. TCP Hardware Requirements 81 1
Page 83 and 84:
5.1. TCP Hardware Requirements 83 C
Page 85 and 86:
5.2. TCP Tuning & Performance Impro
Page 87 and 88:
Page 89 and 90:
Page 91 and 92:
Page 93 and 94:
Page 95 and 96:
Page 97 and 98:
5.3. Network Aid in Congestion Dete
Page 99 and 100:
5.3. Network Aid in Congestion Dete
Page 101 and 102:
5.4. Analysis of AIMD Congestion Co
Page 103 and 104:
Page 105 and 106:
Page 107 and 108:
Page 109 and 110:
5.5. Summary 109 5.5 Summary Even t
Page 111 and 112:
6.1. Survey of New-TCP Algorithms 1
Page 113 and 114:
Page 115 and 116:
Page 117 and 118:
Page 119 and 120:
Page 121 and 122:
Page 123 and 124:
Page 125 and 126:
6.2. Discussion and Deployment Cons
Page 127 and 128:
Page 129 and 130:
Page 131 and 132:
Page 133 and 134:
6.3. Summary 133 6.3 Summary The re
Page 135 and 136:
Chapter 7 Testing Framework for the
Page 137 and 138:
7.1. Metrics 137 various New-TCP al
Page 139 and 140:
7.1. Metrics 139 assuming equilibri
Page 141 and 142:
7.2. Environment 141 of packets in
Page 143 and 144:
7.2. Environment 143 T 0 B, T T 1 F
Page 145 and 146:
Chapter 8 Systematic Tests of New-T
Page 147 and 148:
8.1. Methodology 147 fed into the d
Page 149 and 150:
8.1. Methodology 149 Various other
Page 151 and 152:
8.2. Test Calibration 151 Goodput (
Page 153 and 154:
8.2. Test Calibration 153 cwnd and
Page 155 and 156:
8.2. Test Calibration 155 100 1 Agg
Page 157 and 158:
8.2. Test Calibration 157 1000 FAST
Page 159 and 160:
8.3. Results 159 10 100 Total Throu
Page 161 and 162:
8.3. Results 161 14 12 HSTCP 1 α H
Page 163 and 164:
8.3. Results 163 cwnd (packets) 120
Page 165 and 166:
8.3. Results 165 Fraction Bytes Ret
Page 167 and 168:
8.3. Results 167 1200 1000 HTCP 1 H
Page 169 and 170:
8.3. Results 169 This is due to the
Page 171 and 172:
8.3. Results 171 4000 3500 3000 HTC
Page 173 and 174:
8.3. Results 173 800 700 600 Standa
Page 175 and 176:
8.3. Results 175 1 1 Fairness Ratio
Page 177 and 178:
8.3. Results 177 10 100 Total Throu
Page 179 and 180:
8.3. Results 179 1000 Standard TCP
Page 181 and 182: 8.3. Results 181 600 500 Standard T
Page 183 and 184: 8.4. Discussion of Results 183 Asym
Page 185 and 186: 8.4. Discussion of Results 185 New-
Page 187 and 188: 8.4. Discussion of Results 187 Asym
Page 189 and 190: 8.5. Summary 189 The importance of
Page 191 and 192: 9.1. Transfer Tests Across Dedicate
Page 215 and 216: 9.2. Preferential Flow Handling Usi
Page 229 and 230: 9.3. Internet Transfers 229 can cau
Page 231: 9.3. Internet Transfers 231 Stanfor
Page 235 and 236: 9.3. Internet Transfers 235 1000 90
Page 237 and 238: 9.3. Internet Transfers 237 Fractio
Page 239 and 240: 9.3. Internet Transfers 239 10000 c
Page 241 and 242: 9.3. Internet Transfers 241 achieve
Page 243 and 244: 9.4. Summary 243 performance of sin
Page 245 and 246: 9.4. Summary 245 Nevertheless, a Di
Page 247 and 248: 10.1. Summary 247 PCs. But at 1Gb/s
Page 249 and 250: 10.2. New-TCP Suitability Study 249
Page 251 and 252: 10.2. New-TCP Suitability Study 251
Page 253 and 254: 10.3. Future Directions 253 low ove
Page 255 and 256: 10.3. Future Directions 255 porates
Page 257 and 258: 10.3. Future Directions 257 A more
Page 259 and 260: 10.4. Conclusion 259 This dissertat
Page 261 and 262: A.1. Data Storage 261 with the choi
Page 263 and 264: A.1. Data Storage 263 gether with h
Page 265 and 266: A.1. Data Storage 265 The virtual m
Page 267 and 268: A.1. Data Storage 267 Read Speed (M
Page 269 and 270: A.1. Data Storage 269 Read Speed (M
Page 271 and 272: A.1. Data Storage 271 1600 1600 140
Page 273 and 274: A.2. Network Interface Cards 273 la
Page 275 and 276: A.2. Network Interface Cards 275 de
Page 277 and 278: A.2. Network Interface Cards 277 of
Page 279 and 280: A.2. Network Interface Cards 279 PC
Page 281 and 282: A.2. Network Interface Cards 281 70
Page 283 and 284:
A.2. Network Interface Cards 283 Th
Page 285 and 286:
A.2. Network Interface Cards 285 20
Page 287 and 288:
A.2. Network Interface Cards 287 ha
Page 289 and 290:
B.2. Wide Area Network Tests 289 Da
Page 291 and 292:
C.3. DataTAG 291 Cisco 7606 Cisco 7
Page 293 and 294:
Appendix D Network Paths of Interne
Page 295 and 296:
295 ham02gva.datatag.org:192.91.239
Page 297 and 298:
Bibliography 297 [All03] [AMA + 99]
Page 299 and 300:
Bibliography 299 [BPSK96] H. Balakr
Page 301 and 302:
Bibliography 301 [DC02] [ea02] [FF9
Page 303 and 304:
Bibliography 303 [Flo03] [Flo04] S.
Page 305 and 306:
Bibliography 305 [Hoe96] [IET] [Jac
Page 307 and 308:
Bibliography 307 [KRB05] [Lah00] [L
Page 309 and 310:
Bibliography 309 [MHR03] [MM96] M.
Page 311 and 312:
Bibliography 311 [Pos81a] J. Postel
Page 313 and 314:
Bibliography 313 [tePM] IEPM Intern
show all

An Investigation into Transport Protocols and Data Transport ...

Create successful ePaper yourself

Delete template?

Save as template?