Transiving Global Fairness Packets

A local station receives a global fairness packet via the opposing ringlet. If the received global fair rate is smaller than local fair rate, inter-ring traffic is limited by global fair rate and intra-ring traffic still obeys local fairness mechanism. Otherwise, only local fairness algorithm is implemented. This property ensures the coexistence of local and global fairness. A local station continues to send GF to the upstream node neighbor.

Without virtual destination queues, Bridge can not control certain IIA flow. Once a bridge receives multiple global fairness packets from different bridges, it will choose the smallest GF between itself generated GF and the received GFs.

Chapter 4

Simulation Results and Discussions

4.1 Simulation Environment

In this section, we compare our proposed intelligent global fairness controller (IGFC) with RIAS based global fairness controller (RGFC), [12]. The link capacity is 10Gbps (OC-192) and propagation delay between stations is 100μs. A uniform data packet is 1616 bytes and a fairness packet is 16 bytes. The agingInterval is 100μs and we observe and record the simulation result every agingInterval. We assume all traffic is best effort traffic. Some common parameters specified in IEEE 802.17 are regulated as Table 4.1.

Table 4.1: System Parameters

Parameters Values

Link Capacity 10Gbps

Propagation Delay 100μs

ageCoef 8 rampUpCoef 128 rampDnCoef 128 lpCoef 128 rateCoef 0.125

We concentrate on a BRPR network with small and large topology scenario to justify the coexistence of global fairness and local fairness and to notice the influence of

propagation delay. For focusing on the transient behavior of inter-ring flows, we examine the dynamic traffic scenario which inter-ring flows start to transmit at different times.

Unbalanced traffic scenario is concerned to realize what the weakness of AM leaves the global fairness controller and how IGFC fights against it.

We inspect the transmission rate of each source node to validate the global fairness criteria and the advantage of fair rate adjustment in fuzzy logic. The observation of throughput which is the received rate at the destination node can represent the gain of IGFC. Besides, dropping probability at bridge and averaged access delay at bridge are drawn to declare the significant system measure of IGFC by using fuzzy logic control. We do not perform the buffer overflow prevention scheme of RGFC, as described in chapter 1.4, since it is inadequate for supporting the real non-blocking quality of service. Here are the comparisons between IGFC and RGFC in Table 4.2.

Table 4.2: Comparisons between IGFC and RGFC

Elements and Attributions

IGFC RGFC

Ingress buffer 2 (4MB each) 1 (8MB)

Local ringlet buffer 1 (4MB) 1 (4MB)

Buffer management FIFO FIFO

Traffic Scheduling on ingress and local ringlet buffer

Obey per byte sate machine in IEEE 802.17

Obey per byte sate machine in IEEE 802.17 Traffic Scheduling on

ingress buffers

Dynamic weighted

round robin None (unnecessary) Global congested detection None 1. Ingress buffer length

2. Received local fair rate Global fair rates generator Global fairness criteria

+ Fuzzy control Global fairness criteria

Local fairness algorithm AM AM

4.2 Small Topology Scenario

Figure 4.1 (a) exhibits a simple BRPR network with 7 nodes, where node 0, 1 and 2 are located in Rk and node 3, 4, 5, and 6 are located in Rk+1. All nodes except node 6 will transmit greedy traffic to node 6. Flow (0, 6), flow (1, 6), and flow (2, 6) are inter-ring traffic. Flow (0,6) and flow (1,6) are forwarded by CW ingress buffer and flow (2, 6) is forwarded by CCW ingress buffer of IGFC4. Flow (3, 6), flow (4, 6), flow (5, 6) are local traffic. We want to inspect the affection of intra-ring traffic across the bridge, flow (3, 6) and flow (4, 6) which will transit through the local ringlet buffer of IGFC4.

(a) Scenario Setup

(b) Average Access Delay at Bridge

0

flow(0, 6) flow(1, 6) flow(2, 6) flow(3, 6) flow(4, 6) flow(5, 6) :

flow(0, 6) flow(1, 6) flow(2, 6) flow(3, 6) flow(4, 6) flow(5, 6) :

:

(d) RGFC

Figure 4.1: Small Topology Scenario. (a) Scenario Setup (b) Average Access Delay at Bridge, (c) IGFC, and (d) RGFC.

Figure 4.1 (b) shows the averaged access delay function of time to compare which controller has smaller access delay. We record the averaged access delay per 10ms and Y-axis is the number of agingInterval (100μs). Access delay is measured when a packet enter the buffer until it is served. It means a packet needs to be waited for the number of computational periods to being served. We can find that even if IGFC has two ingress

buffers, the access delay of CW or CCW ingress buffer of IGFC is very far smaller than RGFC’s. Arrival rate of each source node is not fully consistent because of propagation delay; in other words, it does not always match the current GF. When the total arrival rate has been larger than the available bandwidth for some agingIntervals, the buffer length increases and the buffer is inclined to overflow. RGFC with longer access delay means RGFC does not perform effective strategy to throttle buffer occupancy growing. Since the buffer occupancy is concerned into FGFE, IGFC provides the capability to adapt to system dynamics. Fortunately, both IGFC and RGFC do not occur buffer overflow because the defect of propagation delay does not appear apparently in small topology scenario. Upstream nodes can react to the bridge congestion and throttle their transmit rate fast. However, RGFC is close to buffer overflow.

Figure 4.1 (c) and (d) display throughput versus time by taking IGFC and RGFC, respectively. Both IGFC and RGFC not only successfully maintain local fairness but achieve global fairness. Intra-ring flow (3, 6), flow (4, 6), and flow (5, 6) all achieve 2.5Gbps; meanwhile, inter-ring flow (0, 6), flow (1, 6), flow (2, 6) equally share 2.5Gbps, which is 833.33Mbps each. The movements of local traffic by IGFC and RGFC are equivalent even if some intra-ring traffic is across the bridge. This is because AM local fairness algorithm is used in nodes and the bridge as well. If throughput is around 1.5%

deviation of the ideal fair rate, we say that using the controller converges successfully in this scenario. Hence, the convergence time of IGFC for inter-ring traffic is 31ms, but the convergence time of RGFC for inter-ring traffic is 46ms. Even IGFC starts to be gradually stable after 20ms; nevertheless, RGFC is still unstable before 40ms. That is, using IGFC is better than RGFC at small topology scenario.

4.3 Large Topology Scenario

Propagation delay may postpone the convergence time and make severe oscillation since the far upstream nodes need to wait for more time until they receive the global fairness packet and limit their transmit rate. Therefore, we consider a large topology scenario, as illustrated in Figure 4.2 (a), where node 0, 1, 2, 3, 4, 5, and 6 are settled in Rk

and the other nodes are settled in Rk+1. All nodes except node 10 have infinite traffic demands to node 10. Flow (7, 10), flow (8, 10), and flow (9, 10) are local traffic and the others are inter-ring traffic. Throughput, dropping probability, and transmission rate will be emphasized.

(a) Scenario Setup

(b) Throughput by IGFC

(c) Throughput by RGFC

(d) Dropping Probability at Bridge

0

flow(0, 10) flow(1, 10) flow(2, 10) flow(3, 10) flow(4, 10) flow(5, 10) flow(6, 10)

:

0 10 20 30 40 50 60 70 80 90 100

(e) Transmission Rate of each source node by IGFC

(f) Transmission Rate of each source node by RGFC

Figure 4.2: Large Topology Scenario. (a) Scenario setup, (b) Throughput by IGFC, (c) Throughput by RGFC, (d) Dropping Probability at Bridge, (e) Transmission Rate of each source node by IGFC, and (f) Transmission Rate of each source node by RGFC.

Figure 4.2 (b) and (c) demonstrate throughput versus time by IGFC and RGFC, respectively. Regardless of intra-ring traffic, apparently, IGFC has better performance and stability than RGFC for inter-ring flows. The ideal global fair rate is 357.14Mbps and the ideal local fair rate is 2.5Gbps. IGFC trends to converge at 20ms and has a fast convergence time of 40ms, but RGFC has a slow convergence time of 54ms. In addition,

the variation of inter-ring flows by RGFC is more terrible and irregular. This is because too many distinct inter-ring flows enter the ingress buffer of RGFC. It is hard to manage all kinds of traffic due to first-in-first-out discipline plus the effect of propagation delay. In comparison with the last scenario, we can find that the propagation delay problem appears more terrible in large topology scenario, whether we focus on IGFC or RGFC. Since only the bridge computes GF and other nodes just propagate GF to their upstream node neighbor, for example, the bridge has to wait at least 10 round trip time to receive the reacting traffic of node 0 after the bridge has published the GF.

We record packets dropping probability at bridge every 100 agingIntervals in Figure 4.2 (d). Dropping probability is the number of dropped packets over transmitted packets during 10ms. Evidently, IGFC has zero packet loss; nevertheless, RGFC has at most 0.28 packet dropping probability. Fortunately, it does not occur buffer overflow after 20ms.

The reason is that inter-ring traffic converges gradually and the congestion at bridge is solved. However, the utilization of the ingress buffer continues to be fully occupied and it is still not allowed because one of the BRPR targets is without packet loss.

Figure 4.2 (e) and (f) display the transmission rate of each source node, except node 7, 8, and 9. It is due to the same performance of local traffic. The observation at source nodes is eliminated from the damage of propagation delay as more as possible and is obvious if the GF was calculated correctly by the bridge. The convergence time of IGFC and RGFC is almost 26ms. However, IGFC adjusts with fewer and moderate oscillations, but RGFC adjusts with more and intense oscillations. It means IGFC can generate more precise GF. In RGFC, GF is only calculated according to the global fairness criteria.

Unfortunately, when we face a large network, the remote nodes can not adapt their transmit rate as quickly as nodes close to the bridge. This makes the calculated GF is over

raised or reduced because of different arrival rate of each node. In IGFC, not only global fairness criteria but also the buffer occupancy alters the GF. We use fuzzy control to modulate GF since two ringlet buffers may store much enough data to serve out. Hence, fuzzy control is more sensitive and reflects to the real situation of the current network environment.

4.4 Dynamic Traffic Scenario

As illustrated in Figure 4.3 (a), node 0, 1, 2, and 3 are in Rk and the others are in Rk+1. In this scenario, we want to study the transient behavior for inter-ring traffic. Node 1 and node 3 and node 4 start at 0s; node 0 starts at 50ms; node 2 starts at 100ms. All traffic demands are greedy. Our destination is node 5. Figure 4.3 (b) and (c) present throughput versus time by using IGFC and RGFC without ringlet selection, respectively. In other words, each inter-ring flow will choose the shortest path to forward. Otherwise, Figure 4.3 (d) also demonstrates throughput versus time by adopting IGFC with weighted ringlet selector (WRS) and each flow will decide a suitable path when it enters the bridge. We set α to 0.6, that is; to wit, traffic load is more important than distance.

(a) Scenario Setup

0 1000 2000 3000 4000 5000 6000

flow(0, 5) flow(1, 5) flow(2, 5) flow(3, 5) flow(4, 5)

flow(3, 5), flow(1, 5) flow(0, 5), flow(3, 5), flow(1, 5)

flow(2, 5)

52ms 102ms

0 20 40 60 80 100 120 140 160

(b) IGFC

(c) RGFC

(d) IGFC with WRS

Figure 4.3: Simple Rate Changing Scenario. (a) Scenario Setup, (b) IGFC, (c) RGFC, and (d) IGFC with WRS.

Since the scenario only has a local traffic flow (node 4), and has symmetric and few inter-ring traffic flows, there are few oscillations and short convergence time with IGFC and RGFC. In Figure 4.3 (b), IGFC first converges to the ideal global fair rate at 56ms after node 0 begins to transmit traffic at 50ms; it second converges to the ideal global fair rate at 105ms after node 2 starts to transmit traffic at 100ms. In Figure 4.3 (c), the first convergence time by RGFC is 66ms and the second convergence time is 115ms. It can be found that using IGFC has a fast convergence time than using RGFC. The difference of convergence time is about 10ms which is 100 computation periods (agingInterval). This is because when node 0 or node 2 starts to transmit traffic, the ingress buffer of RGFC has keep much traffic and the destination node can not receive the new added traffic immediately.

IGFC has a great advantage on buffer management, but has some fluctuations while the traffic is changing. The reason is that scheduling between CW and CCW ingress buffers is measured by the dynamic weighted round robin. For example, CW and CCW ingress buffers are served at 2:1 ratio as flow (0, 5) reaches the bridge. Meanwhile, flow (0, 5) can not be served yet so flow (1, 5) has larger throughput. Similarly, flow (3, 5) fluctuates but flow (2,5) waits for being served around 100ms.

Figure 4.3 (d) reveals the advantage of using WRS. Inter-ring flows choose a suitable path according to Table 2.1. In the beginning, distance is the dominant factor and flow (1, 5) and flow (3, 5) are selected the shortest path (CCW) by SAS. Consequently, flow (1, 5) and flow (3, 5) are put into CW and CCW ingress buffer of IGFC4, respectively. As flow (0, 5) and flow (2, 5) start to transmit, they are forwarded to the CW path through IGFC3.

Although transmitting traffic along the CW path is twice as far as the opposing path, traffic load dominates the decision at that time. Even though IGFC3 and IGFC4 generate each GFcw and GFccw for node 0 and 1, and node 2 and 3, respectively, SAS would pick the smaller GFcw and GFccw to ensure universal global fairness and broadcast them to upstream nodes. It can be found that inter-ring flows make use of unused bandwidth.

Hence, the total inter-ring traffic throughput by IGFC with WRS is twice more than IGFC without WRS.

4.5 Unbalanced Traffic Scenario

There is a problem with AM in RPR that permanent oscillations occur with low rate downstream flows and unbalanced traffic. Figure 4.4 (a) shows node 0, 1, 2, and 3 are in Rk and node 4 and 5 are in Rk+1. In this case, we assume flow (4, 5) is a lower rate 1.2Gbps for the sake of noting the action of greedy inter-ring traffic, flow (0, 5), flow (1, 5), flow (2, 5), and flow (3, 5) under the unbalanced traffic scenario. Since node 0 and node 1 are symmetric with node 2 and node 3, the former has the same uptrend and downtrend with the latter. So, we use blue line to represent flow (0, 5) and flow (2, 5);

golden line to illustrate flow (1, 5) and flow (3, 5). The ideal global fair rate is 2.2Gbps per flow.

(a) Scenario Setup

(b) IGFC

(c) RGFC

(d) Dropping Probability at Bridge

Figure 4.4: Unbalanced Traffic Scenario. (a) Scenario setup, (b) IGFC, (c) RGFC, and (d) Dropping Probability at Bridge.

Figure 4.4 (b) and (c) illustrate throughput versus time by IGFC and RGFC, respectively. Regardless of using IGFC or RGFC leads to permanent oscillations under the unbalanced traffic scenario. This is because we apply AM as the local fairness algorithm.

The maximum amount of traffic which nodes on Rk can output is subject to global fairness but not themselves local fairness, since they are never locally congested. When node 4 is congested, it sends the local fairness control packet with LF of 1.2Gbps to the bridge.

Accordingly, the available bandwidth for all inter-ring traffic is throttled to 1.2Gbps;

moreover, node 0, 1, 2, and 3 decrease their add rate according to the adjusted GF. When the congestion at node 4 is resolved, it forwards LF of FULL_RATE to the bridge. Thus the bridge can transmit traffic as more as possible below the link capacity, and meantime upstream nodes can increase their add rate until congestion at node 4 takes place again to start another oscillation cycle. Therefore oscillations can not be eliminated.

However, it is obvious that using IGFC has moderate oscillations, but using RGFC has severe oscillations. Figure 4.4 (d) shows dropping probability at bridge in comparison with IGFC and RGFC. IGFC has immunity against buffer overflow, but RGFC has at

most 0.22 packet dropping probability. Since node 4 is periodically congested and the available inter-ring bandwidth is varied, the dropping probability by RGFC can not be eliminated and is also circulated. IGFC, composed of FGFE, provides a soft adaptive capability to avoid buffer overflow and generates an appropriate GF even under the disadvantageous scenario. It seems that the system with RGFC converges fast until the first ripple rises at 95^th round in Figure 4.4 (c). It is an illusion because there is no congestion at node 4 before 95^th round and the transmission rate of nodes on Rk is FULL_RATE. Afterwards, congestion occurs at node periodically and it accompanies violent fluctuations. Figure 4.4 (c) exhibits a weird phenomenon. Node 1 and node 3 which are close to the bridge always oscillate upon the ideal GF after a period; on the contrary, node 0 and node 2 fluctuates below the ideal GF. This is because the number of dropping packets of node 0 and node 2 are larger than node 1’s and node 3’s. When flow (0, 5) and flow (1, 5) (or flow (2, 5) and flow (3, 5)) nearly converge, it means node 4 is not congested.

Chapter 5

Conclusions

In this thesis, we emphasize the importance of global fairness and buffer overflow prevention in a bridged RPR network (BRPR). The current local fairness algorithms can not support global fairness. Design of global fairness controller for BRPR is the major concern in this dissertation. We introduce the global fairness criteria, which are inherited from RIAS local fairness reference model, to ensure the equal share of each inter-ring ingress aggregated (IIA) flow from the available inter-ring traffic bandwidth. Therefore the intelligent global fairness controller (IGFC) is accomplished to realize global fairness for inter-ring traffic, to maintain local fairness for intra-ring traffic, and to prevent from buffer overflow.

There are dual ingress buffers with the dynamic weighted round robin scheduling.

These designs can help reduce the drawback of FIFO discipline only with a single ingress buffer, accommodating inter-ring traffic from CW and CCW, and efficiently serve inter-ring traffic corresponding to the global fairness criteria. IGFC has a local ringlet buffer to contain intra-ring traffic from upstream local nodes and so does a local node. In order to be consistent with local stations, local fair rate at the IGFC is calculated from the forward rates of two ingress buffers.

There are a pre fair rate generator (pFRG) and a fuzzy global fair rate estimator (FGFE). The pFRG is implemented by the global fairness criteria. FGFE estimates two global fair rates (GF) for CW and CCW based on the pre fair rate for next agingInterval.

GF informs upstream nodes to adjust their add rate. With the aid of fuzzy control, IGFC

would behave more aggressive when the buffer occupancy is light, and more conservative when the buffer occupancy is becoming heavy to promote the system stability and to avoid buffer overflow. We also propose a weighted ringlet selector (WRS) according to the traffic load and the hop counts to probably utilize the unused bandwidth of the other ringlet. In other words, each IIA flow may choose a suitable path but not only along the shortest path.

Simulations are performed in various BRPR topology and traffic patterns to measure the proposed IGFC. Simulation results demonstrate that both IGFC and RGFC achieve the performance objectives of BRPR, but RGFC fails to have the immunity against buffer overflow. IGFC has the better performance than RGFC not merely on the convergence time part but also on the oscillation part. Since there is usually packet loss by using RGFC especially in a large topology BRPR network, it is hard to hold the global fairness. Thus the convergence time would be delayed and the margin of oscillations would be enlarged.

The system with IGFC has a defect that it can not converge perfectly by a narrow margin of oscillations under the unbalanced traffic scenario due to the property of AM local fairness algorithm. As a whole, IGFC accomplishes the objectives of BRPR efficiently, and can fight against the influence of propagation delay at the same time.

Bibliography

[1] IEEE Standard 802.17, “Resilient Packet Ring (RPR) Access Method and Physical Layer Specification,” 2004

[2] F. Davik, M. Yilmaz, S. Gjessing, N. Uzun, “IEEE 802.17 Resilient Packet Ring Tutorial,” IEEE Communication Magazine, vol. 42, no. 3, pp. 112 – 118, March 2004.

[2] F. Davik, M. Yilmaz, S. Gjessing, N. Uzun, “IEEE 802.17 Resilient Packet Ring Tutorial,” IEEE Communication Magazine, vol. 42, no. 3, pp. 112 – 118, March 2004.