Packet Coloring for Intelligent Traffic Management In Wireless Networks

Transcription

1 Packet Coloring for Intelligent Traffic Management In Wireless Networks Doru Calin, Bongho Kim Nokia Murray Hill NJ, USA Francis Bourriaud Nokia Villarceaux, France Abstract- Delivering committed traffic Quality of Service (QoS) per Service Level Agreement (SLA) requires support of critical intelligent traffic management features in LTE enodeb (and future 5G access nodes), such as Uplink Traffic Shaping and Congestion and Admission Control (CAC). This paper highlights the motivation and merits of uplink traffic shaping in preserving QoS levels per traffic categories over transport links. Techniques for packet coloring, which are used in the fixed networks, are considered herein to mark packet priorities within the wireless access nodes. In particular, the paper introduces an algorithm for intelligent packet coloring, with the goal to protect the critical traffic while maximizing the return of investments made by network operators. Keywords- Uplink traffic shaping, CAC, Packet coloring, LTE. I. INTRODUCTION Dealing with various applications (e.g., voice, video, Internet gaming, chatting and file transfer) calls for intelligent mechanisms to fulfill their often opposite QoS requirements. For instance, real time services, such as voice, are very sensitive to delay and jitter impairments caused by congestion in the network, while video requires a minimum guaranteed rate. On the other hand, the mix of applications must be supported at minimum cost. Wireless access nodes are bridging the air-interface and the backhaul domains. They must be equipped with mechanisms to control the burstiness of traffic and to secure the relative priority across the mix of traffic flows. While the priority can be guaranteed within the mobile network through appropriate scheduling policies, it remains a challenge to enforce such priority end to end, especially when packets must flow outside the mobile packet core (e.g., through Internet). This is because, in case of congestion along the transmission path, some packets may be either dropped or delayed unacceptably long. Worse, while employing state of the art techniques, packets pertaining to critical traffic flows may be randomly dropped in such conditions, as any other less critical packets. The paper introduces and discusses efficient methods to solve these problems. We are discussing scalable, adaptive, and economically viable techniques of traffic management for the access nodes of the future broadband wireless networks. This relates to intelligent traffic shaping and congestion admission and control within a multi-operator and Service Level Agreement (SLA) context, to ensure that the quality can be controlled per user per application while keeping the overall bandwidth requirements and cost to the minimum. Such aspects are at the crossing boundary between wireless and transport domains of a network, and as such, they are not falling neither under the jurisdiction of the Third- Generation Partnership Project (3GPP) [5] nor under the Metro Ethernet Forum (MEF) [1, 2] umbrella. The undesirable consequence is that the subject is often overlooked, with negative implications on the ability to implement appropriate levels of QoS. There is no surprise that the prior art for that matter is fairly limited. The paper provides an educational perspective on the matter, while articulating the need for intelligent traffic control mechanisms to preserve the QoS network wide. Its organization is as follows: section II describes challenges that must be overcome to enforce QoS end to end in wireless networks; section III is devoted to intelligent traffic management mechanisms introduced through this work and section IV follows through with a quantitative analysis via network simulations to evaluate the merits of the algorithms introduced in section III. The main key takeaways are finally stated in section V. II. CHALLENGES OF PROVIDING END-TO-END QOS IN WIRELESS NETWORKS Future 5G networks will have the ability to provide support for QoS that is superior to today s 4G networks. To make the arguments of this paper, we will refer to the Long Term Evolution (LTE) [3] technology as the reference to QoS support over the air. LTE enodebs are meant to be equipped with some level of traffic management features. We are referring to Uplink Traffic Shaping and CAC within the scope of this paper. It is anticipated that such intelligent traffic mechanisms are even more necessary in the future 5G networks. Uplink Traffic Shaping: Shaping is a method to reduce burstiness of traffic. When done in the enodeb, it is meant to increase the level of uplink bandwidth profile compliance. It is recommended that the enodeb shapes the traffic sent into the Metropolitan Ethernet Network (MEN), so that the shaper matches the Committed Information Rate (CIR), Committed Burst Size (CBS), Excess Information Rate (EIR) and Excess Burst Size (EBS) parameters [1, 2] of the appropriate /17/$ IEEE

2 bandwidth profile. These parameters are further defined in section II.B. Figure 1 reveals a simple representation of the main building blocks of a traffic shaper, consisting of a traffic classifier, which maps multiple traffic applications to different queues based on their relative priority levels, and a scheduler. CIR and EIR metrics. Operator B may support several mobile operators simultaneously; that means that the backhaul managed by the Operator B may be shared by multiple operators such as Operator A. Furthermore, mobile Operator A may employ advanced mechanisms of QoS classification to provide differentiation across its traffic categories (e.g., through QCIs in LTE). Hence, SLA creates an explicit translation from a fine granular definition of QoS - as used by the Operator A (e.g., 9 levels of QCIs exercised eventually in LTE), to a much coarser QoS definition, with only two levels (CIR, EIR) - as used by the backhaul Operator B. Such downsizing in QoS resolution across the transmission path may cause undesirable loss of packets. Figure 1: Example of uplink traffic shaping in LTE enodeb. Congestion and Admission Control: CAC is a preventive load control, which safeguards the QoS of existing Radio Access Bearers (RABs) by refusing additional RABs when necessary. CAC is an open loop control, which does not modify the characteristics of RABs once they are accepted, and so the future consequences of a RAB should be anticipated at connection set-up. In brief, CAC shall ensure that there are enough network resources for a new RAB to be admitted and must operate jointly over the radio as well as over the transport links. Committed Information Rate: CIR corresponds to the effective committed bandwidth for which an Operator is willing to pay the highest price per bit. As the Operator needs exclusive access to the bandwidth corresponding to CIR, the price per information bit carried out through CIR is the highest. Excess Information Rate: EIR corresponds to the additional bandwidth that an Operator could use, but has unguaranteed access to; it is because such additional bandwidth can be temporarily not available (e.g., best effort mode usage in a shared resource environment). Hence, the price per information bit carried out through EIR is lower compared to the equivalent information bit carried out through CIR. II.A Cross-Operator Service Level Agreement CIR and EIR are (typically) configured through binding Service Level Agreements (SLA) across Operators/Service Providers managing different network domains. Note that a Service Provider may choose to operate only based on CIR. Figure 2 illustrates an example of SLA between a mobile operator (Operator A) and a backhaul operator/service provider (Operator B). The backhaul operator can use a combination of technologies such as cable, microwave transmissions, or optical fiber. The Operator B must honor the incoming traffic from the Operator A, as agreed through the CIR and EIR metrics. For this, the edge routers contain policing functions, which control the traffic according to the Figure 2: Example of cross-operator SLA. II.B Packet Coloring Techniques In terms defined by the Metropolitan Ethernet Forum, traffic shaping may be either colored or color-blind. Note that such aspects are not covered by the 3GPP standardization. When packet coloring is enabled, one allows a reference operator (e.g., Operator A) to mark its packets with differentiating colors, which indicate different priority levels. This is important in case of traffic congestion conditions in the network, as an IP packet router would be able to process the colored packets according to the priority levels that are universally understood. A single rate shaper is defined by the parameters (CIR, CBS*, CBS), and the definition of the shaper s metrics in the presence of packet coloring is as following: CIR: the average shaping rate of Green frames; this is the average output rate of the shaper. CBS: the shaping burst of Green frames; this is the maximum output burst of the shaper. CBS*: the accepted burst of Green frames; this is the maximum buffer size for Green frames. CBS* CBS, which means that the shaper accepts larger bursts at its input and generates smaller bursts at its output. A dual rate shaper is defined by the parameters (CIR, CBS*, CBS, EIR, EBS*, EBS), where (CIR, CBS*, CBS) are defined as for the single rate shaper and the other metrics are defined as follows: EIR: the average shaping rate of Yellow frames; this is the average output rate of the shaper.

3 EBS: the shaping burst of Yellow frames; this is the maximum output burst of the shaper. EBS*: the accepted burst of Yellow frames; this is the maximum buffer size for Yellow frames. provides services to two mobile operators, Operator A and Operator C, via SLA_A defined by the parameters (CIR_A, EIR_A) and SLA_C defined by the parameters (CIR_C, EIR_C), respectively. Figure 3 represents a dual rate traffic shaper with packet coloring enabled [1]. All the packets which find the buffer size under the CBS* threshold upon their arrival are colored in green. Packets which find the buffer size over the CBS* threshold, but under a threshold (THS), are colored in yellow. The THS is a threshold which defines the maximum storage size in the traffic shaper, so any packet that arrives when the buffer is already filled up at its maximum storage capacity, THS, is dropped out upon arrival. Figure 4: Example of dual rate traffic shaper with packet coloring (potential wrong packet coloring of the critical traffic flows). Figure 3: Example of dual rate shaper with packet coloring. Figure 4 illustrates the impact of a dual rate traffic shaper that receives traffic from two applications: voice, which has stringent real time constraints, and best effort data. As long as the aggregate traffic volume is under the CBS* threshold, the corresponding packets are shaped at the CIR rate using the train of bursts of CBS size, regardless of their application QoS requirements. On the other hand, as soon as the traffic volume starts exceeding the CBS* threshold, an additional credit running at the EIR rate is activated and the corresponding packets are shaped using the train of bursts of EBS size, and with a frequency of repetition as determined by the EIR and EBS. If packet coloring is supported, packets are colored following the representation in Figure 3 from above, i.e. green and yellow with respect to the CBS* threshold. In this figure, we have assumed some numerical values: CIR = 6 Mbps, EIR = 24 Mbps, CBS = 64 KB and EBS = 16 KB. Note that we have intentionally identified a potential issue with the packet coloring, as illustrated in Figure 4. Indeed, with the standard definition provided by Metro Ethernet Forum, depending on the traffic dynamics, some low priority data packets may end marked as green, while some high priority voice packets may at risk of being marked in yellow. The later causes coloring errors, which may lead to undesirable packet losses in case of network congestion, as represented in Figure 5. In this case, the backhaul Operator B Figure 5 shows a router in the Operator B network that experiences temporal congestion due to temporal traffic overload. This may well happen, since the Operator B has only committed to accommodate CIR_A and CIR_C, and anything else that comes on top through EIR_A and EIR_C is treated as best effort. Hence, the Operator B has sufficient bandwidth provisioned to accommodate the traffic supplied simultaneously through CIR_A and CIR_C, but it is not required to fully accommodate the traffic in excess of the committed rate. Hence, if the router in question does not have enough storage capacity to accommodate all the incoming traffic, it will drop packets that are marked in yellow and coming through the credit lines EIR_A and EIR_C. Figure 5: Example of multi-operators SLA where packets may get dropped due to congestion.

4 Note that in the absence of packet coloring (or packet priority marking and explicit packet inspection while in routing, in more general terms), one cannot differentiate among traffic pertaining to different classes. Thus, in case of congestion as pointed out in Figure 5, one cannot ensure any protection for the high priority traffic over the low priority traffic. Nevertheless, the importance of packet coloring technique as described in this section, with the scope to enforce QoS, is often either ignored or not well understood. This makes the argument that one needs to enable smart packet coloring in the wireless access nodes, such as enodebs. More specifically, one should protect the packets pertaining to the critical traffic from being wrongly colored, and to avoid undesirable packet dropping. III. INTELLIGENT TRAFFIC MANAGEMENT MECHANISMS FOR FUTURE WIRELESS NETWORKS In this section, we introduce an algorithm to eliminate the risk of dropping packets associated with critical traffic flows, for which an Operator having an SLA with another Operator has reserved and provisioned bandwidth through a CIR. The solution consists of the following key concepts: Assume that the buffer size of the shaper for the committed traffic is CBS*. The shaper can have a total buffer size larger than CBS* (e.g., THS >= CBS*), that is used to store packets coming from all traffic flows. Continuously estimate X, the maximum volume of traffic associated with the critical flows that would need to be stored into the shaper s buffer at any given time. This is the maximum instantaneous storage capacity in the shaper that would be required by the critical traffic flows at any time instant. As an example, if VoIP is the critical traffic flow, and each VoIP packet accounts for Nv Bytes, the maximum storage volume required is the maximum number of VoIP packets (e.g., Np) that may need to be stored at a given time, times, Nv, and is measured in Bytes (X = Np x Nv Bytes). If there are multiple such flows associated with the critical traffic that is committed through the CIR, then X is the sum across all the critical traffic flows, and indicates the maximum traffic volume generated by all the critical flows that would need to be stored at any given time. Note that X may vary over time so 0 X CBS*. In the shaper s buffer, reserve a storage capacity amounting to the estimated value X, to be solely used for the critical traffic flows committed through the CIR. This traffic will be marked in green. At any given time, allow a new packet originated by the non-critical traffic flows to make use of the CIR and hence be marked in green, if and only if there is enough instantaneous storage capacity available for the noncritical traffic flows. As an example, assuming best effort data as the non-critical traffic flow and that a data packet accounts for Nd Bytes, a newly originated data packet can be treated with the CIR credit and marked in green if and only if Nd (CBS* - X). Otherwise, the same newly originated data packet must be treated with the EIR credit and it will be marked in yellow. Figure 6 illustrates an example of the proposed dual rate traffic shaping algorithm, which eliminates the wrong packet coloring for the critical traffic flows. Similarly to Figure 4, the shaper receives traffic from two applications: voice and best effort data. A fraction of the shaper s buffer size amounting to the volume X is reserved for the critical traffic flow (voice). The rest of the buffer can be used by both traffic flows. As a result, voice packets are colored in green, and if there is sufficient CIR credit available, some data packets can be colored in green as well (this in particular helps the Operator fully use its committed rate). On the other hand, the method prevents critical voice packets from being colored in yellow. In this figure we have assumed some numerical values: CIR = 6 Mbps, EIR = 24 Mbps, CBS = 64 KB and EBS = 16 KB. Figure 6: Example of dual rate traffic shaper with packet coloring (potential wrong packet coloring of the critical traffic flows eliminated). IV. QUANTITATIVE ANALYSIS VIA NETWORK SIMULATION The algorithm was tested via a network simulation which was created on an OPNET platform [4] using two applications: Voice over IP (VoIP) and Best Effort (BE) data. All simulation results are represented as a function of time. As part of the performance evaluation, we have considered the following parameters for traffic shaping: CIR = 5.35 Mbps, CBS = CBS* = 64 KB EIR = Mbps, EBS = 64 KB, EBS* = 0 PIR = CIR + EIR = 30 Mbps Note that PIR stands for Peak Information Rate. Traffic generation was performed as follows:

5 VoIP traffic is the delay sensitive critical traffic and amounts to a volume of 5.2 Mbps, as represented in Figure 7.a. The BE traffic is generated via UDP traffic and amounts to 30 Mbps at its peak; it is generated through bursts, as illustrated in Figure 7.a. Figure 7.b represents the total input traffic to the shaper, combining VoIP and BE traffic and can reach 35 Mbps at the peak, which is purposely set above the allowed PIR at the output of the shaper. CIR and EIR credits while there are no VoIP packets waiting in the system, resulting in both green and yellow colored BE packets. If the CIR is not temporarily available upon the arrival of some VoIP packets, those VoIP packets will be transmitted through the EIR credit and end up colored in yellow. Figure 7.a Figure 7.b Figure 7: Example of traffic input to shaper: (a) individual VoIP and BE traffic, (b) total combined VoIP and BE traffic. Figure 8 illustrates the effect of the traffic shaping. It contrasts the total traffic at the input of the shaper against the shaped traffic at the output of the shaper. The shaper caps the total traffic at the allowed PIR (30 Mbps). Figure 9.a Figure 9.b Figure 9: Example of packet coloring outcome using standard MEF technique (as in Section II): (a) VoIP traffic, (b) BE traffic. Results produced in Figure 10 are the outcome of the algorithm described in Section III, using X = 2 KB reserved for the critical VoIP traffic out of the total CBS* = 64 KB. Figure 10.a Figure 10.b Figure 10: Example packet coloring outcome using the proposed algorithm (as in Section III): (a) VoIP traffic, (b) BE traffic. Figure 8: Example of traffic shaping: input and output traffic at the shaper. Figure 9 shows colored decomposition of the total shaped VoIP and BE traffic using the packet coloring technique described in Section II. It is remarkable that although the total VoIP traffic amount is under the CIR, still a large fraction of VoIP packets end up colored in yellow (Figure 9.a). On the other hand, there are also BE packets that are colored in green (Figure 9.b). This is because the BE traffic consumes both All the VoIP packets are colored in green this time (no more packet coloring error for the VoIP packet). Data packets still have access to the CIR credit line, and when it happens they get colored green. This enables the Operator to fully use its CIR credit, which offers an implicit advantage to the best effort packets in the event of congestion along their end to end transmission path. In such cases, the best effort packets that are colored in green will have a better chance to go through each time a router along the transmission path needs to drop packets.

6 V. CONCLUSION The article motivates the necessity to design for highly scalable and intelligent traffic management algorithms within the wireless access nodes (e.g., enodeb in LTE). The advanced radio resource management and intelligent airinterface scheduling mechanisms, which are integral components of the modern broadband wireless communication systems, are not sufficient to guarantee that traffic can be delivered with desirable quality of service and quality of experience for the end users, especially for delay sensitive applications such as voice and video. To this extent, we have introduced the uplink traffic shaping and CAC, as essential mechanisms to ensure that any potential congestion in the backhaul links will not create undesirable services disruption. The uplink traffic shaping ensures the following roles: Support priority enforcement across classes of services, in particular support for delay sensitive applications (real time traffic). Differentiate among Guaranteed Bit Rate (GBR) traffic (e.g., voice and video). Maintain the quality for GBR traffic independently of the Transmission Control Protocol (TCP) configuration in the network (which is not under Operator s control in general). CAC is a mechanism which determines the fraction of the bandwidth that is allocated for the GBR service and thus protects the GBR traffic from the Non-GBR traffic. [3] S. Sessia, I. Toufik, and M. Backer (eds.), LTE -- The UMTS Long Term Evolution: From Theory to Practice, John Wiley & Sons, Chichester, West Sussex, UK, Hoboken, NJ, [4] Riverbed, [5] Furthermore, the paper introduces an algorithm which eliminates the risk of dropping packets associated with critical traffic flows for which an Operator having an SLA with another Operator has reserved and provisioned bandwidth through CIR and EIR metrics. More specifically, the algorithm controls the packet coloring in the wireless access nodes, such as enodebs, so that: It protects the packets pertaining to the critical traffic from being wrongly colored (in yellow) and so it avoids undesirable packet dropping. It maximizes the number of packets pertaining to the non-critical traffic flows that can be opportunistically colored in green (the color used for the committed traffic). This in turn helps minimize the number of packets pertaining to the non-critical traffic flows that are at risk to be dropped in case of congestion along the transmission path in the network. Future research work will address more complex traffic shapers that are suitable for scenarios with multiple operators sharing the network infrastructure. REFERENCES [1] Metro Ethernet Forum, Technical Specification MEF 10.3 Ethernet Services Attributes Phase 3, October [2] Metro Ethernet Forum, Technical Specification MEF 41, Generic Token Bucket Algorithm, October 2013.