Fig. 1. Telnet packet size distribution.

Fig. 2. Example of FTP session.

Fig. 3. Time series of entire 2 h VBR video.

Fig. 4. D-ITG platform architecture.

Fig. 5. Data rate analysis.

Fig. 6. Architecture of D-ITG with log server.

Fig. 7. ITGSend maximum data rate analysis. (a) Maximum data rate for the D-ITG sender; (b) loss rate in the generation process.

Fig. 8. Network scenario for ITG-LS TCP implementation.

Fig. 9. MPI D-ITG architecture.

Fig. 10. ITGSend maximum data rate analysis. (a) Maximum data rate for the D-ITG receiver; (b) loss rate in the receiving process.

Fig. 11. Testbed architecture.

Fig. 12. Resulting plots of TCP generation with IDT = Pareto and PS = constant (window size of 1 s).

Fig. 13. Output results of TCP generation with IDT = Pareto and PS = constant.

Fig. 14. Ethereal snapshot.

Fig. 15. Mcalc output.

Fig. 16. Poisson packet size trend (D-ITG output).

Fig. 17. TCP traffic using uniformly distributed PS.

Fig. 18. TCP mcalc snapshot.

Fig. 19. UDP video traffic (D-ITG output).

Fig. 20. Two hours of telnet traffic.

Fig. 21. TCPLib.

Fig. 22. Mcalc telnet output.

Fig. 23. D-ITG telnet PS.

Fig. 24. Mcalc VoIP output. (a) Mcalc VoIP output (Codec G.711 with one voice sample per packet and without VAD); (b) mcalc VoIP output (Codec G.711 with one voice sample per packet and with VAD); (c) mcalc VoIP output (Codec G.711 with two voice samples per packet and without VAD); (d) mcalc VoIP output (Codec G.729 with two voice samples per packet and without VAD); (e) mcalc VoIP output (Codec G.729 with three voice samples per packet and without VAD); (f) mcalc VoIP output (Codec G.723.1 with one voice sample per packet and without VAD).

Fig. 25. Sender buffer size analysis—data rate (c=512; t=60 s): this figure depicts the packet rate achieved by the different implementations of ITGSend as a function of the required packet rate C, for different values of the log buffer size D. The achieved maximum packet rate grows as D increases. For every implementation, it is possible to determine an upper bound for the maximum packet rate. ITG-CV achieves the smallest upper bound (about 28,000 pkt/s when D is at least 30) while the other three implementations perform better (about 30,000 pkt/s when D is at least 40).

Fig. 26. Sender buffer analysis (c=512; t=60 s): this figure shows the percentage of packets the sender is not able to generate as a function of C, for different values of D. ITG-LS and ITG-MPI exhibit a similar behavior (a percentage error smaller than 2% for C<30,000 and D=40) and perform better than ITG-CV. In particular, it is possible to note that ITG-MPI and the UDP version of ITG-LS perform slightly better than the TCP version of ITG-LS (a percentage error almost null for C=30,000 and D=40 against the percentage error of 1.5% achieved by the TCP version of ITG-Log).

Fig. 27. Sender packet size analysis (C=26,000 for ITG, C=28,000 for ITG-LS and ITG-MPI; t=60 s): this figure illustrates the percentage error as a function of c; for all the implementations of D-ITG, the generation error is null for c<1024. For values above this threshold, we can observe a considerable packet loss.

Fig. 28. Sender max data rate analysis (c=1024; D=40): this figure depicts the generated data rate as a function of C, while D and c are equal to their optimal values. For all the implementations of D-ITG, it is possible to determine an upper bound to the generation rate. The smallest upper bound is achieved by ITG-CV and is about 218,500 Kbps. The other implementations perform similarly, having a maximum data rate equal to about 230,000 Kbps (that is, 5% greater).

Fig. 29. Sender max data rate analysis (c=1024; D=40): this figure shows the percentage generation error for the three implementations of D-ITG. We can observe that only ITG-MPI and the UDP version of ITG-LS achieve a generation error smaller than 5% when the expected packet rate is 300,000 pkt/s (which corresponds to an actual data rate of 245,760 kbps).

Fig. 30. Receiver buffer size analysis (c=512; t=60 s): this figure represents the number of packets received by the different implementations of ITGRecv as a function of C, while increasing D. The maximum packet rate increases as D grows. For all the implementations, it is possible to calculate an upper bound. This value is similar for all the component of D-ITG and is equal to 22,000 pkt/s.

Fig. 31. Receiver buffer analysis (c=512; t=60 s): this figure illustrates the percentage of lost packets as a function of C for different values of D. The different implementations of D-ITG exhibit a similar behavior for C<22,000.

Fig. 32. Receiver packet size analysis (C=20,000; t=60 s): this figure shows the percentage error (percentage of packets not received) as a function of c. For all the implementations of D-ITG, the percentage error is null for c<1024. For values above this threshold, we can observe a considerable packet loss.

Fig. 33. Receiver max data rate analysis (c=1024; D=40): this figure depicts the received data rate as a function of C, while D and c are equal to their optimal values. For all the implementations of D-ITG, it is possible to determine an upper bound to the received rate. The smallest value is achieved by ITG-CV and is about 165,000 Kbps. The two versions of ITG-LS and ITG-MPI exhibit a similar behavior and their upper bound is about 180,000 Kbps. The gain achieved with respect to ITG-CV is about 10%.

Fig. 34. Receiver max data rate analysis (c=1024; D=40): this figure shows the percentage of lost packets for the three implementations of D-ITG. The two versions of ITG-LS and ITG-MPI exhibit a null error rate for C<22,000. ITG-LS presents a loss rate of 7% for C=22,000. For C above this value, the error rate grows, but it is still below 8% for C=24,000 for the two versions of ITG-LS.

Hardware characteristics