inwe

Techn

Article history:

deliberated by the IEEE LAN/MAN Standards Committee

the RPR protocol provides a fairness mechanism amongstations in terms of sharing the overall ring bandwidthapplied to the best effort trafc.

ready deployed by service providers in MANs or WANs.T rings consist off the rings is usedd duringse of a faith allocati

network monitoring requirements increase the totof a SONET network. While plain Ethernet does notstatic allocation and provides cost advantages, itprovide desired features such as fairness and auto-restora-tion [2]. In addition, the RPR standard is a media indepen-dent standard which can be used on top of SONET orEthernet network to obtain the above advantages.

1389-1286/$ - see front matter 2014 Elsevier B.V. All rights reserved.

Corresponding author. Tel.: +1 4088530767.E-mail addresses: myilmaz@cisco.com (M. Yilmaz), nirwan.ansari@

njit.edu (N. Ansari).

Computer Networks 62 (2014) 4354

Contents lists available at ScienceDirect

Computer N

journal homepage: www.elshttp://dx.doi.org/10.1016/j.bjp.2014.01.003[1]. The ring is established by using point-to-point bidirec-tional connections between stations in RPR. The RPR net-work has two counter-rotating rings called ringlets. TheRPR protocol utilizes both ringlets to overcome connectiondisturbances, and hence provides resilience. In addition,

However, in these networks, many SONEa dual-ring conguration in which one oas the back-up ring and remains unuseoperation to be utilized only in the cathe primary ring [2]. The static bandwidnormallure ofon andal costrequirecannot1. Introduction

The IEEE 802.17 Resilient Packet Ring (RPR) is a stan-dardized network architecture based on ring topology

The ring network is simpler to operate and administerthan a complex mesh or an irregular network whilevirtually providing connection from each station to everyother station [2]. SONET based ring networks have been al-Received 29 July 2012Received in revised form 16 December 2013Accepted 6 January 2014Available online 10 January 2014

Keywords:MANWANRing networksSpatial reuseFairnessActive queue managementThe IEEE 802.17 is a standardized ring topology network architecture, called the ResilientPacket Ring (RPR), to be used mainly in metropolitan and wide area networks. This paperintroduces destination differentiation in ingress aggregated fairness for RPR and focuses onthe RPR MAC client implementation of the IEEE 802.17 RPR MAC in the aggressive mode ofoperation. It also introduces an enhanced active queue management scheme for ring net-works that achieves destination differentiation as well as higher overall utilization of thering bandwidth with simpler and less expensive implementation than the generic imple-mentation provided in the standard. The enhanced scheme introduced in this paper pro-vides performance comparable to the per destination queuing implementation, which isthe best achievable performance, while providing weighted destination based fairness aswell as weighted ingress aggregated fairness. In addition, the proposed scheme has beendemonstrated via extensive simulations to provide improved stability and fairness withrespect to different packet arrival rates as compared to earlier algorithms.

2014 Elsevier B.V. All rights reserved.a r t i c l e i n f o a b s t r a c tAchieving destination differentiationfairness for resilient packet rings bybased fair dropping

Mete Yilmaz a,, Nirwan Ansari ba Edge Routing Business Unit, Cisco, CA 95134, USAbDepartment of Electrical and Computer Engineering, New Jersey Institute ofingress aggregatedighted destination

ology, NJ 07102, USA

etworks

evier .com/ locate/comnet

weighted fairness and its use cases in [9,10]. The basic

An RPR network can have a maximum of 255 stations.Each station on the ring removes a frame from the ringletif that frame is destined for that station (except multicastframes, which are removed by the sender). This facilitatesthe spatial reuse property of an RPR network since thebandwidth will not be consumed by that frame aroundthe ringlet.

Fig. 1 shows an example scenario. Stations 4 and 3 aretransmitting to Station 2 on the inner ringlet, while Sta-tions 2 and 1 are transmitting to Station 7. In addition Sta-tion 2 has trafc destined to Station 1. Each time a frame isreceived at Station 2 from the ring, the frame will bechecked to see if that frame is destined to the station itselfor not. The frame will be removed from the ring if it is des-

44 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354algorithm relies on identifying a single congestion pointon the ring. This decreases the ring utilization and spatialreuse in certain trafc scenarios. The standard providesmeans to keep track of all the congested downstreamnodes and their fair rates. A generic way to utilize thismechanism fully is to implement 255 separate queues(one queue per destination, including the multicast queue)within the MAC client. Supporting multiple destinationqueues at the MAC layer is not efcient and requires addi-tional memory or complex shared memory queuing imple-mentations. In addition, this may not be even possible toimplement on the already deployed hardware. This paperpresents a simpler and less expensive MAC client imple-mentation that does not require separate queues whileproviding ring utilization comparable to the generic imple-mentation. Our previous work [11] showed that it is possi-ble to utilize an active queue management scheme at theMAC client level. However, the implementation that willbe presented in this article achieves destination differenti-ation at the source stations by allowing different weightsper destination while achieving RPR ingress aggregatedfairness. In addition, the proposed implementation pro-vides a higher degree of stability under different packet ar-rival rates and patterns. This is achieved by penalizing theows which exceed their fair shares more than the others.

The rest of this paper is organized as follows. In Section2, the RPR operation and fairness algorithm will be dis-cussed with emphasis on MAC client implementations.Section 3 will discuss the MAC client implementation uti-lizing our proposed active queue management schemebased on the approximate fair dropping algorithm [12].In Section 4, behavior of the current (aggressive) fairnessalgorithm with MAC client implementations of singlequeue, multiple queues, and two other algorithms withactive queue management of a single queue will be illus-trated through simulations. Section 5 provides simulationresults for the modied scenario that provides destinationdifferentiation. In Section 6, comparison of a large ring hubscenario will be provided. Finally, the conclusion will bedrawn in Section 7.

In order to demonstrate different operational modes,some performance gures of merits are included and dis-cussed. The scenarios have been executed on the RPR sim-ulator model developed at Cisco during the RPRstandardization process. The simulation model is imple-mented in Opnet. The suggested modication has beenincorporated into the simulator model and its behavior isveried through simulations.

2. RPR operation and fairness

The operation of the RPR protocol has been discussed indetail in [2]. In this paper, a brief overview based on anexample will be provided.The fairness aspects of RPR have been investigated indepth in light of interesting scenarios as those describedin [35]. Improvements for the current fairness algorithmof the IEEE 802.17 have also been proposed as reportedin [3,4,68]. In addition, in our earlier work we introducedtined to the station (in this case, any frame from Station 3or 4). In parallel, Station 1 will receive frames from Station2, and Station 7will be able to receive the frames from Sta-tions 1 and 2 without being impacted by the trafc gener-ated at Stations 3 and 4.

The objective of the fairness algorithm is to distributethe unallocated bandwidth around the ring among stationsin a fair manner. In the case of Fig. 1 (assuming that all thestations have equal weights), Stations 3 and 4 will get anequal amount of the link bandwidth on the link betweenStations 2 and 3, while Stations 1 and 2 will get an equalamount of the link bandwidth on the link between Stations1 and 7. More detailed explanation of the standardized fair-ness algorithms can be found in [1,2].

In short, the fairness algorithm provides a fair sharing ofthe link bandwidth according to the weights of the stationswhen there is more trafc than that can be transmittedthrough that link. RPR fairness is based on ingress aggrega-tion. This fairness is referred to as Ring Ingress Aggregatedwith Spatial Reuse (RIAS) fairness in an earlier article [3].This denition follows the same methodology used in [13]for maxmin ow control. The RIAS fairness denition,however, does not include the station weights in the gen-eralized formula while this is included in the IEEE 802.17standard in the calculation of the estimated fair rate of anode. In addition, the RIAS denition assumes equal shar-ing of the bandwidth among ows originating from thesame station, while the standard does not require that. Inour previous works [9,10], we have expanded this deni-tion with the inclusion of source station weights. In thissection, a more general denition will be provided withthe inclusion of the destination station weights along with

Fig. 1. Destination stripping and spatial reuse illustrated on the inner ring(ringlet 1).

both destination ows and ingress aggregation,respectively.

For the scenario given in Fig. 1, if Station 4 is weighedtwo times that of Station 3, it will get two times morebandwidth out of the ring than Station 3. In this case, if Sta-tion 3 increases its share, (9) will not be satised. Destina-tion differentiation can be provided at Station 2 if thedestination weights (qst21 and q27) are adjusted so thatthe destination Station 1 has two times more weight thanStation 7. Then the Station 1 will receive two times moretrafc from Station 2 compared to Station 7 in order to sat-isfy Eq. (8).

The standard denes two mechanisms to distribute thefairness information around the ring. The rst mechanismis used to distribute the fair rate of the nearest congestedstation to the upstream stations. This fairness message iscalled the single-choke fairness frame (SCFF) in the stan-

M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354 45the source station weights to provide a more comprehen-sive representation of fairness as in the IEEE 802.17standard.

Denote N as the total number of stations on a ringlet.Let the capacity of each link on the ringlet be C. EachStation s on the ringlet is given a weight ws for providingthe weighted fairness. On this ringlet, a ow vector is de-ned by F = {fst}, in which each ow from Station s to Sta-tion t is denoted by fst which is also referred to as the pathof the ow. For each ow a weight vector is dened byP = {qst}, in which the weight of each ow from Station sto Station t will be denoted by qst. A fair rate vector is de-ned by R = {rst} in which the fair rate of ow fst is denotedby rst. By using the above denitions, the total allocatedrate on link n of the ringlet is given by (1).

Tn X

8s;t: link n2fstrst 1

On this ringlet, the vector R is said to be feasible if the fol-lowing conditions in (2) and (3) are met.

rst > 0 8s; t : f st 2 F 2

Tn 6 C 8n 2 N : 0 < n 6 N 3The sum of all ows originating from Station s and passingthrough link n is

Ans X

8t2N: link n2fstrst 4

For a feasible vector R, the link n is a bottleneck link,Bn(s, t), with respect to R for fst crossing link n if the follow-ing conditions in (5)(7) are met with respect to all owsfs0t0 crossing link n.

Tn C 5

rs0t0 6 rst 8s0; t0 : s s0 & t0 t & link n 2 fs0t0 6

Ans0 6 Ans 8s0; t0 : s0 s & link n 2 fs0t0 7Note that if there are no other ows originated from anystation other than Station s going through link n, An(s0) willbe zero and (7) will be satised by default.

The vector R is said to be weighted ingress aggregatedfair with destination differentiation if it is feasible asdened in (2) and (3) and if for each fst, rst cannot be in-creased while maintaining feasibility without decreasingthe fair rate rs0t0 of some ow fs0t0 for which

rs0t0qs0t0

6 rstqst

8s0; t0 : s0 s & f s0t0 2 F 8

Ans0ws0

6 Answs

8s0; t0;n; : s0 s & f s0t0 2 F &linkn 2 fs0t0 & linkn 2 fst

9

Eq. (8) ensures the fairness among the ows originatingfrom the same station with destination differentiation,while (9) ensures the fairness among ingress aggregatedows. The weights q and w are used to normalize the com-parisons, and hence to achieve the weighted fairness fordard [2]. The second mechanism is used to propagate thefair rate of each station to all the other stations. This fair-ness message in RPR is called the multi-choke fairnessframe (MCFF) [2]. Each station on the ring broadcastshow much of its output link is used by the station itselfin its MCFF and sends it to all the stations on the ring. Areceiving station may collect these messages and build aglobal view of the congestion situation on the ring, andschedule the trafc to add to the ring accordingly. Asshown in our previous article [16], this information caneven be used to support non-uniform capacity links in anRPR network efciently.

Fig. 2 shows the separation of RPR MAC and its client.RPRMAC transfers MCFF and SCFF messages via the controlpath indication, while the MAC client transmits andreceives the packets via data path request and indicationmessages, respectively. MAC Datapath Sublayer handlesthe transmission and reception of the frames to and fromthe dual ringlets. The implementation details of an RPRMAC client layer is not part of the standard; however, inAppendix J of the standard [1], some examples are pro-vided for a single queue implementation that utilizes SCFFand another that utilizes MCFF with virtual output queues.By shaping trafc according to the MCFF messages at theMAC client, one can increase the bandwidth utilizationby avoiding single congestion points. In Section 3, an RPR

Fig. 2. RPR MAC services model.

queue instead of waiting for tail drops based on queue

standard). Thus, the hardware implementation is simpli-ed and allows microcode based implementations on pres-ently deployed hardware.

For the wDBFD algorithm, consider a ring with thesource Station s and the destination Station d as shown

46 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354occupancy and RPR fairness messages. Furthermore, byadding weights to the drop probabilities among differentdestination stations, destination differentiation isachieved. This algorithm is more stable in terms of fairnesswhen subject to different packet arrival rates than thealgorithm proposed in [11] as shown in Section 4.

3. Weighted destination based fair dropping for RPR

Similar to the other active queue management tech-niques (e.g., RED [14]), weighted DBFD also utilizes thequeue size to accept or drop packets on arrival. However,wDBFD not only relies on past measurements of the queuesize but also on recent observed rates of ows. By usingthis additional information wDBFD can provide fairnessamong different ows [15].

Approximation of virtual output queuing using singleFIFO queue with active queue management is shown withthe simulation results in Section 4. The implementation ofwDBFD in RPR requires a modication of the AFD algo-rithm since in RPR the received fair rate from a congestedstation changes the packet drop probabilities of all owsdestined to stations downstream to that congested station(excluding the congested station itself) once the aggre-gated rate of ows exceeds the received fair rate. There-fore, the MAC client needs to actively adjust the dropprobability of each packet to each destination by consider-ing the received fair rates from congested stations. Whileproviding fairness among destinations, the MAC client willnot need to implement 255 destination queues with thewDBFD algorithm. Instead, this scheme utilizes per desti-nation counters in the MAC client (most of which are al-ready necessary for a multi-queue implementation of theMAC client implementation that utilizes a modiedApproximate Fair Dropping (AFD) algorithm [12] will beshown.

In a simple RPR MAC client, a single low priority queuehandles the packets as they are waiting to enter the ring.After a station sends its fair share of trafc to the network,the RPR MAC will not allow any more packets to be sentand this will cause head of line (HOL) blocking. The RPRMAC client can also support separate queues per destina-tion for low priority packets to eliminate HOL blocking.This multi-queue implementation of an RPR MAC clientkeeps track of SCFF and MCFF messages, and schedulespackets in a decit round robin fashion as long as theyare allowed by the RPR MAC. During this scheduling, theimportant point to consider is to avoid sending more pack-ets beyond the congested nodes than allowed. Otherwise,the RPR MAC will not allow transmission of any packetfor a period of time. However, implementing multiplequeues at an RPR MAC client is costly. This article proposesan RPR MAC client with an active queue managementscheme and will be referred to as weighted DestinationBased Fair Dropping (wDBFD). Similar to a multi-queueRPR MAC client, this MAC client also needs to keep trackof SCFF and MCFF messages. Instead of implementing mul-tiple queues, only a single queue is needed, and packets areprobabilistically dropped before they are accepted into thein Fig. 3. On this ring, assume Station i as the station thathas the minimum fair rate in between Station s and Stationd. Also dene Station j as any arbitrary downstream stationbeyond Station s.

On this ringlet, Us is denoted as the received fair ratevector at Station s, where Us is the set of fair rates obtainedfrom SCFF or MCFF sent by all the downstream stations inbetween Station s and Station d at time t. If ui is the mini-mum of all fair rates received, then Station i is the mostcongested station in between the source Station s and thedestination Station d.

ui minUs 10Dene a ow vector, F = {fsd}, in which the arrival rate of aow to Station s destined to Station d is denoted by fsd. De-ne another ow vector, R = {rsj}, in which a ow sourcedby Station s destined to Station j is denoted by rsj. By usingthe above denitions, the total trafc sourced by Station sthat is destined beyond Station i is given by (11).

ksi X

8j: Station j>Station irsj 11

Dene Qtarget, a1, a2 as arbitrary constants, dbfd0s as the pre-

vious value of the wDBFD rate, qs as the length of the queueat Station s in bytes, and q0s as the previous value of thequeue length, then the current wDBFD rate at Station s isgiven by (12). Note that the details of adjusting parametersa1 and a2 can be found in [15].

dbfds Q target qs 0dbfd0s a1qs Q target a2q0s Q target qs 0

(

12Dene b as an arbitrary constant to normalize the wDBFDrate with respect to the rate measurement, which is equalto LINK_RATE [1] divided by Qtarget. Dene fsd as the arrivalrate of a ow at Station s destined to Station d, then thedrop probability of a packet destined to Station dwill be gi-ven by (13).

psd1 ws ui< ksi1 qsd bdbfdsfsd ws uiP ksi and fsdPqsd b dbfds0 ws uiP ksi and fsd:

13As dened in (13), if the received fair rate (ui) is less thanthe total trafc sourced by Station s destined beyond Sta-tion i, then the trafc destined beyond Station i will bedropped. If the Station s is still allowed to send more trafcbeyond Station i, then based on the wDBFD rate and the ar-rival rate (fsd) the trafc may be randomly dropped. The

Fig. 3. Multi destination scenario.

drop probability goes up if the arrival rate of a ow is muchhigher than the wDBFD rate. If a ow continues to send atthe higher rate, it will be penalized more by increasing thedrop probability. This allows the algorithm to be more sta-ble with respect to different packet arrival rates. If thearrival rate of trafc is higher than the departure rate fromthe queue, the algorithm will operate at buffer occupancy

adding an additional level to the scheduling hierarchy isnot possible without requiring new hardware.

While the implementation of wDBFD algorithm doesnot require 255 queues, it requires additional calculationstages on top of the current RPR fairness algorithm. Specif-ically, the station will need to keep track of fairness mes-sages from all other stations in a 255 entry array. Inaddition, at each RPR parameter calculation interval calleddecay interval, the dbfd_fair rate needs to be decayedand smoothed out. The highest computational complexitycomes from the case when a queuing decision needs tobe made at each packet arrival. One can employ the sam-pling algorithm as proposed in [12] to allow rate estima-tions at certain intervals so as not to burden the systemwith calculating the rates at each packet arrival.

4. Performance evaluation of weighted destinationbased dropping

The example scenario shown in Fig. 4 will be used to

Fig. 4. Multi destination scenario.

M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354 47of Qtarget. The algorithm will establish a certain packetmix in the buffer so that the ratio of packets destined toeach station will correspond to the ratio of allowed fairrates to each destination. In addition, the drop probabilitiesare also adjusted by weight qsd to allow different destina-tions receive different proportions of the availablebandwidth.

The current single queue implementation of RPR mayresult in severely underutilized rings in some scenarios,and oscillations in utilization are also observed in thosescenarios [3]. With the proposed mechanism, the ring uti-lization can easily approach the theoretical limit of thenumber of links multiplied by the bandwidth with a rela-tively simple implementation. The advantage of this algo-rithm for RPR is that it improves the performance of anRPR ring, and it is backward compatible with the standardwhich is not true for the previously proposed solutions. Inaddition, it does not require 255 independent queues to beimplemented in the scheduling hierarchy. In generalFig. 5. Actual trafc sourced at Stations 2, 3, 4 and 5 with single MAC client queue.compare the performance of different MAC client imple-mentations. In this scenario, Stations 2, 3, 4 and 5 have traf-c destined to Station 1. Station 5 has also trafc destinedto Stations 3 and 4. Stations 2, 3, and 4 start sending trafcto Station 1 at time 1 s. Station 5 starts sending trafc toStations 1 and 4 at time 1 s. Station 5 then starts sendingtrafc to Station 3 at time 2 s. In this scenario, the trafcdemand of all but one ow at each node is OC12 rate peractiveow from one station to another. Station 5 receivestwo times more trafc to destination Station 3 than theother stations in order to demonstrate the stability of thewDBFD algorithm.

Per RIAS fairness [3], the 620 Mbps bandwidth on thelink between Stations 1 and 2 should be equally sharedamong four stations resulting in 155 Mbps per station. Sta-tion 5 can utilize more bandwidth without impacting thisfairness by utilizing the unused 465 Mbps bandwidth onlink between Stations 5 and 4 and also a maximum avail-able bandwidth of 310 Mbps on link between Stations 4

Fig. 6. Trafc received at Stations 1, 3 and 4 with single MAC client queue.

Fig. 7. Actual trafc sourced at Stations 2, 3, 4 and 5 with VoQ.

Fig. 8. Trafc received at Stations 1, 3 and 4 with VoQ.

48 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354

M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354 49and 3. In the ideal case, Stations 1, 3 and 4 should receive620 Mbps, 232.5 Mbps, and 232.5 Mbps, respectively.Therefore, the total bandwidth utilization on the ring willbe 1085 Mbps based on be weighted ingress aggregatedfair with destination differentiation denition. This sce-nario will be used to compare the performance of differentMAC client implementations.

The scenario is simulated using the single queue RPRmodel implemented in the Opnet simulator. An OC12 ringwhich is composed of 9 stations is created with 20 km ofdistance between every two adjacent stations. Each stationis congured as a dual-queue station with the aggressivefairness mode enabled. The size of the secondary transitqueue (STQ) at each station is 512 KB and the LP_COEFF[1] parameter of the RPR MAC is set to 4.

Fig. 9. Actual trafc sourced at Stat

Fig. 10. Trafc received at StatioFig. 5 shows the total trafc sourced by Stations 2, 3, 4and 5 to the outer ringlet. As expected, the available band-width is being shared equally by Stations 2, 3, and 4, whileStation 5 is able to get more bandwidth out of the ring byutilizing the unused bandwidth on the links. However, attime 2 s, once Station 5 starts sending trafc to Station 3,the fairness message generated by Station 4 limits the totaltrafc that can be sourced by Station 5.

As shown in Fig. 6, Station 1 receives the full 620 Mbpsof trafc, while as shown in Fig. 5 Station 5 is not able toutilize the unused bandwidth fully. In addition, once Sta-tion 5 starts sending packets to Station 3 at time 2 s, thefairness is lost and Station 5 is not able to get its fair shareof the bandwidth, and Stations 2, 3 and 4 start sourcingmore trafc to Station 1 than Station 5. The total ring utili-

ions 2, 3, 4 and 5 with DBFD.

ns 1, 3 and 4 with DBFD.

50 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354zation is 810 Mbps instead of the expected 1085 Mbps.Oscillations observed around the steady state behaviorshow less stability.

One other observation is that after time 2 s, Station 3 re-ceives more trafc than Station 4 even though the trafc issourced by the same Station 5. The main reason for thisbehavior is the imbalanced trafc demand used in this sce-nario and the simple single queue implementation notbeing able to maintain fairness at the source station perdestination. Therefore, destination fairness is not achieved.

The same scenario is simulated with the RPR MAC clientmodel using virtual output queues (VoQ) as explained inthe standard. Fig. 7 shows the actual trafc sourced at sta-tions 2, 3, 4, and 5. In this case, the oscillations are mini-mized, and a steady response is observed at time 2 s,when Station 5 starts sending trafc to Station 3.

Fig. 8 shows the trafc received at Stations 2, 3, and 4,respectively. The observed bandwidth matches the ex-

Fig. 11. Actual trafc sourced at Stat

Fig. 12. Trafc received at Statiopected values, and provides a maximum bandwidth utili-zation of 1085 Mbps. In addition, the destination fairnessis achieved with respect to trafc received at Station 3and Station 4, since virtual output queuing is able to main-tain destination separation with respect to different packetarrival rates for each destination.

Next, the same scenario is simulated using RPR MAC cli-ent with DBFD as explained in [11]. Figs. 9 and 10 show theactual trafc sourced and received from and at respectivestations. The steady response is observed after time 2 swhen Station 5 starts sending trafc to Station 3 and thetotal ring utilization reaches up to the expected1085 Mbps. Similar to the single queue RPR implementa-tion, the destination fairness is not achieved for the trafcsourced by Station 5 to the destination Stations 3 and 4.Specically, Station 3 receives almost two times more traf-c than Station 4. Since the packet arrival rate destined toStation 3 after time 2 s is two times more than the packet

ions 2, 3, 4 and 5 with wDBFD.

ns 1, 3 and 4 with wDBFD.

M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354 51arrival rate destined to Station 4 and the arrival rate is notregulated per destination, this undesirable behavior is ex-pected for the DBFD algorithm as well.

Next, the same scenario is simulated with the wDBFDalgorithm as explained in Section 3. The results are shownin Figs. 11 and 12.

The results shown in Figs. 11 and 12 are very similar tothe ideal case with virtual output queues. Destination fair-ness at Station 5 is maintained regardless of its packet ar-rival rate difference from destination Stations 3 and 4. Thisis made possible by increasing the probability of packetdrop per ow according to the ratio of the arrival rate ofa ow to the DBFD rate. This provides stability and fairsharing of the queues even when the arrival rates and/orpacket sizes are different among ows.

Another aspect to consider is the convergence speed.The interval required for the ring trafc to converge to

Fig. 13. MAC client queue occupancy of

Fig. 14. Actual trafc sourced at Statthe fair rate is also impacted by the desired queue occu-pancy (Qtarget). Based on the RPR fairness messages andthe destination weights, the MAC client queue will havethe right mix of packets to match the fair rates to each des-tination. As the desired queue occupancy increases, theconvergence to fair rates will take longer. If the desiredqueue occupancy (Qtarget) is set too low, unnecessary pack-et drops can be observed. This scenario has been testedwith different packet arrival patterns per destination.Regardless of the packet arrival rates and packet sizes,the destination stations receive a similar amount of trafcas shown in Figs. 11 and 12.

For this case, the transmit buffer occupancy of Stations2, 3, 4 and 5 are shown in Fig. 13. In this scenario, the de-sired queue occupancy (Qtarget) is set to 150000 bytes. Asshown in Fig. 13, the MAC client queues converge to thedesired queue occupancy after a short transition phase.

Stations 2, 3, 4 and 5 with wDBFD.

ions 2, 3, 4 and 5 with wDBFD.

Statio

52 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354Fig. 15. Trafc received at5. Providing destination differentiation through wDBFD

When desired, the proportion of trafc destined fromstation s to station d can be adjusted by the qsd parameteras described in Section 3 with the wDBFD algorithm pre-sented in this article. This adjustment relies on adjusting

Fig. 16. Large ring hub scenario.

Fig. 17. Actual trafc sourced at Stations 1, 2, 3the drop probabilities of ows per destination with respectto each other as well as buffer occupancy. The qsd parame-ters will dictate the ratio of packets in the MAC clientqueue destined to different stations. In this section, thescenario is modied such that the destination Station 4 isgiven a weight of 2 (q54 = 2) while the ows destined toStations 1 and 3 are each assigned to weight of 1.

Figs. 14 and 15 show the simulation results with thedestination weight adjustment (q54 = 2) utilizing thewDBFD algorithm. In this case, Station 5 can send twotimes more trafc to Station 4 than the other destinations.Specically, as shown in Fig. 15, the destination Station 5 isable to transmit approximately 310 Mbps of trafc to Sta-tion 4, while it transmits 155 Mbps of trafc to the Stations1 and 3. This shows that the wDBFD algorithm can ef-ciently provide destination differentiation as required evenwhen the amount of trafc destined to Station 3 is muchhigher than the amount of trafc destined to Station 4.

ns 1, 3 and 4 with wDBFD.Note that while this is a static weight allocation, thewDBFD algorithm does not restrict any bandwidth when

, 50 and 60 with single MAC client queue.

modications to the standardized IEEE 802.17 RPR fairnessmechanism and allow a simpler and computationally less

Statio

M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354 53there is no bottleneck and provides fair sharing of the ringbandwidth among destination nodes from one station.

6. Hub scenario in large RING

In this section, performance results will be providedwith single MAC queue and DBFD algorithm for a largeRPR ring in a hub scenario as shown in Fig. 16.

This scenario requires the control algorithm of RPR torun over the large ring with the ring being composed of128 stations. The weight (w) of each station is assignedto 1. Each of the stations from Station 1 to 60 has25.6 Mbps of trafc destined to Station 128 and the qsd ofeach ow is set to 1. On an OC12 ring, each station shouldbe able to transmit 10 Mbps of the trafc destined to Sta-tion 128 when bandwidth is shared equally on the ring.

Figs. 17 and 18 show the results for trafc sourced byeach station from Station 1 to 60 with single MAC clientqueue and DBFD algorithms, respectively. Both of theimplementations are able to converge to the desiredthroughput of 10 Mbps. The details of the parameters thatcan affect convergence were studied earlier in [5,10]. Dur-ing the initial start of the trafc the Station 1 is able totransmit up 25 Mbps, which is close to the maximum traf-

Fig. 18. Actual trafc sourced atc sourced by Station 1 to Station 128. However, the DBFDalgorithm shows a quicker convergence enabled by the ac-tive packet dropping mechanism at the MAC client levelbased on the system load instead of relying solely on theRPR MAC algorithm convergence.

7. Conclusion

In this paper, we have introduced a new fairnessscheme that works with the ingress aggregated RPR fair-ness by including destination differentiation at the sourcestation. In addition, this paper has discussed the proposedimplementation of an efcient active queue managementmechanism to utilize multi-choke fairness in an RPR net-work while providing weighted destination based fairness.We also showed the stability of this algorithm in a verylarge ring as well in Section 6. This algorithm and the prin-ciples can be applied to other ring networks to provide fairintensive implementation than the generic multi queueimplementation discussed in the standard.

References

[1] IEEE Standard 802.17-2004, Resilient Packet Ring.[2] F. Davik, M. Yilmaz, S. Gjessing, N. Uzun, IEEE 802.17 resilient packet

ring tutorial, IEEE Commun. Mag. 42 (3) (2004) 112118.[3] V. Gambiroza, P. Yuan, L. Balzano, Y. Liu, S. Sheafor, E. Knightly,

Design, analysis, and implementation of DVSR: a fair high-performance protocol for packet rings, IEEE/ACM Trans. Netw. 12(1) (2004) 85102.

[4] F. Davik, A. Kvalbein, S. Gjessing, Performance evaluation andimprovement of non-stable resilient packet ring behavior, in: Proc.distribution of bandwidth across the ring as long as thefairness information is distributed among stations. As com-pared to our earlier implementation, this algorithm pro-vides better isolation of ows with respect to differentarrival rates. As shown in this paper, while preserving fair-ness among stations, this approach has improved the utili-zation of the underlying network as compared to the singlequeue implementation of the standard. In addition, themechanisms discussed in this article do not require any

ns 1, 2, 3, 50 and 60 with DBFD.Part II of the 4th International Conference on Networking (ICN05),Ser. LNCS 3421, Reunion Island, April 1721 2005, pp. 551563.

[5] F. Davik, A. Kvalbein, S. Gjessing, An analytical bound forconvergence of the resilient packet ring aggressive mode fairnessalgorithm, in: Proc. 40th Annual IEEE International Conference onCommunications (ICC05), Seoul, Korea, May 1620 2005.

[6] F. Alharbi, N. Ansari, Distributed bandwidth allocation for resilientpacket ring networks, Comput. Netw. 49 (2) (2005) 161171..

[7] F. Alharbi, N. Ansari, SSA: simple scheduling algorithm for resilientpacket ring networks, IEE Proc. Commun. 153 (2) (2006) 183188.

[8] W. Tang, C. Chang, A Fuzzy inter-ring route control with PRNNpredictor bridged resilient packet rings, in: Proc of IEEE Globecom,December 2010.

[9] M. Yilmaz, N. Ansari, Weighted fairness in resilient packet rings, in:Proc. of the 2007 IEEE International Conference on Communications(ICC07), Glasgow, June 2428 2007, pp. 21922197.

[10] M. Yilmaz, N. Ansari, Weighted fairness and correct sizing ofsecondary transit queue in resilient packet rings, J. Opt. Commun.Netw. 2 (11) (2010) 944951.

[11] M. Yilmaz, N. Ansari, J.H. Kao, P. Yilmaz, Active queue managementfor MAC client implementation of resilient packet rings, in: Proc. ofthe International Conference on Communication (ICC09), Dresden,Germany, June 1418 2009, pp. 15.

[12] R. Pan, L. Breslau, B. Prabhakar, S. Shenker, Approximate fairnessthrough differential dropping, SIGCOMM Comput. Commun. 33 (2)(2003) 2339.

[13] D. Bertsekas, R. Gallager, Data Networks, Prentice-Hall, 1987.[14] S. Floyd, V. Jacobson, Random early detection gateways for

congestion avoidance, IEEE/ACM Trans. Netw. 1 (4) (1993) 397413.[15] C. Hollot, V. Misra, D. Towsley, W. Gong, On designing improved

controllers for AQM routers supporting TCP ows, in: Proc. of IEEEINFOCOM, April 2001.

[16] M. Yilmaz, N. Ansari, Resilient packet rings with heterogeneouslinks, in: Proc. of the 2012 IEEE Symposium on Computers andCommunications (ISCC12), Cappadocia, July 14 2012.

Mete Yilmaz received his B.S. (1995) and M.S.(1999) degrees from Bogazici University,Istanbul, Turkey, in electrical engineering andcomputer engineering, respectively. Hereceived his Ph.D. in computer engineeringfrom New Jersey Institute of Technology,Newark, in 2013. He works at Cisco Systems,San Jose, California as a Technical Leader withresponsibilities ranging from architecting newline cards to developing packet processingFPGAs for Ciscos high-end routers. He hasbeen working on communication networks

for fteen years. His research interests include quality of service, band-width reservation, and fairness algorithms. He holds three patents onfairness algorithms.

communications (Springer, 1994) with B. Yuhas. His current researchfocuses on various aspects of broadband networks and multimediacommunications. He has also contributed over 350 technical papers, overone third of which in widely cited refereed journals/magazines. Forexample, one of his seminal works was the sixth most cited article pub-lished in the IEEE Transactions on Parallel and Distributed Systems. Hewas/is serving on the Advisory Board and Editorial Board of eight journals,including as a Senior Technical Editor of IEEE Communications Magazine(20062009). He had/has been serving the IEEE in various capacities suchas Chair of IEEE North Jersey COMSOC Chapter, Chair of IEEE North JerseySection, Member of IEEE Region 1 Board of Governors, Chair of IEEECOMSOC Networking TC Cluster, Chair of IEEE COMSOC Technical Com-mittee on Ad Hoc and Sensor Networks, and Chair/TPC Chair of severalconferences/symposia. He has been frequently invited to deliver keynoteaddresses, distinguished lectures, tutorials, and talks. Some of his recentawards and recognitions include an IEEE Fellow (Communications Soci-ety), IEEE Leadership Award (2007, from Central Jersey/Princeton Sec-tion), the NJIT Excellence in Teaching in Outstanding ProfessionalDevelopment (2008), IEEE MGA Leadership Award (2008), the NCEExcellence in Teaching Award (2009), and designation as an IEEE Com-munications Society Distinguished Lecturer (20062009, two terms).

54 M. Yilmaz, N. Ansari / Computer Networks 62 (2014) 4354Nirwan Ansari received the B.S.E.E. (summacum laude, gpa = 4.0/4.0) from the New JerseyInstitute of Technology (NJIT), Newark, in1982, the M.S.E.E. degree from University ofMichigan, Ann Arbor, in 1983, and the Ph.D.degree from Purdue University, West Lafay-ette, IN, in 1988. He joined NJITs Departmentof Electrical and Computer Engineering asAssistant Professor in 1988, tenured AssociateProfessor in 1993, and Full Professor since1997. He has also assumed various adminis-trative positions at NJIT. He authored Com-

putational Intelligence for Optimization (Springer, 1997, translated intoChinese in 2000) with E.S.H. Hou, and edited Neural Networks in Tele-

Achieving destination differentiation in ingress aggregated fairness for resilient packet rings by weighted destination based fair dropping1 Introduction2 RPR operation and fairness3 Weighted destination based fair dropping for RPR4 Performance evaluation of weighted destination based dropping5 Providing destination differentiation through wDBFD6 Hub scenario in large RING7 ConclusionReferences