Experimental Evaluation of DOCSIS1.1Upstream Performance Radim Bartoˇs1,Chaitanya K.Godsay1,2,and Steve Fulton2
1Department of Computer Science
2InterOperability Laboratory
University of New Hampshire,Durham,NH03824,USA
E-mail:rbartos@cs.unh.edu
ABSTRACT
Data-Over-Cable Service Interface Specification (DOCSIS)is one of the many last mile technologies intended to provide Internet access and packet-bad rvices to the customer.DOCSIS us the widely de-ployed hybridfiber/coax(HFC)network as the physical link between multiple cable modems(CMs)and the cable modem termination system(CMTS).
This paper prents the upstream performance of DOCSIS1.1in the physical layer and MAC layer across various traffic patterns.The experiments are conducted on real devices and not simulators as us
ed in all previous re-arch.The u of real devices allows us to capture the complete complexity of the protocol,and gives us realis-tic results but it also limits our control over different pa-rameters.The goal of the project is to study the impact of different parameters that can be controlled by the cable r-vice provider and to compare different CMs with respect to upstream performance.The performance metrics ud are upstream data rate and channel utilization.The results can be ud by the cable operators to optimize their networks, by the CM and CMTS manufacturers to enhance their prod-ucts and they may help in identifying protocol bottlenecks for upstream performance.
KEYWORDS
Data over cable,DOCSIS,upstream performance evalua-tion.别具匠心的意思
1Introduction
Cable operators,in the early nineties,envisioned the growth of cable networks and were driven to explore possi-bilities for transmitting data from the residential ur to the rvice provider.By providing this capability,packet-bad rvices,such as high-speed Internet access,cheaper tele-phone connections,and video-conferencing could be de-ployed easily.This led to the formation of many rearch groups and Multimedia Cable Network System(MCNS),a collaboration of cable comp
南京风水anies,was thefirst to come up with a specification.MCNS relead the t of standards known as DOCSIS1.0(Data Over Cable Service Interface Specification)in March1997.CableLabs,a non-profit re-arch and development consortium,worked in collabora-tion with MCNS,and is now responsible for developing new specifications and product certification.
Specification[1]describes a DOCSIS network as a tree bad network with the Cable Modem Termination System(CMTS)as the root of the tree and the Cable Modems(CMs)as the leaves of the tree.The CMTS is at the rvice provider facility and the CM is at the res-idential ur home.The transmission of data from the CMTS to CM,termed as downstream,is a point to multi-point broadcast,whereas the transmission from the CM to CMTS,termed as upstream,is controlled by the CMTS and is multipoint to point TDMA.DOCSIS defines an asym-metric network in terms of upstream and downstream data rate,with downstream being substantially larger than the upstream.The residential ur has the Customer Premi Equipment(CPE),such as computer,telephone,etc.,con-nected to the CM.Data in upstream goes from the CM to the CMTS,which is then forwarded appropriately.Sim-ilarly,data in downstream,pass from the CMTS to the CM,which is forwarded to the CPE.Typically there are 1500to2000CMs connected to a CMTS with distance be-tween the CMTS and CM going up to50miles.
Let us briefly discuss a typical data transmission in upstream.Once the CM has registered with the CMTS,it is allowed to transmit data upstream.However,it can nd data only when it is allowed by the CMTS to do so.Since upstream is multipoint to point TDMA arbitrated by the CMTS,the CMTS nds bandwidth allocation MAPs,sim-ply termed as MAPs,at regular intervals.The MAP explic-itly describes the time when a CM is allowed to transmit data upstream.
Transmitting data upstream is a three-step process. First the CM has to nd a request for a data grant to the CMTS,then it has to wait to get a data grant from the CMTS(in the MAP)and then it must nd the data(at time specified by the MAP).We call this process the RDS cycle (Request-Data grant wait-Send cycle).
When the CPE nds some data to the CM,the CM looks in the most recent MAP for the REQ or REQ/DATA region.It then creates a data grant request message indi-cating the grant size and transmits the request in the time specified for the REQ or REQ/DATA region in the MAP. The regions are subject to collisions as many CMs can try to nd a data grant request message.If the data grant request message reaches the CMTS then it either nds a long data grant or short data grant to the CM in the fol-lowing MAP.A long or a short data grant depends on how
much data the CM wants to nd.The CMTS will nd a data grant pending message in the following MAP,if the CMTS has received the data grant request from the CM but cannot allocate a data grant to it.The CM detects a colli-sion when it does not get a short data grant,long data grant, or a data grant pending in the next MAP.In the ca of a collision,the CM starts exponential backoff for collision resolution.The CM will then defer certain number of re-quest opportunities before requesting again.
However,if the CM gets a short/long data grant suc-cessfully from the CMTS,then it extracts the time to nd the data from the MAP and transmits the data to the CMTS at the time specified by the MAP.The CM thus has to go through one or more RDS cycles to transmit data upstream.
It can be en that DOCSIS is a complex protocol that is difficult to model.This paper attempts to capture the full complexity of the protocol by measuring the performance of real devices.
2Motivation and Goals
In the early nineties many rearch groups were formed to develop a specification for delivery of data in the last mile using the widely deployed cable networks.Organizations involved in this effort were Data Over Cable Service Inter-face Specification group(MCNS-DOCSIS),IEEE802.14 working
group,Society of Cable Telecommunications En-gineers(SCTE),Digital Video Broadcasting(DVB),Digi-tal Audio Video Council(DA VIC),and ATM Forums Res-idential Broadband Working Group(RBWG).The initial rearch in DOCSIS was on comparing different aspects of specifications developed by different groups,mainly IEEE 802.14and MCNS-DOCSIS[2].The effect of different up-stream allocation and scheduling algorithms,MAP rates, MAP length,etc.,were studied in[3,4,5].Efforts to sta-tistically predict the upstream requests and allocate data grants was prented in[6].Recently there has been a study on performance evaluation of DOCSIS1.1using simulator [7].All of the above rearch have employed simulators for experimentation with most of them using Opnet as the DOCSIS simulator.
The prented study has been carried out by using real devices with the goal to evaluate the impact of different upstream parameters,that can be controlled by the cable rvice provider,on upstream performance.Different pa-rameters are considered from the PHY layer,MAC layer, and traffic patterns.
In the PHY layer the effect of different modulation formats and modulation rates is considered.In the MAC layer we consider performance enhancers,such as concate-nation and piggybacking,that improve the upstream per-formance.However,there has been no rearch on the behavior of the e
nhancers.How much performance im-provement do they provide?Is there a situation where us-ing the enhancers might be detrimental?Are there bot-tlenecks in using a combination of the enhancers?This paper prents the experimental evaluation of the aspects of the protocol.
Another aspect of the study is to understand the be-havior of the DOCSIS network under different kinds of traffic loads.Devices are generally tested with a constant bit rate stream.However,a constant bit rate stream is hardly ever generated in the real world.Since Internet is the biggest application driving the cable modem market we decided to concentrate on generating Internet-like traffic. Some advanced testing devices now support traffic distri-butions for packet length.A traffic pattern that cloly ap-proximates realistic traffic was generated and transmitted through the DOCSIS network to aid in understanding the nsitivity of the DOCSIS network to different traffic pat-terns.
The change in performance due to the addition of CMs across different CM manufacturers is also analyzed. We hope that the results will be ud by the vendors to en-hance their products.
It should be noted that we do not test for the confor-mance to the DOCSIS1.1protocol.The conformance and certification testing of CMs and CMTSes is done by Ca-bleLabs only.We u certified and conformant CMs and CMTSes to get statistical results and analyze the impact of different protocol parameters.
3Parameters considered in this study
We u upstream channel utilization and upstream data rate as the performance metrics.Latency,another measure of performance,is not considered in this paper,however,it is a subject of our current rearch.The parameters considered for evaluation of upstream performance are enlisted below. Modulation Formats and Modulation Rates:DOCSIS
1.1allows two modulation formats for upstream
transmission,QPSK(2bits/symbol)and16QAM (4bits/symbol).DOCSIS 1.1also supportsfive modulation rates for upstream transmission,160 ksym/s,320ksym/s,640ksym/s,1280ksym/s,and 2560ksym/s which correspond to the channel widths of200kHz,400kHz,800kHz,1600kHz,and 3200kHz respectively.The product of modulation rate and modulation formats gives us the theoretical maximum upstream data rate for each combination of modulation rate and modulation format.We thus have
0.32Mbps,0.64Mbps,1.28Mbps,2.56Mbps,and
5.12Mbps as the theoretical maximum data rate for
QPSK and0.64Mbps,1.28Mbps,2.56Mbps,5.12 Mbps,and10.24Mbps for theoretical maximum data rate for16QAM.The product of modulation rate and modulation format is also referred to as the channel rate.
Concatenation:The CM can nd a concatenated burst of packets instead of small packets if allowed by the con-figurationfile and the CMTS.The configurationfile provides operational parameters to the CM when it
registers with the CMTS.We u the configurationfile to control the performance enhancers(on/off).The improvement in upstream performance by using con-catenation,in terms of maximum data rate achieved without dropping any packets,is studied. Piggybacking:If the CM wants to nd data upstream,it has to request a data grant in the REQ or REQ/DATA region.Alternatively,it can request a data grant by piggybacking a request with the data packet being nt.The CM does so by adding an extended header to the data packet being nt.Piggybacking is enabled by using the configurationfile,and should be allowed by the CMTS.
Traffic profiles:We conduct experiments on Constant Packet Length-Constant Bit Rate(CPL-CBR)traffic, as it provides uful insights into the behavior of CMs.
However,since the biggest application driving the ca-ble modem market is high-speed Internet connectivity, we study the behavior of a DOCSIS network with the source,transmitting Internet-like traffic.We define Distributed Packet Length-Constant Bit Rate(DPL-CBR)as traffic that would transmit different packet sizes at the micro level but maintain a constant bit rate at the macro level.Several advanced traffic generators now support quad-modal packet length distribution (four Gaussian distributions superimpod,each with adjustable mean and half-point width).The above fea-ture is ud to compare,understand and analyze the behavior of the DOCSIS network under realistic traf-fic loads.
Number of CMs:Since the upstream performance is lim-ited by the RDS cycle,having more active CMs on the network with the potential of transmitting data up-stream increas the probability of collisions.Since the number of RDS cycles increa on collision,we study the effect of upstream performance for different numbers of CMs on the network.
CM chipt manufacturers:We conduct experiments on devices bad on different chipts.This provides im-portant insights into the nsitivity of different CMs to different performance parameters.The results can also be ud by the CM manufacturers to enhance their devices.
4Experimental tup
The experiments prented in this paper were conducted on the testbed shown in Figure1.The test network consists of a CMTS connected via coaxial cable plant to one or more CMs(the number of CMs varies with the experiment as described in the next ction).The upstream traffic is gen-erated by a traffic generator connected to the CM over an Ethernet network.The output of the CMTS is routed over another Ethernet gment to the traffic analyzer.Two traffic
Coaxial cable
Ethernet
Figure1.Experimental tup.
generators/analyzers were ud in the experiments:Smart-Bits600chassis with LAN3101A cards for CPL-CBR ex-periments and Adtech AX4000chassis with10/100BaT Ethernet interfaces for DPL-CBR experiments1.
Sigtek ST-260B DOCSIS1.0/1.1RF sniffer/traffic an-alyzer was ud to ascertain that the traffic generated by the CMs adhered to the t parameters.This helped us elimi-nate veral CMs that incorrectly implemented certain as-pects of the ,piggybacking)and led to the final decision to limit the study to only CableLabs certified CMs.
Two ts of CMs,each consisting of identical devices, were ud in the prented study.One t consisted of two CMs bad on Broadcom BCM3300QAMLink chipts. The CMs in the cond t were bad on Texas Instru-ments TNETC4040chipts.
We u the above test tup tofind the maximum data rate(throughput)at which the CM can transmit upstream without packet loss2.A script was ud to t the CM pa-rameters,to control the traffic generator,and to process the results from the analyzer.A binary arch algorithm was ud tofind the maximum throughput of the modem.The arch algorithm starts with0as the minimum data rate and theoretical maximum data rate as the maximum.It then av-erages the minimum and maximum tofind
the current data rate to transmit.If the transmission succeeds then the cur-rent is made the minimum and the process continues.If the current data rate transmission fails then the current is made the maximum and the process of averaging the mini-mum and maximum,followed by transmission and update continues.This process will terminate only when the dif-ference between the maximum and minimum is within the tolerance specified.
The test tup is such that there is no injected physi-cal noi or interference,only a few feet of cable and all
T h r o u g h p u t (M b p s )
Packet length (bytes)
Figure 2.Experiment 1–Maximum data rate (single Broadcom-bad CM,modulation format 16QAM,piggy-backing and concatenation on).
C h a n n e l u t i l i z a t i o n (%)鱼腥草
Packet length (bytes)
Figure 3.Experiment 1–Channel utilization (single Broadcom-bad CM,modulation format 16QAM,piggy-backing and concatenation on).
the devices clo to each other.This is almost an ideal condition.In the deployed cable networks the distances are more and there is significantly more noi and inter-ference.However,experimental tup for generating and controlling such an environment was not available for the experiments.We have thus lected only tho parameters that are not affected,or minimally affected by distance and interference.
5Experimental evaluation
This ction prents a lection of the results that were obtained in the study.A full t of results can be found at www.cs.unh.edu/cnrg/cgodsay/DOCSIS-study.
0.5 1 1.5 2 2.5 3 0 200 400 600 800 1000 1200 1400 1600
T h r o u g h p u t (M b p s )
Packet length (bytes)
All On All Off Piggy On Concat On
Figure 4.Experiment 2–Performance enhancers (single Broadcom-bad CM,channel rate 5.12Mbps,modulation format 16QAM).
Experiment 1:Channel rate
The purpo of this experiment was to study the impact of varying channel rates by changing the channel widths and keeping the modulation format the same.Figure 2shows the throughput performance of a network with one modem.It can be en that as the channel rate increas the modem throughput increas.Figure 3displays the results of the same experiment as channel utilization percentages (ratio of obrved data rate and channel rate).The channel uti-lization is well below the theoretical maximum data rate and decreas as the channel rate increas.Experiment 2:Performance enhancers
Figure 4shows the results of experiments that evaluate the impact of performance enhancers,piggybacking and con-catenation.It can be en that concatenation significantly improves the throughput for small packet lengths.In our experiments,piggybacking did not have a significant im-pact on the performance.We have conducted the same ex-periment for all possible channel rates and obtained similar results.
Experiment 3:Modulation format
This experiment evaluated the impact of the two available modulation formats on the performance (s
ee Figure 5).As outlined earlier,the experiment tup did not allow injec-tion of physical layer impairments that would truly test the benefits of each modulation format.Instead,the experi-ment concentrated on the protocol-level aspects.The most pronounced difference in performance was obrved for channel rate 2.56Mbps where QPSK clearly outperformed 16QAM.The results were mixed for the remaining channel rates.
T h r o u g h p u t (M b p s )
Packet length (bytes)
结婚的意义
Figure 5.Experiment 3–Modulation format (single Broadcom-bad CM,piggybacking and concatenation on).
Experiment 4:CM chipt
Modems bad on two different chipts were available for the experiments (2bad on Broadcom BCM3300QAM-Link and 3bad on Texas Instruments TNETC4040).Fig-ure 6compares the results for single modem experiments.The performance is comparable for lower channel rates while for higher channel rates the performance of the TI-bad CM drops dramatically when packet length exceeds 780bytes.This behavior was obrved consistently over a wider range of parameters than that shown in the figure.Since we can only obrve the external behavior of a CM,we are so far unable to determine the cau of the perfor-mance drop.
Experiment 5:Number of modems
This experiment evaluated the impact multiple simultane-ously transmitting CMs.Figures 7and 8sho
w the perfor-mance for modems bad on both chipts.In both cas,the per-CM throughput remains roughly the same even if two modems transmit simultaneously.This is not surpris-ing given the results of the previous experiments where sole CMs were unable to achieve channel utilization better than a fraction of the theoretical maximum data rate.In the ca of three modems,the carrying capacity of the network is reached and the per-CM throughput decreas.Experiment 6:Internet traffic mix
In this experiment,the network was subjected to traffic with packet length distribution that approximates packet length distribution of real Internet traffic [8,9].We have ud quad-modal distribution that superimpos four Gaussian distributions with parameters shown in Table 1.Table 2gives the results for the four possible combinations of per-formance enhancers and both models of CMs.By com-
T h r o u g h p u t (M b p s )
Packet length (bytes)
Figure 6.Experiment 4–Comparison of CMs bad on dif-ferent chipts (single CM,channel rate 5.12Mbps,mod-ulation format 16QAM,piggybacking and concatenation on).
描写祖国山河的诗
Distribution number 24Mean in bytes
300
1500
0.1
0.1
Weight
20%15%Table 1.Internet traffic mix parameters.
paring the outcomes of this experiment with the results ob-tained in Experiment 2for fixed-size packets,it can be con-cluded that the packet distribution does not significantly af-fect the CM performance.The values obtained can be com-pared to Figure 4for the Broadcom-bad CM and we can conclude that the behavior was approximately equivalent to passing 500-byte packets with CPL-CBR.
锦绣集团6Conclusions and Future Work
三年级下册语文古诗DOCSIS is a complex protocol with complex interactions.This makes it difficult to provide generalized data through-put projections bad solely on channel capacity and of-fered data rate.However,we have analyzed different pa-rameters involved in each layer of DOCSIS,namely PHY and MAC layers across different traffic profiles,with focus on upstream channel utilization and upstream data rate.
We have obrved that one CM is unable to utilize all the available bandwidth even in nearly ideal condi-tions.The throughput of 16QAM was not twice QPSK and as channel rate incread,channel utilization decread.Concatenation and piggybacking helped significantly for packet sizes below 800bytes.The comparison between different CM manufacturers becomes more pronounced as the channel rate increas.Running the tests for more num-ber of CMs gave expected results.The per CM data rate remained constant until the carrying capacity of the net-work was reached after which the per CM data rate dropped
T h r o u g h p u t (M b p s )
Packet length (bytes)
Figure 7.Experiment 5–Per-modem throughput for one and two Broadcom-bad CMs (modulation format 16QAM,piggybacking and concatenation on).
T h r o u g h p u t (M b p s )
Packet length (bytes)
Figure 8.Experiment 5–Per-modem throughput for one,two,and three TI-bad CMs (modulation format 16QAM,piggybacking and concatenation on).
as we added more CMs.Passing distributed packet length traffic did not significantly effect the CM performance.
The main contribution of this project is the experi-mental evaluation of upstream performance of DOCSIS 1.1networks.There have been numerous studies on the subject that utilized either analytical models or simulators.This is the first study that attempts to evaluate the performance of real devices.
个人网站设计Our future work will focus on latency experiments and on finding additional measures to compare CM per-formance.The percentage of MAP opportunities ud by a CM as a comparison measure ems promising.Experi-ments will also be run on different CM and CMTS vendors with more number of CMs on the network.
We are working on a method to predict the upstream performance,to be ud by the cable rvice providers.By using this method the cable rvice providers can study var-
Piggybacking on on Concatenation
off on Broadcom-bad CM
1.08
1.48
0.84
0.92
