INTERNET-DRAFT Hassan Naser Diffserv Working Group Alberto Leon-Garcia Expires: June 1999 Dept. of Electrical and Computer Eng. University of Toronto Toronto, ON M5S 3G4, Canada Osama Aboul-Magd Nortel Networks Ottawa, ON K1Y 4H7, Canada December 1998 Voice over Differentiated Services Status of This Memo This document is an Internet-Draft. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." To view the entire list of current Internet-Drafts, please check the "1id-abstracts.txt" listing contained in the Internet-Drafts Shadow Directories on ftp.is.co.za (Africa), ftp.nordu.net (Northern Europe), ftp.nis.garr.it (Southern Europe), munnari.oz.au (Pacific Rim), ftp.ietf.org (US East Coast), or ftp.isi.edu (US West Coast). Abstract We study the performance of the Constant Bit Rate (CBR) and the interpolated ON-OFF voice over the differentiated services Internet. Both the Premium and Assured services are tested, and their relative performance is investigated. Using these services, at the DS3 rate, the call handling capacity can be doubled, at the expense of packetization delay and a few milliseconds packet delay. Also, at a reasonable load level, if the Assured flows are smooth their playout delay could be limited to a few milliseconds. This is true even in the presence of the higher-priority Premium flows. The loss rate of the Assured packets at the RIO buffer is observed to be within the order of 10^(-06), where as the best effort packet loss rate is within the order of 10^(-02). Finally, shaping the Premium traffic reaches diminishing returns; shaping could often exceedingly increase the end-to-end delay and jitter of the Premium traffic. A pdf or ps version of this document is available at: http://www.comm.toronto.edu/~hnaser/page2.html#pub96 Naser, et. al. Expires: June 1999 [Page 1] INTERNET-DRAFT Differentiated Services Field December 1998 1. Introduction We present a simulation study of voice transmission over the two-bit differentiated services [2BIT] for the Internet. The ``Premium'' service is proposed by Jacobson [2BIT] to meet the demand for high quality service. This service essentially creates a ``virtual line'' with the bandwidth equal to the desired peak bit-rate being negotiated. The Premium flows are expected to experience a very small queuing delay and delay-jitter inside the network. This service could potentially yield the levels of quality that the Constant Bit Rate (CBR) applications require. The ``Assured'' service proposed by Clark et al. [Clark1] [Clark2] provides a ``better level'' of service than the ordinary ``best effort'' service, during period of network congestion. The level of service assurance is determined by an expected bit-rate that is negotiated with the network. If the level of assurance is not satisfactory, then a higher bit-rate must be (re)negotiated. This service allows more statistical multiplexing of data, at the expense of lower quality, than the Premium service. The Assured service might be useful for transmitting the Variable Bit Rate (VBR) traffic, such as the interpolated ON-OFF voice. However, the achievable quality of service needs to be investigated. Reference [2BIT] lists the benefits of deploying both Premium and Assured services, and provides a set of common mechanisms in the access, core, and border routers of an internet in order to implement them. The network routers that deploy both services always serve the Premium traffic with a high priority over the Assured and best-effort traffic. This, however, should not significantly deteriorate the performance of Assured flows as long as the Premium allocations remain small. For the combined architecture to work, the access routers (or other edge devices) classify the flows, and check their conformity to the contracted profile. The Premium flows need to be shaped in order to generate sufficiently smooth traffic. Regardless of what shaping mechanism is chosen, one can speculate that the Premium flows must initially be smooth, or else, shaping can have a effect. For example, our study of bursty video traffic (the results are not reported here) indicates that shaping can introduce large delay and delay-jitter into the Premium traffic. Currently, there are two mechanisms favored by the DiffServ community to implement the conformity check (i.e, policing/reshaping) at the access routers. The ``rate estimator'' mechanism [Clark2] measures the sending-rate of the corresponding TCP flow, and allows the policer to mark packets once the measured rate exceeds a certain threshold. On the other hand, the ``leaky bucket'' mechanism measures the amount of data that a flow generates Naser, et. al. Expires: June 1999 [Page 2] INTERNET-DRAFT Differentiated Services Field December 1998 over any time interval. If the amount of data exceeds a certain threshold the leaky bucket marks (or discards in some implementations) the flow's packets. We chose to implement the ``leaky bucket'' mechanism as a means of checking the traffic conformity. For voice and video traffic, the performance of the leaky bucket mechanism has long been investigated. We also use this mechanism to implement the border policing functions. The performance of the rate estimator mechanism needs to be studied for voice and video traffic. In this draft, we study the performance of Premium service and Assured service for CBR and ON-OFF voice traffic. We will examine the relative performance of the Premium and Assured services in a congested network. We will identify the situations where the Premium performance does not change significantly, while the Assured traffic performance deteriorates. The parameters of our study are the packet length, the level of network load (i.e, the fraction of time that the network nodes are busy, servicing the packets), the relative portion of the link bandwidth assigned to each service class, and the number of sources that compete for service at each node. First, we briefly review the common architecture used to implement the differentiated services [2BIT]. We refer the reader to the reference for a more detailed explanation. We then introduce our network model and the simulated parameters. Finally, we end our report with a discussion of results. 2. Implementation of Differentiated Services Figure 1 shows our implementation of the differentiated services architecture proposed in [2BIT]. The network elements deployed both Premium and Assured services in addition to the default best effort service. These network elements are: leaf routers, core routers, and border routers of the Internet. Figure 1: (see pdf or ps form of document for all figures and tables). Leaf routers were configured with ``traffic profiles'' of individual differentiated services flows. Each traffic profile set up a ``traffic Marker'' with the appropriate per flow traffic descriptors. In our implementation, the Markers were leaky buckets that compared the generated user traffic against the configured rate and burst traffic descriptors. Unless otherwise stated, the Premium bucket rate was configured with the flow's peak bit-rate. The burst size (bucket depth) of the Premium Marker was set to the size of the largest packet the flow generated. The Premium Marker shaped the arriving traffic by enqueuing the user data generated faster than the bucket consumption rate. Thus, all packets that emerged from Naser, et. al. Expires: June 1999 [Page 3] INTERNET-DRAFT Differentiated Services Field December 1998 the Marker were considered conforming Premium packets, albeit to a probable (large) shaping delay. The Assured (leaky bucket) Marker was set up with the flow's requested rate and burst size. Unlike, the Premium flows, the Assured flows were allowed to inject large traffic bursts into the network. This was done by demoting the non-conforming packets to best effort. Thus, the Assured Marker acted as a policer leaving non-conforming packets to be discarded inside the network. The network routers do not keep per flow information but rather act on the aggregate differentiated traffic. At each output interface of a router, two priority queues were deployed. The arriving Premium packets were assigned to the high-priority queue, whereas the Assured and best-effort packets were assigned to the other queue. The former queue has a simple ``non-preemptive'' priority over the later queue. A pair of Random Early Discard (RED) buffer management mechanisms were implemented on the low-priority queue. One RED mechanism acted upon the Assured packets, while the other RED mechanism acted upon the best-effort packets. As in [Clark1], we refer to the twin RED mechanisms as RED with In and Out (RIO). At the input interface of an ingress border router, two profile meters were employed. One meter policed the arriving Premium packets, while the other meter policed the arriving Assured packets. The meters were token buckets configured with the aggregate Premium and Assured rates negotiated at the border, respectively. The function of the Premium token bucket was slightly different from the Assured token bucket. The non-conforming Premium packets were discarded immediately by the Premium token bucket, while the non-conforming Assured packets were reclassified and sent as best effort by the Assured token bucket. 3. Network Model We have simulated the end-to-end network shown in Figure 2. The network consists of three core routers each modeled as a node with two-level priority queues. The last node is also augmented with the border policing functions described before. At each node, three classes of traffic-flows compete for the network resources; the Premium flows, the Assured flows, and the best effort flows. At each node, the number of flows within each class is represented by N_P, N_A, and N_BE, respectively. For simplicity, we ignore the queueing delay at the leaf routers. For this, the only points of queuing/multiplexing are within the core routers. However, the leaf routers still reshape (in case of Premium) or police (in case of Assured) flows with the Markers described in the last section. Figure 2: (see pdf or ps form of document for all figures and tables). Naser, et. al. Expires: June 1999 [Page 4] INTERNET-DRAFT Differentiated Services Field December 1998 At the output of each core node, some 40% of the flows are exited, and the remaining flows sustain the end-to-end path either partially or completely. We choose to collect the statistics of two reference flows: a Premium flow and an Assured flow, both taking the same end-to-end path. There are two types of traffic inside the network. The traffic generated by Constant Bit Rate (CBR) sources, and the traffic generated by ON-OFF sources. Both sources have shown to be good models for generating simulated voice traffic, depending on the coding or shaping schemes used at the application level. In this study, the traffic sources generate constant size packets. Table 1 summarizes the sources' traffic descriptors used as our baseline study. Throughout the discussion, some of these parameters are changed in order to study their effects on the network performance. In this draft, unless otherwise stated, all parameters are represented in unit of bytes or bytes/second. In Table 1, the token bucket parameters represent values assigned to each flow's traffic Marker. The token bucket parameters of the ON-OFF source were obtained so that 90% of the generated packets were conforming. Thus, when the source generated Assured flow, 10% of the packets were reclassified as best effort by the corresponding Marker. Table 1: (see pdf or ps form of document for all figures and tables). The core links ran at 45Mbps (5.625MBps). The proportion of the link assigned to each differentiated class is 20% for the Premium, 40% for the Assured, and the remaining 40% for the best effort. For the Premium and Assured classes, the aggregate token rate of the border policers was set to the above proportions: the aggregate Premium token rate was set to 1.125MBps, whereas the Assured token rate was set to 2.25MBps. The bucket depth of the policers however varied for each simulation scenario; they were set to the sum of the bucket depth of all flows that reached the border. For instance, where there were N_P = 140 Premium CBR flows, and N_A = 703 Assured ON-OFF flows, the corresponding bucket depth of the border policers were 140 * 576 bytes (for Premium) and 703 * 3744 bytes (for Assured). The size of the premium queue at each core node was set to 30 packets. For this, the maximum queuing delay of the reference Premium packets (each of size 576 bytes) was bounded by 10 milliseconds. For the baseline case, where the non-Premium packets were also 576 bytes, the size of the RIO queue at each node was set to 650 packets. Assuming there is no Premium flow in the network, the maximum queuing delay of a reference Assured packet is bounded by 200 milliseconds. Throughout this work, we choose this delay bound in order to find the RIO queue size at each node. Naser, et. al. Expires: June 1999 [Page 5] INTERNET-DRAFT Differentiated Services Field December 1998 The RIO parameters were set as the following. Like [Ibanez], the minimum and maximum thresholds for the ``OUT'' packets were set to 40% and 80% of the total RIO queue size, respectively. The minimum and maximum thresholds for the ``IN'' packets were set to 45% and 80% of the total RIO queue size, respectively. The OUT and IN probabilities were set to 0.05 and 0.02, respectively. Finally, both OUT and IN queue weights were set to 0.004. Having said the RIO queue size for the baseline case, the simulated RIO parameters are listed as follows: 260/520/0.05/0.004 for OUT parameters, and 295/520/0.02/0.004 for IN parameters. 4. Delay Results The baseline case includes (N_A = 703, N_BE = 703) ON-OFF flows, and a varying number of CBR flows (N_P) at each node. This fixes the total Assured and best effort loads at each node to 0.36 and 0.44, respectively. Note that, 10% of the traffic generated by each Assured flow is best effort. For this, the observed Assured (best effort) load at each node is 10% less (more) than the corresponding link proportion assigned to that class (i.e, 0.4). Figure 3 illustrates the end-to-end delay performances obtained for the reference Premium (shown on the left side) and the reference Assured (shown on the right side) flows, as functions of N_P. The corresponding total load, averaged over the three core nodes, is shown on top of the Figure. The delay performances are the average and the alpha-percentile; where alpha = 99% for the Premium flow, and 95% for the Assured flow. At all load levels, the Premium delay performances are less than the corresponding Assured delay performances. This is especially true at the high load levels, where the difference reaches to two order of magnitude. Moreover, the Premium delay curves begin to rise at a lower load level than the Assured delay curves. One reason for this is that the Premium bandwidth is scarce (compared to the Assured bandwidth), thus, the Premium performance can easily be deteriorated when the Premium class is over-subscribed. Figure 3: (see pdf or ps form of document for all figures and tables). In terms of the playout delay performance, at the highest load level shown, 99% of the Premium packets see an end-to-end delay of less than 1.15 milliseconds. At this load level, however, 95% of the Assured packets see an end-to-end delay of less than 130 milliseconds--- still within the 200 milliseconds acceptable bound for the voice packets. At the lowest load level, the Assured playout delay is only 2.6 times greater than the Premium playout delay. Therefore, the user of the Assured service will have to allocate a much larger playout buffer at the destination than the Premium user will. Naser, et. al. Expires: June 1999 [Page 6] INTERNET-DRAFT Differentiated Services Field December 1998 One encouraging result is that the number of accepted voice calls (flows) is almost two times larger than the number of voice calls that is normally accepted at the DS3 telephone lines. Compare the total accepted number of flows here--- this is between 1441 and 1546 depending on the load level--- with 45Mbps/64kbps ~= 703 flows that can be accommodated at the DS3 rate. Even with this radical increase in the accepted traffic, the delay performances of the Assured flows stay within a reasonable few milliseconds range, as long as the network load remains below 95% (see Figure 3). Figure 4 illustrates the standard deviation of the same data shown in the previous Figure. As the premium load increases, the Assured packets tend to experience more and more delay-variation. In contrast, the delay-variation of the Premium packets begins to decrease after a certain load level. This is again attributed to the over-subscription problem described earlier. At high load the majority of Premium packets arrive late at the destination, whereas at low load, they arrive early. In either case the standard deviation is low. Figure 4: (see pdf or ps form of document for all figures and tables). In terms of the packet loss performance, we have not seen any Premium packet loss at the core routers. However, at the highest load level shown in Figure 3, the Assured packets were discarded by the RIO algorithm at the rate approximately equal to 1*10^(-06). The rate with which the best effort packets were discarded by the RIO algorithm was limited to 2*10^(-02). In Figure 5 we compare the effects of the Premium load with the Assured load on the performance parameters. We simulated the case where the Premium and Assured flows were all CBRs. The best effort flows were ON-OFFs as in the previous case. Each CBR generated packets at the rates shown in Table 1. Further more, all packets generated by an Assured CBR flow were considered conforming, as opposed to the previous scenario where 10% of the packets generated by any Assured flow were non-conforming. Figure 5: (see pdf or ps form of document for all figures and tables). On the left side of Figure 5, we show the delay performances (both Average and 95% Percentile) of the Assured reference flow, whereas on the right we show the maximum RIO buffer occupancy averaged over the three core nodes. As RIO is concerned, we took two measurements: the maximum Assured buffer occupancy, and the maximum total (i.e, Assured plus best effort) buffer occupancy. For the curves labeled ``PRL20'', the Premium load was fixed at 20% and the Assured load was varied from 25% to 40%. For the curves labeled ``ASL40'', the Assured load was fixed at 40% and the Premium load was varied from 5% to 20%. Naser, et. al. Expires: June 1999 [Page 7] INTERNET-DRAFT Differentiated Services Field December 1998 As Figure 5 shows, for the same total load, more Premium load (PRL20) results in an slight increase in the Assured delay performances, compared with the more Assured load case (ASL40). In terms of the RIO performance, at very high load levels (> 0.95), the PRL20 case creates slightly more congestion in the RIO buffer than the ASL40 case. At all other load levels the opposite is almost true. One reason for this is that the PRL20 case consistently utilizes ``all'' allocated premium bandwidth. This case generates more Premium packets than the ASL40 case. At a very high load, these Premium packets impede the flow of non-Premium packets at the core nodes. This results in a large RIO congestion when the network operates at high load, under the PRL20 case. We have also measured the rates at which the packets were discarded at the Premium and RIO buffers. At all load levels shown in Figure 5, there was no Premium or Assured packet loss inside the network. Only at the highest load level shown in the figure were the best effort packets discarded at the rates of 5.98*10^(-5) and 1.54*10^(-5) for the PRL20 case and ASL40 case, respectively. At other load levels, none of the best effort packets were lost. Returning to the delay performance of the Assured reference flow, the following interesting observation can be made. The delay performance--- either the average or the 95% Percentile--- of the ``CBR'' Assured flow (shown in Figure 5) is less than the corresponding delay performance of the ``ON-OFF'' Assured flow (shown in Figure 3). This is particularly true at the high network load, where the playout delay of the CBR Assured flow is bounded by 20 milliseconds, as compared to the 100 millisecond playout delay of the ON-OFF Assured flow. This means that when the Assured flows are smooth, the Assured service can yield a reasonably good delay performance. In fact, as long as the load remains below 95%, the playout delay of the Assured CBR flow is within a few millisecond, slightly larger than what is shown for the Premium CBR flow in Figure 3. 5. Effect of Shaping The premium flows are assumed to inject a very smooth traffic into the core network. For this, the Premium Markers at the leaf routers should reshape the user generated traffic. In this section, we study the effect of (re)shaping on the Premium and Assured flow performances. First, we describe the baseline scenario where N_P=140 Premium CBR flows were multiplexed with N_A=703 Assured and N_BE=703 best effort ON-OFF flows. This maintained the load at each node at nearly 100%. Regarding the amount of reshaping of the Premium flows at the leaf Naser, et. al. Expires: June 1999 [Page 8] INTERNET-DRAFT Differentiated Services Field December 1998 routers, we compared the following four cases shown in Table 2. With CBR-1, the CBR flows were essentially not reshaped. The token bucket parameters were chosen from Table 1: the token rate was equal to the CBR rate and the bucket depth was equal to the length of one CBR packet. These parameters were large enough to pass the CBR packets intact. With CBR-2, the CBR packets initially had a small jitter in their inter-arrival times. We chose this jitter to be uniformly distributed in [T-T/50, T+T/50] interval, where T is the period of a CBR flow. The corresponding bucket parameters were chosen as in the previous case. CBR-3 is the same as CBR-2 with the only difference being the bucket depth now chosen to be the size of two packets (i.e, 2*576 bytes). Finally, CBR-4 differs from CBR-2 in that the token rate of CBR-4 was chosen (1+2/50) larger than the token rate of CBR-2 in order to accommodate the initial jitter in the packet inter-arrivals. Table 2: (see pdf or ps form of document for all figures and tables). The delay performances of these four cases are shown in Table 2. For CBR-1, we show the 98.8% percentile, as the 95% percentile was not available. For this, we should assume a lower percentile value than is shown when we compare it with the percentile delays of other cases. With CBR-2, the delay performances are unbelievably high. This means that when CBR traffic is shaped, a small jitter in the packet inter-arrival times could lead to a large end-to-end delay and delay-variation. CBR-1, CBR-3, and CBR-4 essentially have the same average/percentile delay performances. The large Standard Deviation of CBR-3 or CBR-4 is greatly attributed to the initial jitter in the packet inter-arrival times. In the next simulation scenario, the Premium sources were also ON-OFFs. This scenario consisted of (N_P=351, N_A=703, N_BE=703) ON-OFF sources with the rates indicated in Table 1. The operating load at each node was therefore very close to 100%. Like an Assured Marker, each Premium Marker generated tokens at 5000 bytes/sec--- 37.5% below the maximum rate. Should the Premium bucket-depth be chosen small, some percentage of the Premium packets would be delayed at the Marker. For example, if the bucket-depth was chosen to be equal to the Assured bucket-depth (see Table 1), for 10% of the Premium packets there would not be enough tokens in the bucket, and thus, they must be delayed at the bucket. To further investigate the effect of shaping, we varied the Premium bucket-depth, normalized by the packet length (576 bytes). For each selected bucket-depth, we also adjusted the border Premium bucket-depth to be the sum of the bucket-depth of all Premium flows reaching the border. Naser, et. al. Expires: June 1999 [Page 9] INTERNET-DRAFT Differentiated Services Field December 1998 Figure 6 shows the delay performances (i.e, the average and the 95% percentile) obtained for the Premium and the Assured reference packets, as functions of the Premium (reshaper) bucket depth. The Figure conveys a clear deterioration in the (delay) performances of the Premium flows. When the Premium flows are hard limited their delay performances are quite high--- in fact as high as a couple of seconds! The Premium delay values only begin to fall when the Premium bucket depth is chosen larger than the size of ten packets. This number is equal to the average number of back-to-back packets generated by a Premium ON-OFF source during the ON interval. The Premium bucket depth must be chosen even larger than ten packets (i.e, between 15 to 20 packets), in order to allow the delay performances of the Premium flow to reach the Assured flow performances. Figure 6: (see pdf or ps form of document for all figures and tables). There is a transition phase in Figure 6 where the Premium delay performances drop from their unacceptably high values to the reasonably low values (in the range of one or a few milliseconds). After the transition phase, the Premium delay performances do not change significantly. This is quite interesting, because, we allow the Premium flows ``not'' to be reshaped where the flow's bucket depth is chosen to be 300 packet-size (the rightmost point of Figure 6). Therefore, in the case of ON-OFF sources studied here, the shaping does not result in an acceptable delay performance. Now, let us investigate other issues involving the shaping. Figure 7 shows the number of non-conforming Premium packets that arrived at the border policer. Again, the variable is the Premium (reshaper) bucket depth. The simulation setup is as explained earlier. When the Premium flows are shaped hard, the number of non-conforming Premium packets arriving at the border grows. This is firstly due to the large shaping jitter, and secondly to the small (aggregate) bucket size chosen for the border policer when the Premium flows are smooth. For example, this bucket size is equal to 576*351 bytes when each Premium flow bucket size is 576 bytes, whereas it is 4*576*351 bytes when the individual Premium flow buckets are 4*576. In the former case, the small bucket at the border cannot accommodate as many Premium packets as the later case can. Figure 7: (see pdf or ps form of document for all figures and tables). Finally, Figure 8 shows the aggregate rate at which the Assured and Best-effort packets are discarded at all three RIO queues. For the best-effort packets, this rate remains essentially constant (~= 2.X * 10^(-02)), when the Premium flow burstiness is changed. The Assured loss rate is however varied, but, still remains within the order of 10^(-06). Naser, et. al. Expires: June 1999 [Page 10] INTERNET-DRAFT Differentiated Services Field December 1998 Figure 8: (see pdf or ps form of document for all figures and tables). 6. Large Premium Allocation So far, the reserved bandwidth for the Premium class was kept as 20%. In some cases, though, this bandwidth was not fully utilized. In this section, we study different bandwidth reservations for the Premium class. We varied the reserved bandwidth for this class from 10% to 50%. The reserved Assured bandwidth was kept to 40%. Each class fully utilized its reserved bandwidth, while the remaining bandwidth was occupied by the best effort traffic. We used the baseline case where the Premium flows were CBR types, and the non-Premium flows were ON-OFF types. The observed delay performances of the Premium reference flow were quite encouraging. While the average delay performance varied from 0.88 millisecond to 1.06 millisecond the 98% percentile varied from 0.93 millisecond to 1.18 millisecond, when the Premium reservation changed from 10% to 50%. For the Assured reference flow, the observed average and 95%-percentile were in the [55, 100] and [107, 200] milliseconds, respectively. Note that the 200 millisecond delay bound (set by us) for the Assured packets was only reached when the Premium reservation was 50%. 7. Future work Currently, we are studying the admission control algorithms for the combined differentiated services architecture. We anticipate that the measurement-based algorithms work better than the parameter-based algorithms, in the differentiated services environment. This is due to the fact that with the Premium and Assured services, the flows are characterized by only one (or two) parameter(s). A measurement-based admission control mimics this inaccuracy by providing the real dynamics of the network. Nevertheless, our preliminary results have shown an excessive packet loss at nodes where admission control was not exercised. Naser, et. al. Expires: June 1999 [Page 11] INTERNET-DRAFT Differentiated Services Field December 1998 8. References [2BIT] K. Nichols, V. Jacobson, and L. Zhang, "A Two-bit Differentiated Services Architecture for the Internet", Internet Draft , November 1997, ftp://ftp.ee.lbl.gov/papers/dsarch.pdf [Clark1] D. Clark and J. Wroclawski, "An Approach to Service Allocation in the Internet", Internet Draft, draft-clark-diff-svc-alloc-00.txt http://diffserv.lcs.mit.edu/Drafts/draft-clark-diff-svc-alloc-00.txt [Clark2] D. Clark and W. Fang, "Explicit Allocation of Best Effort Packet Delivery Service", MIT Lab for Computer Science, http://diffserv.lcs.mit.edu/Papers/exp-alloc-ddc-wf.ps [Ibanez] J. Ibanez and K. Nichols, "Preliminary Simulation Evaluation of an Assured Service", Internet Draft, draft-ibanez-diffserv-assured-eval-00.txt 9. Author's Addresses Hassan Naser Department of Electrical and Computer Engineering University of Toronto Toronto, ON M5S 3G4, Canada Phone: (416) 978- 6876 Email: hnaser@nal.utoronto.ca Alberto Leon-Garcia Department of Electrical and Computer Engineering University of Toronto Toronto, ON M5S 3G4, Canada Phone: (416) 978- 4766 Email: alg@nal.utoronto.ca Osama Aboul-Magd Nortel Networks P.O. Box 3511, Station C Ottawa, ON K1Y 4H7, Canada Email: osama@nortel.ca Naser, et. al. Expires: June 1999 [Page 12]