Chapter 8 Experimental Results – TCP Highlights
8.7 E FFECT OF LEAKY BUCKET INITIAL SIZE
In sections 8.3 and 8.4, we showed the effect of using a token bucket policer on TCP sources. TCP is a bursty protocol in nature, and the main problem of using such policer types is that they impose bursty losses (and thus poor performance) on TCP if the TCP burst does not fit in the configured CBS parameter. There are several papers that discuss the impact of
token bucket policers on TCP and some even propose more TCP-friendly policers [87 (along with its referenced papers)].
In this section, we will not analyze the effect of CBS on TCP due to prior art as mentioned above, however we will briefly show the effect of the Gold CBS parameter for the various models. We will fix the CBR8 (Silver Background rate) value at 1,450 KBps (which we observed earlier as a high enough value to cause reordering in the SLAR model without congesting the network) while varying the Gold CBS. All other parameters are the same as in section 8.5 i.e. the Gold policer parameters set to [CBS=variable, CIR=330,600], the Silver policer parameters set to [CBS=32,000, CIR=230,600], and the Bronze policer parameters set to [CBS=32,000, CIR=230,600].
Figure 112 and Figure 113 show Client2’s Gold goodput versus the Gold CBS for TCP packet sizes of 1,000 Bytes and 1,500 Bytes, respectively. The corresponding maximum approximated TCP rates are 20x1,000/120 = 166KBps, and 20x1,500/120 = 250KBps for the mentioned packet sizes.
0 20000 40000 60000 80000 100000 120000 140000 160000 180000
0 5000 10000 15000 20000 25000
(Bps)
Gold_CBS(Bytes)
DS/Gold2/Good SLAR/Gold2/Good SLA/Gold2/Good
Figure 112 Gold TCP goodput Versus the Gold CBS (Packet size = 1,000B)
0 50000 100000 150000 200000 250000 300000
0 5000 10000 15000 20000 25000 30000 35000 40000
(Bps)
Gold_CBS(Bytes)
DS/Gold2/Good SLAR/Gold2/Good SLA/Gold2/Good
Figure 113 Gold TCP goodput Versus the Gold CBS (Packet size = 1,500B)
Both figures show that the CBS parameter severely affects the TCP goodput as
predicted (due to TCP burstiness). What is noticeable about the above figures is the similarity in the shape of the curves where the goodput increases, then we observe a small bump, then another increase, a jump and then a flat line. What could cause such bumps (at 7,000 Bytes for 1,000 Bytes packets and at 10,500 Bytes for 1,500 Bytes packets) and early saturations (flat line at 15,000Bytes for 1,000 Bytes packets and 22,500Bytes for 1,500 Bytes packets), given that the cwnd is equal to 20 packets (in NS-2, the cwnd is in units of packets) and that the full burst should have been 20x1,000=20,000Bytes for 1,000 Bytes packets and
20x1,500=30,000Bytes for 1,500 Bytes packets?
Remember that TCP slow start is actually an exponential curve, where the window size doubles every RTT (first packet is sent, upon ACK reception 2 packets are sent, upon both ACK reception, 4 packets are sent and so on). It seems that the bumps coincide with CBS = 7 packet-sizes (7x1,000=7,000 and 7x1,500=10,500) and that the early saturations coincide with CBS = 15 packet-sizes (15x1,000=15,000 and 15x1,500=22,500). In fact, 7 is
the sum of 1, 2, and 4, and 15 is the sum of 1, 2, 4 and 8. Notice also in the graphs that there is a small bump at CBS = 3,000Bytes in Figure 112 and at CBS=4,500Bytes in Figure 113, both corresponding to the sum of 1 plus 2 (=3 packet-sizes). So it seems that as CBS increases we have some preferred values corresponding to the sum of the discrete
exponential series starting at 1; these values being 3 (1+2), 7 (1+2+4), and 15 (1+2+4+8). As explained earlier, the slow start property of TCP triggered upon timeout detection of losses causes this exponential increase in window size.
Figure 114 shows Client2’s Silver TCP goodput versus the Gold CBS. The figure shows that the Silver TCP goodput takes advantage of upgrades for the SLAR and SLA cases as soon as there is at least 2 packet sizes worth of tokens (2x1,500=3,000Bytes). The Gold token refresh rate being at 330KBps, the Gold bandwidth gap of 330KBps minus the Gold- rate as per Figure 113 for a particular Gold CBS value was enough for the Silver to get the upgrade needed to reach the approximated maximum TCP rate of 250KBps.
140000 160000 180000 200000 220000 240000 260000
0 5000 10000 15000 20000 25000 30000 35000 40000
(Bps)
Gold_CBS(Bytes)
DS/Silver2/Good SLAR/Silver2/Good SLA/Silver2/Good
Figure 114 Silver TCP goodput Versus the Gold CBS (Packet size = 1,500B)
To summarize this section’s findings, we showed that the CBS value severely impacts TCP performance, as observed in several prior art, due to limiting the TCP burst size. We also showed that TCP prefers some CBS values, corresponding to the sum of the discrete exponential series (Sum(2n) , e.g. 3, 7, 15), due to the exponential slow start behavior upon timeout detection of packet loss. Finally, we observed that it doesn’t require a large upper class CBS value for the lower class to benefit from upgrades (an upper CBS value of 2 packets was enough).