Chapter 8 Experimental Results – TCP Highlights
8.8 E FFECT OF R OUND T RIP T IME
In the previous sections, we used a propagation delay of 10ms over all links in the network. We also said that the maximum TCP throughput is approximated at cwnd/RTT (or 20x1500/(2x6x10ms) = 250KBps. In this section, we show the Round Trip Time (RTT) effect on the TCP throughput for the various models by varying the per-link propagation delay.
Figure 115 shows Client2’s Gold TCP goodput versus the per-link propagation delay.
The minimum RTT is estimated at 2x6x10ms = 120ms (ignoring transmission, queuing and processing delays); this is the time it takes the packet to traverse 6 links in the forward and backward direction. The figure shows that when the propagation delay is small, the TCP goodput is lower than expected, then it peaks out at 8ms to drop back down as the
propagation delay is further increased. Remember that we are policing Client2’s Gold traffic at 330KBps, and thus even though with a smaller RTT the TCP throughput should be higher, the policer keeps the TCP throughput below the CIR.
180000 200000 220000 240000 260000 280000 300000 320000
0 2 4 6 8 10 12
(Bps)
Prop_delay(ms)
DS/Gold2/Good SLAR/Gold2/Good SLA/Gold2/Good
Figure 115 Gold TCP Goodput Versus the per link propagation delay
Figure 116 shows the same data as Figure 115, except that we add the theoretical approximated maximum TCP throughput for RTT = 2x6xProp_delay labeled min_RTT, the estimated RTT with an extra 10ms delay (due to queuing/transmission for example) labeled RTT_10 and the policing rate at ingress labeled CIR. The min_RTT and the RTT_10 curves show the approximated TCP throughput had we not policed on ingress for 2 different values of RTT. From this figure, we can see that when the per-link propagation delay is smaller than 8ms, the RTT is small enough to result in a TCP throughput higher than the policed rate.
Which means that, for propagation delay values of less than 8ms, the TCP Gold traffic has observed drops, due to policing, and that these drops forced TCP into congestion avoidance (thus lowering its rate). As the RTT is increased, the TCP throughput is decreased resulting in lower packet drops due to policing (Figure 117 shows the drops due to policing). For a per-link propagation delay of 8ms or more, the TCP maximum throughput becomes smaller than the policed rate and all three models throughput match the approximated theoretical
curve (labeled min_RTT in the graph). This explains the behavior in Figure 115 where the goodput peaks out at 8ms and then drops down.
0 500000 1e+06 1.5e+06 2e+06 2.5e+06
0 2 4 6 8 10 12
(Bps)
Prop_delay(ms)
DS/Gold2/Good SLAR/Gold2/Good SLA/Gold2/Good min_RTT/Gold2/Good RTT_10/Gold2/Good CIR/Gold2/Good
Figure 116 Max Theoretical TCP goodput Versus the per link propagation delay
0 5000 10000 15000 20000 25000 30000
0 2 4 6 8 10 12
(Bps)
Prop_delay(ms)
DS/Gold2/Drop SLAR/Gold2/Drop SLA/Gold2/Drop
Figure 117 Gold ingress drop rate due to policing Versus the per link propagation delay
Figure 115 shows that if the RTT is over- or under-estimated, then the customer would be wasting some of his/her paid-for reserved bandwidth. For the customer to fully use
his/her paid-for reserved bandwidth, s/he needs to accurately estimate the RTT. Or in other words, if the customer does not know the bandwidth he needs to purchase from a provider, the most efficient bandwidth value (efficient in terms of utilization) would be dictated by an accurate estimation of the RTT. In general, the TCP RTT can not be estimated a priori unless the end-points within the network are predefined; Even if the end-points are defined, the network queuing, routing, etc. behavior is not always deterministic. So, could the automatic upgrades SLA benefit from such a TCP characteristic?
Figure 118 shows Client2’s Silver TCP goodput versus the per link propagation delay (we also left the DS Gold goodput for convenience, the SLAR and SLA Gold goodput being the same as per Figure 115 – we don’t upgrade the Gold traffic). Note that both the SLAR and SLA models show that they indeed benefited from holes created by a “bad” estimate of RTT when compared to the DS model. Remember that we are policing the Gold at 330KBps, and that for a per-link propagation delay of less than 8ms, the Gold TCP could not reach the CIR rate due to ingress dropping. However, the Gold gap generated by this TCP
characteristic was used by the Silver traffic to get higher throughput (the Silver traffic was policed at 230KBps; for the DS case, for a small RTT, the Silver traffic was dropped due to ingress policing resulting in a TCP rate much less than 230KBps). At 8ms, the Gold gap is at its minimum since the Gold traffic is almost completely using the reserved Gold bandwidth and thus both the SLAR and SLA dip to minimum upgrades (as also shown in Figure 119).
For a propagation delay higher than 8ms, the RTT becomes large enough to make the maximum TCP rate less than the CIR as already mentioned, and thus the SLAR and SLA models use this gap to upgrade Silver packets.
120000 140000 160000 180000 200000 220000 240000 260000 280000 300000 320000
0 2 4 6 8 10 12
(Bps)
Prop_delay(ms)
DS/Gold2/Good DS/Silver2/Good SLAR/Silver2/Good SLA/Silver2/Good
Figure 118 Silver TCP Goodput Versus the per link propagation delay
0 50000 100000 150000 200000 250000 300000 350000
0 2 4 6 8 10 12
(Bps)
Prop_delay(ms)
DS/Gold2/Good DS/Silver2/Upgr SLAR/Silver2/Upgr SLA/Silver2/Upgr
Figure 119 Silver TCP Upgrade rate Versus the per link propagation delay
As a summary to this section, we showed that in the DS model, for the client bought bandwidth to be fully utilized, it needs to be derived from an accurate estimation of the RTT.
Since in a network a client is not necessarily communicating with the same endpoint, and since the RTT depends on the state of the network for a specific endpoint, the bandwidth
bought can be severely underutilized. We also showed that the SLAR and the SLA models can use the gap generated from such a TCP behavior to upgrade lower-class packets to fill that gap, resulting in a much better utilization of the paid-for reserved bandwidth.