Chapter 8 Experimental Results – TCP Highlights
8.4 L OOSE I NGRESS P OLICING WITH UPGRADES
In the previous section, we established that the low performance of TCP was due to the ingress edge policing. In this section, we will increase both the CBS and CIR policer parameters for Gold, Silver and Bronze classes to 32,000 bytes and 330,600 Bps,
respectively, in order to compare the TCP performance of the DS, SLAR and SLA models (previous CBS was 5,000 bytes, and previous CIR was 153,600 Bps). Note that we
previously approximated the maximum TCP rate to be around (20x1500)/121ms = 250KBps for a max cwnd = 20 packets, average packet size of 1500 bytes and an average round trip time of 121ms.
Figure 97 shows the DS, SLAR and SLA Client2’s Silver TCP goodput versus the background Silver rate. The figure shows that when there is no congestion in the Silver pipe (i.e. CBR8<1,400KBps), all three models have the same performance. As we mentioned earlier, all three models are bounded by the maximum TCP rate for a round trip time of 121ms which was calculated to be 250KBps, regardless of upgrades (the upgrades rates is shown in Figure 98). The forward delay is shown in Figure 99 (the backward delay is assumed to be 60ms). Notice in Figure 99 that the delay for the non-congestion zone is the same for all 3 models and is equal to 61ms.
0 50000 100000 150000 200000 250000 300000
0 500000 1e+06 1.5e+06 2e+06 2.5e+06
(Bps)
CBR8(Bps)
DS/Silver2/Good SLAR/Silver2/Good SLA/Silver2/Good
Figure 97 Silver TCP Goodput Versus Silver Background rate
0 10000 20000 30000 40000 50000 60000 70000 80000
0 500000 1e+06 1.5e+06 2e+06 2.5e+06
(Bps)
CBR8(Bps)
DS/Silver2/Upgr SLAR/Silver2/Upgr SLA/Silver2/Upgr
Figure 98 Silver TCP Upgrades Versus Silver Background rate
61 62 63 64 65 66 67 68 69 70
0 500000 1e+06 1.5e+06 2e+06 2.5e+06
(ms)
CBR8(Bps)
DS/Silver2/Delay SLAR/Silver2/Delay SLA/Silver2/Delay
Figure 99 Silver TCP average forward delay Versus Silver Background rate
Notice in Figure 97 that the SLAR model has lower goodput at CBR8=1,450KBps than the other 2 models and then for CBR8 > 1,750KBps the SLAR model gets a better goodput than the other 2 models. At CBR8=1,450KBps, both the DS and the SLA models did not observe any significant losses, so there is no reason why the SLAR model would
observe losses, especially that the SLAR model uses the Gold pipe for upgrades bypassing Silver congestion. Looking at the 3 or more duplicate ACK count in Figure 100, we see that the SLAR model has a peculiarly higher count than the other 2 models. Remember that for the SLAR model, this count is both lost packets and reordered packets that could cause a transmitter rate reduction (fast retransmit). This is surely an indication of the latter (reordered packets) since neither the DS and the SLA models have observed losses at the same rate.
This surely explains why the SLAR model performs poorly at CBR8=1,450KBps (a difference of 70 KBps with the other 2 models).
-1 0 1 2 3 4 5 6
0 500000 1e+06 1.5e+06 2e+06 2.5e+06
(Count)
CBR8(Bps)
DS/Silver2/3Ack SLAR/Silver2/3Ack SLA/Silver2/3Ack
Figure 100 Silver TCP 3+ duplicate ACK count Versus Silver Background rate
The reason why the SLAR model outperforms both the DS and the SLA models during congestion periods (CBR8 > 1,750KBps), is due to the fact that the SLAR model uses the Gold pipe to bypass the Silver congestion as explained in section 8.3. For relatively the same amount of 3 or more duplicate ACKs (as indicated in Figure 100) and relatively lower Round Trip Times (as indicated in Figure 99) the SLAR model has a better goodput.
To summarize this section’s findings, we showed that when we don’t tightly police the TCP rates on ingress, all three models, DS, SLA and SLAR, have the same performance when there is little congestion in the network, even if we allow ingress upgrades. The reason for this was that the TCP rate is bounded by a maximum rate approximated by cwnd/RTT and would not take advantage of upper class upgrades beyond this limit. We also reinforced the findings of the previous section, in particular that the SLAR model has an advantage when a certain class (e.g. Silver) is congested, since upgraded packets flow on the upper class (e.g. Gold) bypassing the congested (i.e. Silver) class. Finally, we showed that the SLAR reordering characteristic affected the TCP throughput under certain conditions where the network is on the verge of congestion where the probability of upgraded packets reordering is maximized (Silver packets are delayed so that subsequent Gold packets bypass them).
However, the TCP receiver dampened the reordering impact due to buffering out of order packets; only when reordering causes 3 duplicate ACKs that the TCP transmitter rate is affected.