Ethernet PON (epon): Design and Analysis of an Optical Access Network.

Transcription

1 Ethenet PON epon: Design and Analysis of an Optical Access Netwo. Glen Kame Depatment of Compute Science Univesity of Califonia, Davis, CA 9566, USA Tel: ; Fax: Biswanath Muhejee Depatment of Compute Science Univesity of Califonia, Davis, CA 9566, USA Tel: ; Fax: Gey Pesavento Advanced Technology Lab. Alloptic, Inc. Livemoe, CA 94550, USA Tel: ; Fax: Abstact August, 2000 With the expansion of sevices offeed ove the Intenet, the last mile bottlenec poblems continue to exacebate. A Passive Optical Netwo PON is a technology viewed by many as an attactive solution to this poblem. In this study, we popose the design and analysis of a PON achitectue which has an excellent pefomance-to-cost atio. This achitectue uses the time-division multiplexing TDM appoach to delive data encapsulated in Ethenet pacets fom a collection of Optical Netwo

2 Units ONUs to a cental Optical Line Teminal OLT ove the PON access netwo. The OLT, in tun, is connected to the est of the Intenet. A simulation model is used to analyze the system s pefomance such as bounds on pacets delay and queue occupancy. Then, we discuss the possibility of impoving the bandwidth utilization by means of timeslot size adjustment, and by pacet scheduling. Keywods: access netwo, local loop, passive optical netwo, PON, time-division multiple access, TDMA, self-simila taffic. Intoduction Passive Optical Netwos PON ae point-to-multipoint optical netwos with no active elements in the signals path fom souce to destination. The only inteio elements used in such netwos ae passive combines, couples, and splittes. PON technology is getting moe and moe attention by the telecommunication industy as the last mile solution [,2]. Advantages of using a PON fo local access netwos ae numeous: A PON allows fo longe distances between cental offices and custome pemises. While with the Digital Subscibe Line DSL the maximum distance between the cental office and the custome is only 8000 feet appoximately 5.5 m, a PON local loop can opeate at distances of ove 20 m. A PON minimizes fibe deployment in both the local exchange and the local loop. A PON povides highe bandwidth due to deepe fibe penetation. While the fibe-tothe-building FTTB, fibe-to-the-home FTTH, o even fibe-to-the-pc FTTPC solutions have the ultimate goal of fibe eaching all the way to custome pemises, fibeto-the-cub FTTC may be the most economical deployment today. As a point-to-multipoint netwo, a PON allows fo downsteam video boadcasting. A PON eliminates the necessity of installing multiplexes and demultiplexes in the splitting locations, thus elieving netwo opeatos fom the guesome tas of maintaining them and poviding powe to them. Instead of active devices in these locations, a PON has passive components that can be buied into the gound at the time of deployment. 2

3 A PON allows easy upgades to highe bit ates o additional wavelengths. The clea advantages of using PON technology in access netwos dictate that we mae impotant design decisions in implementing it. Because an access netwo aggegates taffic fom a elatively small numbe of subscibes compaed to meto o egional netwos, it is vey cost sensitive. Theefoe, a PON design should not equie ove-povisioning and should allow fo incemental deployment. In this study, we popose the design and analysis of a PON achitectue that has an excellent pefomance-to-cost atio. This achitectue Section 2 uses the time-division multiplexing TDM appoach to delive data encapsulated in Ethenet pacets fom a collection of Optical Netwo Units ONUs to a cental Optical Line Teminal OLT ove the PON access netwo. The OLT, in tun, is connected to the est of the Intenet. A simulation model descibed in Section 3 is used to analyze the system s pefomance such as bounds on pacets delay and queue occupancy. The simulation analysis was pefomed using Bellcoe taces that exhibit the popety of self-similaity [3]. Self-simila o factal taffic has the same o simila degee of bustiness obseved at a wide ange of time scales. Using self-simila taffic is extemely impotant as it povides ealistic bounds on pacets delay and queue occupancy. See Section 4. We continue ou investigation of the bandwidth utilization of ou poposed model and consideed two ways to impove it. In Section 5 we conside the timeslot size adjustment to achieve the best opeating paametes utilization and delay. We employ the powe function [4] as a convenient measue of system pefomance. In Section 6 we conside an altenative appoach to impove utilization; specifically, we examine pacet scheduling. In doing so, we pay special attention to the effect of pacet eodeing on TCP/IP connection behavio. Section 7 concludes this study. 2. PON Design Altenatives Thee ae seveal topologies suitable fo the access netwo: tee, ing, o bus Fig.. A PON can also be deployed in edundant configuation as double ing o double tee; o edundancy may be added only to a pat of the PON, say the tun of the tee Fig. 2. Fo the est of this aticle, we will focus ou attention on the tee topology; howeve, most of the conclusions made ae equally elevant to othe topologies. 3

4 All tansmissions in a PON ae pefomed between Optical Line Teminal OLT and Optical Netwo Units ONU. Theefoe, in the downsteam diection fom OLT to ONUs, a PON is a point-to-multipoint netwo, and in the upsteam diection it is a multipoint-to-point netwo. The OLT esides in the local exchange cental office, connecting the optical access netwo to an IP, ATM, o SONET bacbone. The ONU is located eithe at the cub FTTC solution, o at the end-use location FTTH, FTTB solutions, and povides boadband voice, data, and video sevices. ONU ONU2 OLT ONU3 OLT a Tee topology ONU5 ONU4 ONU ONU5 ONU4 ONU2 ONU4 ONU2 OLT ONU3 c Ring topology ONU b Bus topology ONU3 ONU5 Fig.. PON topologies. 2:N ONU ONU2 OLT ONU3 ONU4 ONU5 Fig. 2. Tee with a edundant tun. 4

5 Access netwos based on PON technology face seveal design challenges, egadless of the physical topology. The fist design decision to be made is the data-lin technology. The Table summaizes the advantages and disadvantages of diffeent data-lin technologies. Data Lin Advantage Disadvantage SONET Fault toleance, fault management, lage installed base. Expensive hadwae too expensive fo the local loop. Also not efficient fo data taffic. ATM Queues in the OLT and ONUs can easily implement vaious QoS policies and guaantees poviding bette suppot fo eal-time taffic voice and video. At the custome side and at the netwo side, data has the fom of IP pacets. In ode to tavese the PON, IP pacets should be boen into cells and eassembled at the othe end. This intoduces additional complexity and cost. Ethenet Vey convenient to cay IP pacets see ATM disadvantage; ubiquitous and cheap hadwae; scalable 00 Mbps, Gbps, 0 Gbps. Requies QoS techniques fo ealtime taffic. Table. Advantages and disadvantages of data-lin technologies in PON. Anothe design challenge is the sepaation of upsteam channels belonging to diffeent ONUs. Without such sepaation, two ONUs may stat tansmitting not necessaily simultaneously such that thei tansmissions, when they each the tun combine, may ovelap most liely, only patially and thus will collide. The available solutions fo multiplexing ae WDM, TDM, and CDM. Table 2 descibes the advantages and disadvantages of each solution. 5

6 Advantage Disadvantage WDM Povides high bandwidth. This could be the best appoach as it is vey simple to implement. Cost and scalability: the OLT has to have a tansmitte aay with one tansmitte fo each ONU. Then, adding a new ONU could be a poblem, unless tansmittes wee ovepovisioned in advance. Each ONU must have a wavelengthspecific lase. TDM CDM Allows each ONU to have a faction of a wavelength s capacity. Only one tansmitte needed in the OLT, no matte how many ONUs ae connected. No fixed limit on numbe of uses; povides secuity. Moe complicated than WDM. Requies ONUs to be synchonized. Inte-channel intefeence inceases with numbe of uses; Most impotantly, physical components must be able to handle signal ate much highe than the use s data ate [5]. Table 2. Advantages and disadvantages of media-access technologies in PON. It should also be mentioned hee that WDM and TDM appoaches may be combined when a subset of ONUs shae a common wavelength. Fo example, if the PON combines home and business uses, business uses may shae one wavelength, and home uses may shae the othe wavelength. This solution still emains scalable as new uses can be added to each goup without adding new hadwae to the OLT. 6

7 3. Model Desciption Based on the discussion above, it seems that an Ethenet and TDM combination has the best of all qualities. An epon is a PON that caies Ethenet taffic. Because Ethenet is boadcasting by natue, it fits pefectly with the epon achitectue in the downsteam diection fom netwo to use: pacets ae boadcast by the OLT and extacted by thei destination ONU based on thei media-access contol MAC addess. In the upsteam diection fom use to netwo, each ONU will use a sepaate TDM channel. In this study, we conside a model with N ONUs. Evey ONU is being assigned a timeslot. All N timeslots togethe compose a fame. A fame typically would have a small ovehead used fo synchonizing the ONUs to the OLT s cloc, but we conside it to be negligibly small fo the puposes of ou analysis. In all numeical examples pesented in this study, the default value fo N was chosen to be 6. Fom the access side, taffic may aive to an ONU fom a single use o fom a gateway of a local-aea netwo LAN, that is, taffic may be aggegated fom a numbe of uses. Pacets should be buffeed in the ONU until the coect timeslot fo this ONU aives. Then, pacets will be tansmitted upsteam. Tansmission speed of the PON and the use access lin may not necessaily be the same. In ou model, we conside R D Mbps to be the data ate of the access lin fom a use to an ONU, and R U Mbps to be the bit ate of the upsteam slotted lin fom the ONU to the OLT see Fig. 3, with default values of R D and R U being 00 Mbps and 000 Mbps espectively, in ou numeical examples. OLT R U Mbps slotted lin ONU R D Mbps fame N timeslots Fig. 3. System model. The questions we shall ty to answe in this study ae: what is the aveage delay the pacets will expeience in the ONU buffe, how big this buffe should be, and what lin utilization we can achieve. 7

8 To pefom tace-diven simulations, we used Bellcoe taces [3]. The aival timestamps wee conveted into bytestamps one unit coesponds to a time needed to tansmit one byte. Even though the oiginal taces wee obtained on a 0-Mbps lin, using bytestamps instead of timestamps allowed us to use the taces as if they wee collected on a 00-Mbps lin. To simulate netwo behavio with highe loads, we scaled the taffic up, i.e., we popotionally deceased evey inte-pacet gap to achieve the desied load. While doing so, we ept the minimum inte-pacet gap to be 8 bytes fo peamble, as specified in the IEEE standad. It is impotant to notice that the oiginal taffic was only scaled up to simulate highe load. Scaling the taffic down does not peseve the popety of self-similaity same o simila degee of bustiness obseved at diffeent timescales efe to [6] fo an extended bibliogaphy on the subject. Speading pacets fa apat will educe the bustiness obseved on vey fine scales. Scaling taffic up, on the othe hand, peseves the bustiness, as the degee of bustiness obseved at lage scale now will manifest itself at smalle scale. Also, bustiness will disappea at a vey high load. This is due to the fact that most of the inte-pacet gaps ae educed to a minimum of 8 bytes. This is in no way epesentative of the eal netwo taffic and we did not simulate the ONU load of moe then 62.5%. This paticula value was chosen fo the following eason: maximum bandwidth available to an ONU is R U / N, which, based on default values, equals 000 Mbps / Mbps. When offeed load pe ONU exceeds 62.5 Mbps o 62.5% of input ate of 00 Mbps, the system becomes unstable. In eality, such a system should dop the pacets, thus educing the effective caied load. 4. Analysis of Pacet Delay and Queue Size In this section, we discuss how delay and queue size depend on the netwo load. We fist conside a simple FIFO queue. Only the pacets that aive befoe the timeslot may be sent in the timeslot, i.e., the system uses gated sevice. If the next pacet to be sent is lage than the emaining timeslot, then this pacet and all pacets that aived afte it will wait fo the next timeslot. Befoe we pesent ou esults, let us conside what ae the constituents of the delay expeienced by a pacet. Pacets aive to the ONU at andom times. Evey pacet has to wait fo the next timeslot to be tansmitted upsteam. This delay is temed TDM delay. TDM delay is 8

9 the time inteval between pacet aival and the beginning of the next timeslot. In [7], this delay is called slot synchonization delay. Due to the busty natue of netwo taffic, even at light o modeate netwo load, some timeslots may fill completely and still moe pacets may be waiting in the queue. Those pacets will have to wait fo late timeslots to be tansmitted. This additional delay is called Bust delay. Bust delay may span multiple fames ecall that a fame consist of N timeslots whee N is the numbe of ONUs. In ou simulation, we define pacet delay to be the time inteval between the end of eception of the last byte and the beginning of the tansmission of the fist byte. Thus, pacet tansmission time is excluded fom ou calculations. Ethenet taffic can be consideed to be an ON/OFF pocess whee an ON peiod coesponds to pacets being tansmitted, and an OFF peiod coesponds to inte-pacet gaps. Then busts of taffic, as seen by some buffe, can be chaacteized as a combination of busts of ON intevals and OFF intevals. Bust of ON intevals is a bust of pacet sizes, when the netwo suddenly sees goup of pacets of lage size. Bust of OFF intevals is a bust of intepacet gaps, when the netwo sees a goup of inte-pacet gaps of vey small size. Of couse, in eal taffic, both mechanisms affect the oveall taffic shape, and ae not sepaable. Howeve, in ou fist simulation esult, we will attempt to obseve netwo behavio when taffic is only subjected to pacet size busts and not to inte-pacet gap busts. We will use the following taces: Tace A: Tace B: Tace C: This is the oiginal Bellcoe tace [3], but scaled up to simulate the necessay netwo load. It still peseves the popety of self-similaity, i.e., simila degee of bustiness can be obseved at diffeent timescales see Fig. 4, tace A. This tace uses pacet sizes fom tace A, but has pacet timestamps modified to obtain inte-pacet gaps of equal size. Thus, effectively, this tace gets id of inte-pacet gap busts. The bustiness obseved in this tace is only due to busts of pacet sizes see Fig. 4, tace B. This tace also uses pacet sizes fom tace A. It has pacet timestamps modified to mae inte-pacet gap popotional to the size of the pacet that follows it, i.e., 9

10 a lage pacet follows a lage gap. Thus the ON busts and OFF busts ae exactly in anti-phase. If we loo at this taffic at a scale of few pacet sizes, we will see a constant taffic ate expessed in bytes pe unit of time, i.e., ON and OFF busts cancel each othe see Fig. 4, tace C. Tace A 600 bytes 200 bytes bytes Tace B 600 bytes 200 bytes bytes Tace C 600 bytes 200 bytes bytes Time Fig. 4. Illustation of taces used in simulations. We need to emphasize hee that taces B and C do not bea any esemblance to the eal netwo taffic. Howeve, compaing pacet delays in simulations using taces A, B, and C will let us visualize how much ON busts and OFF busts contibute to the delay. The following ae ou simulation paametes: Each tace contained million pacets. Fame time 2 ms. This is the time between the aivals of successive timeslots fo each ONU. Thus, the expected TDM delay is ms. Timeslot size 5625 bytes. This value is defined by the fame time, numbe of ONUs, and the line ate. 3 Fame s 2 0 s 9 bit byte Timeslot bytes Line _ Rate bytes N 6 s 8 bit No pacets wee dopped, i.e., infinite buffe at each ONU. Load was vaied fom % to 63% with % incements. A load of % is the oiginal netwo load. Highe loads wee obtained by scaling the oiginal load up. Fig. 5 shows the esults of ou fist simulation. 0

11 0 Delay s Tace A Tace B Tace C Offeed Load Fig. 5. Aveage pacet delay. Hee we can see that tace C intoduces the shotest delay. In fact, only TDM delay is pesent up to about 59% load. The eason fo such a nice behavio is that evey timeslot is getting appoximately the same numbe of bytes. Then, when one timeslot finally oveflows, all of them oveflow, and we have avalanche-lie incease of pacet delay. Tace B shows mostly TDM delay up to a load of 40%. At this load, the numbe of timeslots that oveflow stat inceasing exponentially. Tace A shows exponential incease almost fom the beginning. It is mostly due to inte-pacet gap busts, i.e., pacets ae aiving close to each othe in a pacet tain. The fact that we have some timeslots oveflowing means that thee wee some busts of taffic that deliveed moe than 5625 bytes timeslot size in 2 milliseconds fame time. This means that while the oveall netwo load was only % o Mbps, thee wee peiods when the bust ate achieved at least 5625 bytes x 8 bit/byte / 2 ms 62.5 Mbps. Fig. 6 epesents the aveage queue size sampled just befoe evey timeslot aival. This behavio is vey simila to that of the aveage pacet delay. The queue size fo tace C gows linealy up to timeslot size 5625 bytes. On loads above 59%, evey timeslot sees the queue of size lage than the timeslot can accommodate. Tace B shows that busts of pacet sizes can be toleated up to a load of 40%. Howeve, it is the inte-pacet gap busts that mae buffeing less efficient. Even at vey low load, they intoduce a fai amount of pacet tains such that lage buffes ae needed. With the incease of load, the needed buffe space inceases exponentially

12 tace A. Simila queue behavio was obseved in [8] in a model employing multiple souces of powe-tail Maov-Modulated Poisson Pocesses..E+08 Queue size bytes.e+07.e+06.e+05.e+04 Tace A Tace B Tace C.E Offeed Load Fig. 6. Aveage queue size. The above esults show that the Bust delay dominates the TDM delay in the oveall delay expeienced by a pacet. As such, it would not mae much sense to educe the TDM delay, say by educing the timeslot size. It is also tue that inceasing the timeslot size will not intoduce a lot of delay. We also illustated hee that pacet loss could not be pevented. It only can be mitigated at the expense of exponential incease of buffe space and pacet delay. This popety of selfsimila taffic povides a statling contast to models employing Poisson aival pocess. Unde those models, it was possible to incease the buffe and timeslot size just enough, so that taffic aveaged ove the fame time would appea smooth and would entiely fit in the buffe, so that no pacet loss will be obseved. In a eal netwo, the taffic busts have a heavy-tail distibution [9, 0]. Tail of distibution function fo such distibutions deceases sub-exponentially, unlie Poisson whee decease is exponential. This leads to the fact that the pobability of extemely lage busts is geate than in Poisson. This also means that no taffic smoothing is possible in eal netwos. 2

13 The next section will evaluate the effective egess bandwidth and will examine the possibility of finding optimal paametes that minimize the delay while maximizing the available bandwidth. 5. Bandwidth Utilization Peviously, we have shown that the maximum egess bandwidth available to an ONU is R U / N which equals 62.5 Mbps with ou default paametes. In this model, we assume that Ethenet pacets cannot be fagmented, i.e., if the next pacet to be tansmitted is lage than the emainde of the timeslot, the pacet will wait fo the next timeslot. This also means that the timeslot will be tansmitted with an unused emainde at the end. Then, the question is how lage the mean emainde is? Let T timeslot size R andom vaiable epesenting unused emainde andom vaiable epesenting pacet sizes A, B ange fo pacet sizes: A size B In Ethenet A 64 bytes, B 58 bytes By definition, the expected emainde is B E R P R Obviously, the emainde can only be in the ange fom to B-. If the emainde is moe than B- bytes long, then we ae guaanteed that the next pacet will fit into the cuent timeslot, thus educing the emainde. Hee, we assume that we always have pacets waiting, i.e., load is heavy. Assuming that we have placed pacets in the timeslot, what is the pobability of getting a emainde of size? Taing S + + K+ 2, this pobability is P R K P S P S P S T I S T I + + > T K > K T K P + > K 2 Since is independent and identically distibuted, 3

14 4 P K P > > + 3 To get pobability R, we sum 2 fo all, i.e., 4 K P K T S P P K P K R P R P > We sum it fo all because we do not cae how many pacets fit in a timeslot. All we cae is that, afte we have added some numbe of pacets, we still have the unused emainde of size. Stictly speaing, we do not need to sum fo all. Timeslot of a specific size T can only accommodate m pacets, whee A T m B T Now, the summation in 4 denotes the pobability that the sum of seveal pacet sizes equals to T - without any efeences to numbe of pacets used in the summation. In othe wods, this is the pobability that any numbe of pacet sizes sums to T -. Thus, we have T S P P R P > 5 We can view S as a enewal pocess with inte-enewal times. Thus, we expect to have one enewal evey E bytes. The pobability that some enewal will occu exactly at epoch T - is, theefoe, /E, i.e., E T S P 6 Substituting 5 and 6 into, we get ] [ F E E P R E B B > 7 And, fo the pobability density function, we have othewise B E F f R 0, 0, 8 The amazing esult hee is that ER does not depend on the timeslot size. It only depends on the distibution of pacet sizes. This agees vey well with ou simulations.

15 As an example, if we assume unifom distibution fo, A64, and B58, we get B A + 3A B 2A E R A + B B A + It follows that the maximum utilization achieved by an ONU is U T E R 9 T Obviously, inceasing the timeslot size should esult in inceased utilization. Fig. 7 plot a shows the utilization as the function of timeslot size. But what about the delay? As was mentioned befoe, the pe-pacet delay consists of two components: TDM delay and Bust delay. The TDM delay is popotional to the fame size and, as such, inceases linealy with the timeslot size. The Bust delay on the othe hand, deceases exponentially as the timeslot inceases. The eason fo this is that, with smalle timeslots, the utilization deceases see Equation 9, but mostly due to fact that shote busts now can cause timeslot oveflow. Fig. 7 plot b depicts a typical combination of TDM and Bust delay as a function of timeslot size. Next question we as is how do we choose the optimal timeslot size such that utilization is maximized and aveage delay is minimized. The powe function descibed in [4] can be employed as a convenient measue of optimality. The powe function is defined as U T P T, 0 D T whee UT is a utilization as a function of timeslot size and DT is a delay as a function of a timeslot size. The optimal timeslot size would be the one whee the powe function is maximal. Below we pesent ou easoning in expecting the powe function to have a maximum. TDM delay asymptotically behaves as D TDM T ~ C T whee C is a positive constant. Clealy, we expect TDM delay to be half of the fame size whee fame size itself is a multiple of timeslot size. Bust delay D BURST T is expected to decay exponentially as a function of timeslot size. Thus, we have 5

16 C3T DBURST T ~ C2 e 2 whee C 2 and C 3 ae positive constants. Total pacet delay is equal to the sum of TDM and Bust delays: D T D T D T 3 TDM + BURST The esulting delay as a function of timeslot size is shown in Fig. 7, plot b. Utilization is calculated accoding to Equation 9 and thus behaves as the plot a in Fig. 7. Plot c in Fig. 7 shows the expected behavio of the powe function see Equation 0. The functions in Fig. 7 wee obtained analytically. They do not epesent simulation esults, but athe they give an idea of the elative shapes of utilization, delay, and powe functions. a Utilization b Delay c Powe Function Timeslot size bytes Fig. 7. Optimal timeslot size. The only paamete in calculating the powe function that depends on offeed load is Bust delay. The easonable question to as is how ou optimal opeating timeslot size changes when the offeed load changes. Fig. 8 pesents a family of nomalized powe functions calculated fo vaious netwo loads. 6

17 % load 20% load 30% load 40% load 50% load 60% load Timeslot size bytes Fig. 8. Nomalized powe function fo diffeent loads. This is an inteesting discovey. Not only does the delay incease exponentially with inceased load, but the maximum of the powe function also shifts with exponential scale. This means that, if we attempt to eadjust the timeslot size in eal time to eep the optimal opeational point, not only the Bust delay will be exponential, but TDM delay will also incease exponentially. This may have a negative impact on the pefomance, as the TDM delay affects not only the pacets that aived inside the busts, but also the pacets that aived between the busts. Fig. 9 pesents the dependency of the best timeslot size vesus the offeed load as measued by the powe function. This is again to illustate the exponential dependency. The simulation was pefomed with timeslots vaying fom 2000 to bytes with 000-byte incements. The figue also shows the extapolated function to see the timeslot incease above bytes. 7

18 80000 Timeslot size bytes measued extapolated Load Fig. 9. Dependency of timeslot vs. offeed load. In this section we showed that the ONU could not completely utilize the slotted lin available to it. Thee is an unused emainde at the end of the timeslots. The mean value of the emainde is independent of the timeslot size and only depends on the distibution of pacet sizes. This explains why we have a nee of the delay plot fo tace C Fig. 5 aound 59% and not aound 62.5%. This is because lin utilization is appoximately 59/ %. We also demonstated that adjusting the timeslot size could not be a solution to optimizing the utilization-to-delay atio. The timeslot size would need to be adjusted exponentially with espect to changed load. That would incease the TDM delay fo all the pacets sent by the ONU. Moe impotant, it would also incease the fame size and cause additional TDM delay fo pacets sent by othe ONUs. Thee is one moe eason why the timeslot size should not be changed, namely fo QoS issues. While the Bust delay can be avoided fo high pioity pacets by using cleve scheduling schemes, the TDM delay is a fundamental delay that affects all pacets. Alteing the timeslot size would mean that we ae unable to povide any guaantees to the delay value o vaiation. In the next section we will conside a scheduling appoach in ou attempt to educe the size of the unused emainde, thus inceasing the utilization. 8

19 6. Scheduling In ou delay and buffe size simulations, we used FIFO queues. A smate appoach may be to attempt to eode some pacets waiting in the buffe. If in the peviously discussed method, a pacet that is cuently at the head of the queue does not fit in a patially occupied timeslot, this pacet and all following pacets will wait fo the next timeslot. Howeve, if some late-aiving pacet in the queue is small enough to fit into the cuent timeslot, then why wait? Fig. 0 illustates these two appoaches. Hee thee timeslots ae needed without pacet eodeing, but only two timeslots will suffice with eodeing. buffe a No scheduling timeslot 2 timeslot timeslot 3 5 buffe b With scheduling timeslot timeslot Fig. 0. Illustation of scheduling. This is a vaiation of the bin-pacing poblem. Diffeent flavos of the algoithm may be used: fist fit, best fit, pediction, etc. Fig. pesents the compaison of lin utilizations fo no eodeing scheme FIFO and eodeing using fist fit. These esults wee obtained by pefoming simulations with tace C at vey high load appox. 73%. This was done in ode to have moe timeslots satuated. 9

20 ..05 Utilization FIFO Fist Fit Timeslot size bytes Fig.. Maximum lin utilization FIFO vs. Fist Fit. This eodeing can be easily implemented in hadwae as well as in softwae. While we have not obtained a mathematical equation fo ER in this case, it is easy to see that ER will depend not only on the pacet size distibution, but also on the netwo load. Indeed, the highe the load, the moe pacets will be waiting in the queue, and the highe is the possibility of finding one pacet that fits in the emainde of the timeslot. Howeve, as it tuns out, fist-fit scheduling is not such a good appoach. To undestand the poblem, we need to loo at the effects of pacets eodeing fom the pespective of the TCP/IP payload caied by Ethenet pacets. Even though TCP will estoe the pope sequence of pacets, an excessive eodeing may have the following consequences: Accoding to the fast etansmission potocol, the TCP eceive will send an immediate ACK fo any out-of-ode pacet, wheeas fo in-ode pacets, it may geneate a cumulative acnowledgement typically fo evey othe pacet []. This will lead to moe unnecessay pacets being placed in the netwo. 2 Second, and moe impotant, pacet eodeing in the ONU may esult in a situation whee n late pacets ae being tansmitted befoe an ealie pacet. This would geneate n ACKs n- duplicate ACKs fo the ealie pacet. If n exceeds a pedefined theshold, it will tigge pacet etansmission and eduction of the TCP s congestion window size the cwnd paamete. Cuently, the theshold value in most TCP/IP potocol stacs is set to thee efe to the Fast Retansmission potocol in [] o elsewhee. 20

21 Even if special cae is taen at the ONU to limit out-of-ode pacets to only one o two, the netwo coe may contibute additional eodeing. While tue eodeing typically geneates less than thee duplicate ACKs and is ignoed by the TCP sende, togethe with eodeing intoduced by the ONU, the numbe of duplicate ACKs may often exceed thee, thus focing the sende to etansmit a pacet. As a esult, the oveall thoughput of use s data may decease. So, what is the solution? As we mentioned ealie, we assume that the taffic enteing the ONU is an aggegate of multiple flows. In the case of business uses, it would be the aggegated flows fom multiple wostations. In the case of a esidential netwo, we still may expect multiple connections at the same time. This is because, as a conveged access netwo, a PON will cay not only data, but also voice-ove-ip VOIP and video taffic. Also, home appliances ae becoming netwo plug-and-play devices. The conclusion is that, if we have multiple connections, we can eode pacets that belong to diffeent connections, and neve eode them if they belong to the same connection. The outline of the algoithm is given in Fig. 2. 2

22 Let Q be the queue of pacets q, q 2,.. q n waiting in the ONU Cq i connection Id of pacet q i P set containing Ids of pacets that wee postponed R slot emainde Repeat fo evey timeslot { i P Clea the set P R timeslot While i n and R min { If { q i R then pacet fits into timeslot if { C q i P i.e. pacets fom this connection send q i. R R qi wee not postponed yet } else } pacet does not fit into timeslot P P C q add connection Id to P i } } i i + Fig. 2. Algoithm fo connection-based pacet eodeing. The algoithm in Fig. 2 peseves the ode of pacets within a connection by eeping tac in set P, see Fig. 2 of all the connection identifies of pacets that wee postponed. Obviously, the fine the ganulaity of connection identifies, the moe eodeing possibilities the ONU will have, but moe memoy would need to be allocated fo the set P which pobably should be implemented as a hash table. So, if connection ID is identified only by the souce 22

23 addess, then in case of a single use with multiple connections, the ONU will not be able to eode any pacets. Looing at the destination addess instead of the souce addess may impove the situation fo ONUs with a single use, but this has a potential dawbac in the situation when multiple uses send pacets to the same addess. In this situation, even though pacets ae fom diffeent sendes, the ONU will not eode them. A easonable solution may be to loo simultaneously at a souce-destination pai, plus also to include the souce and destination pot numbes. Then, the ONU will have maximum flexibility. Moe studies need to be done to detemine the statistical popeties of the connections to estimate the advantages of fine ganulaity connection identifies. Howeve, an impotant point is that the impovement achieved by this scheduling will be between the FIFO case and Fist Fit case. Clealy, the above algoithm will eode some pacets, which will mae its utilization bette than in FIFO. It is also tue that some pacets that belong to the same connection and that Fist Fit will eode will not be eodeed in the given algoithm; thus its pefomance will be lowe than that of Fist Fit. But unless we use the extemely small timeslot size less than 5000 bytes, the diffeence in lin utilization between the FIFO case and Fist Fit algoithms is only in the ange between.5% and 4.5 % efe to Fig.. The impotant question then is whethe it maes sense to invest in additional hadwae and softwae cost to implement something that can impove utilization by maybe 4% maximum. It would not mae too much economical sense to implement this algoithm if these 4% of impovement should bea the whole cost of implementation. But we should emembe that PON is viewed as a technology fo full-sevice access netwos. As such it should be able to povide QoS suppot. It may be eithe DiffSev, RSVP, o MPLS flow switching. In any case, the ONU should have the ability to eode pacets based on the TOS field in the IP heade, Fowading Equivalence Class FEC, o some othe pioity identification. Then adding the algoithm to impove utilization may come at the vey low additional cost. 7. Conclusion In this study we discussed and evaluated design issues that must be dealt with in a PON access netwo. In Section 4, we investigated the pacet delay. We found that the Bust delay consideably exceeded the TDM delay even at vey light loads. At highe loads, this diffeence became even moe damatic. Simila situation was shown fo the queue size: lage busts wee 23

24 pesent at vey low aveage load. These obsevations led us to a conclusion that pacet loss could not be pevented. Having lage buffes will slightly educe the congestion, but will incease the Bust delay, as moe pacets will be accumulated duing busts. Then we investigated ways to incease the bandwidth utilization. In Section 5 we showed that the ONU does not use all the bandwidth available to it. Thee will be an unused emainde at the end of a timeslot. Inteestingly, we found the expected value of this emainde to be independent of the timeslot size except the timeslots compaable in size with pacet sizes. Then, we analyzed the possibility of adjusting the timeslot to find the optimal opeating value that optimizes the delay and utilization. We found this appoach to be not feasible, as it equies the timeslot adjustment to have exponential magnitude with espect to effective load. That would mae the vaiations of TDM delay to have exponential amplitude. In Section 6, we consideed an altenative appoach to impove the utilization: pacet scheduling. We showed that the Fist Fit algoithm may slightly impove the utilization, but will have negative impact on the TCP/IP connection behavio. We then suggested a connectionoiented fist-fit algoithm. That algoithm was found to be too computationally expensive weighting against its benefits. Howeve, it may be implemented as pat of QoS scheduling. This study has some shotcomings. Fist, even though the Bellcoe taces ae of high quality, the sevices and usage of the Intenet since the time the taces wee collected have changed. Second, we have not yet veified the esults pesented hee on a hadwae pototype. And finally, we showed that pacet loss is unavoidable, but have not yet simulated ONUs with finite buffes and vaious pacet dop policies. This wo is still in pogess. Bibliogaphy [] G. Pesavento and M. Kelsey, PONs fo the boadband local loop, Lightwave, PennWell, vol. 6, no. 0, pp , Septembe 999. [2] B. Lung, PON achitectue futuepoofs FTTH, Lightwave, PennWell, vol. 6, no. 0, pp , Septembe 999. [3] W. Leland, M. Taqqu, W. Willinge, and D. V. Wilson, On the self-simila Natue of Ethenet Taffic extended vesion, IEEE/ACM Tansactions on Netwoing, vol. 2, no., Febuay 994, pp

25 [4] L. Kleinoc, On Flow Contol in Compute Netwos, Poc. of ICC 78, Toonto, Ontaio, June 978, vol. 2, pp [5] B. Muhejee, Optical Communication Netwos, McGaw-Hill, New Yo, 997. [6] W. Willinge, M. Taqqu, and A. Eamilli, A Bibliogaphical Guide to self-simila taffic and pefomance modeling fo moden high speed netwos, in Stochastic Netwos: Theoy and Applications in Telecommunication Netwos F. P Kelly, S. Zachay, and I. Ziedins, eds., vol. 4 of the Royal Statistical Society Lectue Notes Seies, Oxfod Univesity Pess, Oxfod, 996. [7] J. Hammond and P. O Reilly Pefomance Analysis of Local Compute Netwos, Addison- Wesley Publishing Co., Reading, MA, 986. [8] P. M. Fioini, On Modeling Concuent Heavy-Tailed Netwo Taffic Souces and its impact upon QOS, Poc. of ICC 99, Vancouve, June 999, vol. 2, pp [9] W. Willinge, V. Paxson, and M. Taqqu, Self-Similaity and Heavy Tails: Stuctual Modeling of Netwo Taffic, A Pactical Guide to Heavy Tails: Statistical Techniques and Applications, R. Adle, R. Feldman, M. S. Taqqu, eds. Bihause, Boston, 998. [0] K. Pa and W. Willinge, Self-Simila Taffic: An oveview, Pepint. To appea in Self-Simila Netwo Taffic and Pefomance Evaluation, Wiley-Intescience, [] W. R. Stevens, TCP/IP Illustated, Volume, Addison-Wesley Publishing Co., Reading, MA,