

HARDWARE IMPLEMENTATION OF MULTIMEDIA DRIVEN HFC MAC PROTOCOL

H.-C. Leligou, J. Sifnaios, G. Pikrammenos*

Abstract--Tree-shaped topologies present attractive cost advantages for broadband access networks by allowing many customers to share the expensive head-end equipment and the feeder section and provide a graceful upgrade path towards the photonization of the local loop. In addition they offer reuse of the copper last drops to the customer at least during the crucial introductory phase and probably for many years to come. Typical examples are Hybrid Fiber Coaxial (HFC) and Passive Optical Network (PON) systems. The TDMA multiplexing of traffic entering such a system is governed by the MAC protocol, which arbitrates the allocation of bandwidth to the shared feeder. At the same time the need to integrate telecom services presenting different quality requirements with plain best effort services over the same infrastructure brings new issues to the design of such an access mechanism. The MAC protocol as the only arbiter of the upstream bandwidth directly affects the QoS provided to each upstream traffic flow and must meet several constraints. Such constraints include the adequate speed of operation exploiting in the highest degree the speed of H/W implementation, flexibility to support efficiently the largest number of services and applications offering an adequate number of QoS classes and independence of higher layers, protocols and future extensions to traffic management specifications. The implementation of a MAC protocol targeting these goals in the framework of the AROMA research system is presented in this paper. We discuss the details and the implementation cost of the solutions followed and we evaluate the implemented mechanisms using computer simulation.

Index terms—HFC, MAC, QoS, FPGA controller

I.INTRODUCTION

Hybrid Fiber/Coaxial (HFC) systems, re-using the existing infrastructure of CATV systems and their coaxial medium beyond the fiber node, reduce the cost of the transition to a broadband access network offering a graceful solution in the cost-sensitive residential access domain. In such shared access systems, sharing the upstream channel (from customer to network) is usually effected through TDMA multiplexing. From a traffic engineering point of view, the downstream direction exhibits a broadcast nature while in the upstream, the traffic from all active modems is multiplexed/ concentrated as in the output port of any switching node. To arbitrate the access to the slotted common medium according to the demand, a Medium Access Control (MAC) has to be employed.

Figure 1. HFC system

Bandwidth allocation to the system users may be static, based on a pre-assigned unsolicited number of upstream time slots, or on demand, based on a request-grant exchange protocol. Different standard organization [4], [5], [6] have proposed several mechanisms to support all possible bandwidth allocation schemes trying to optimize the trade-off between the flexibility and performance, and the overhead and complexity of the MAC protocol. The allocation of upstream slots in the AROMA system is based on a reservation method which allows to dynamically adapt the bandwidth distribution to traffic fluctuations. The implementation of mini-slots [8] in the physical layer in order to efficiently support contention mechanisms was not followed, since the focus of the project was in providing QoS guarantees and predictable performance to differentiated services. Contention mechanisms for access to the upstream channel even in mini-slots cannot guarantee worst case performance and could optionally be employed as a companion mechanism for not time critical services [7]. Polling of requests was

instead employed providing the opportunity to trade-off utilization for low reaction time and controlled delay. The MAC controller collects access requests and then allocates the upstream slots by sending access permits. This results in a distributed queuing system characterized by the long time required to pass control information from the queuing points to the service controller residing at the head-end. It is important to note that the service policy of the MAC governs the distributed multiplexing from a central point situated miles away at the head-end. This has the important implication that, as regards delay control, the acquisition of arrival information and buffer fill levels includes a considerable delay element not found in the centralized multiplexing typically operating in a switch queuing point. In contrast, drop policies must be distributed over all network terminations (cable modems) where the flow identity and fill levels are known.

The need to differentiate the handling of traffic classes became apparent by the emergence of requirements for ISPs to offer QoS to certain flow in the presence of disturbing plain best-effort traffic without complex maintenance of state information in the core routers. It was obvious that it was not acceptable to mix delay-sensitive and best-effort services without unacceptable loss of grade of service for the former class, which employed preventive flow control. In contrast TCP/IP traffic uses flow control on the transport layer based on the detection of losses. Thus a TCP flow increases gradually the rate submitted to the network up to the point that buffer overflow occurs and re-transmission is triggered. This action will disturb the other traffic classes, which can not benefit from re-transmission.

As the MAC protocol requires a significant hardware part, a flexible strategy that can deal with behavior aggregates, guarantee quality to demanding traffic and reduce complexity is needed. The basis of the approach presented is the use of access priorities in the reservation system. Logically separate queues for each priority connection are necessary in the cable modem side for the proper operation of the prioritization scheme.

Services with a similar set of requirements can be grouped into 4 behaviour aggregates (classes) supported by 4 different priority connection:

  • All delay sensitive periodic Constant Bit Rate traffic is treated as the highest priority traffic (1st priority) and is supported by unsolicited permits scheduled during the connection establishment stage leaving the three lower priorities employ the reservation scheme.
  • Real-time Variable Bit Rate flows, such as video services are grouped into the 2nd priority of the presented algorithm. The 2nd priority requests are inspected first, preventing a possible disturbance of the two lower priority traffic that also employ the reservation scheme. At the same time, a peak-rate policing mechanism guarantees that malicious users cannot disturb the system performance.
  • The third priority traffic class is suitable for services requiring a better than best-effort behaviour. A minimum cell rate is guaranteed for this priority, while traffic exceeding this rate is relegated to the 4th priority.
  • Plain best-effort services are grouped into the 4th priority, relying their rate adaptation to their loss based flow control mechanisms at the TCP level.

Prioritizing the different requirements connection classes, QoS can be guaranteed to the services that need it and TCP/IP services are supported taking advantage of the statistical bandwidth multiplexing without disrupting other services. In other words, if we did not have the lower class we would have either to accept less demanding- quality traffic or leave the excess bandwidth unutilized in an effort to avoid performance degradation during statistical extremes. It is the availability of closed loop congestion control of TCP that allows us to offer both guaranteed performance to demanding traffic and good system resource utilization, at the same time. This results to a bigger number of services support and a more efficient bandwidth exploitation. The presented approach is in line with both the different classes of service according to the ATM forum philosophy and the Differentiated Services strategy of the IETF [1], [2],[3].

II.Implemenation of the MAC Controller

The MAC controller was implemented on the Access Network Adaptation (ANA) board of the AROMA system. Since the MAC layer is considered a sub-layer of the physical layer and is closely related to the physical layer functions (TDMA, framing, etc.), the permit generation function was selected to be integrated with the downstream framer component, which was responsible for the construction of the downstream frame including the transmission permits. The whole design was programmed and placed on a Field Programmable Gate Array (FPGA) chip. This decision made possible the re-design of the implementation to adjust to modified frame structures, rates etc. An embedded processor (On Board Controller) was used to calculate and modify all non-real time parameters such as the programmed (pre-allocated or provisioned) bandwidth distribution, which may vary with time due to the switch on or off of the modems. An external static RAM chip was used for buffering necessary protocol information in addition to the available on-chip memory provided by the FPGA, which was exploited to keep state information and speed-up the permit generation process. The MAC algorithm works as follows:

The OBC executes the Call Acceptance Control (CAC) scheme and prepares a list of 512 permits as well as a credit list. The 1st and the 2nd priority connection service at the constant and minimum bit rate respectively is supported by pre-allocated unsolicited permits written by the OBC at the time of the subscription in the permit list. The modems that have established such priority connections do not need to be polled since the programmed permits will enable the announcement of new arrivals at any queue. Polling permits are inserted for modems that have not established a 1st or 2nd priority connection. A variable number of positions in the permit list are left empty. Techniques to space the permits in the list are given in [9]. At the same time to support the 2nd priority connection service up to the peak-rate and to guarantee that the 3rd priority connections will achieve their guarantied minimum rate, credits are allocated and grants will be issued on request basis. Both the credit and the permit lists are stored on the external SRAM chip. Since ATM signaling is also supported in AROMA, the permit and credit lists can be updated dynamically to add new connections using a second copy.

Figure 2: MAC design

An empty position in the list (representing unallocated bandwidth) enables the dynamic service of the three lower priorities, which is based on requests. The requests per queue/priority arrive at the MAC controller in the head-end piggy backed in the upstream transmitted cells. The 1-byte wide piggy back field has been segregated in 3 parts, to offer the chance to all the three lower priority queues to announce the arrival of new cells, albeit with a delay equal to the time for the propagation of the information. This information is used to increment the 2 outstanding request totals for each modem and queue, which are stored in the 16 bit wide external SRAM creating a reflection of the corresponding buffer fill levels. Due to the employed round-robin service strategy, the piggybacked request buffer should be accessed frequently and traversed in order to find the next permit that will be scheduled. To avoid a redundant number of accesses to the external SRAM, which is time-consuming, and to expedite the search process, on chip flags are used to quickly detect and skip all empty locations.

The permit generation process results in the construction of a field pre-pended in each downstream frame, which includes all the grants for the corresponding time interval. The permit list is cyclically read out by the MAC controller H/W. The scheduled permits are inserted in the permit field and the empty positions are filled with permits produced by an engine scanning in a round robin fashion the outstanding request counters and reducing them by one for each permit scheduled. The higher priority counters are inspected first and only if all are empty the same process is repeated for the immediately lower priority. The permit generation is subject to some additional checks and guarantees by force of a companion credit check process. The MAC H/W subtracts the credits as it issues permits and stops serving any modem queue that exceeded its allocated apportionment. This policing action guarantees that malicious users can not disturb complying traffic exceeding their contracted peak cell rate. The credit check is used differently in the 3rd priority: it guaranties a minimum cell rate service, since a programmed number of credits guaranties the existence of the corresponding unallocated bandwidth. When 3rd priority credits are exhausted, left-over requests are added to the 4th priority ones and treated as plain best-effort traffic request, in accordance with AF rules. The operation of the permit scheduler and the execution of the MAC algorithm are depicted in Figure 2.

It worth noting that the permits sent downstream indicate only the modem and not the queue that is allowed to transmit. The modem is responsible for selecting the appropriate queue for transmission in an increasing order of priority, so that the higher priority traffic achieve a better performance in terms of delay.

III.Performance evaluation

For the evaluation of the above MAC mechanisms a computer simulation model was created using the PTOLEMY platform for modeling the AROMA MAC function under certain loading conditions. The scenarios used 10 cable modems loaded with uniform traffic for each priority. Each source used a common ON-OFF model, generating traffic at the slot/cell level with exponentially distributed on and off periods. The peak rate was the system rate simulating the arrival of IP packets with geometrically distributed lengths, which were instantaneously segmented into slots.

The 10% of the bandwidth was provided by the list as necessary for the polling of new bursts and the rest equally distributed among the priorities and modems in a way that the total was the one shown on the horizontal axis.

The 1st priority is not presented since it exhibits deterministic behaviour, as the transmission permits are pre-programmed in the list [7], and thus, the delay never exceeds the fixed permit distance. The most important result caused by the prioritization scheme is the delay advantage provided to the higher priorities. The transmission of almost all the 2nd priority cells delay less than 250 slots (i.e. 43ms) while those of the 3rd less than 350 slots (60ms). Of course there is no bound for the 4th priority which can exceed any limit depending on the total loading. In real-life though, this would cause no harm to the TCP-based user applications, because the rates of the sources would adapt to the possible bottleneck sharing the spare capacity equally among them due to TCP layer congestion control tools.

Figure 3: Mean delay vs. total load

At the same figure we can observe the behavior of the system when no priorities are available except the unsolicited permits of the 1st priority, i.e. the sources which were used for the three other classes are all feeding the same queue, marked “all” in figure 3. The total load is the same with 10% used for the 1st priority in all cases. The benefits of prioritization are clear since no traffic class can enjoy bounded delay in this case. In contrast, when priorities are enforced the 2nd and the 3rd classes enjoy a seemingly lightly loaded medium. Only the last priority sees a performance reduction, which however is well equipped to handle using the TCP layer congestion control tools.

IV.Conclusion

Tree-shaped shared medium access networks such as PONs and HFC effect a distributed multiplexing function which concentrates traffic from many users and many services with diverse requirements. The system thanks to prioritization can guarantee the QoS required by sensitive traffic while exploiting any unreserved bandwidth for the support of best-effort traffic. The DiffServe architecture is quite advantageous in this context and a MAC protocol taking advantage of this approach was implemented and evaluated with satisfactory results.

Acknowledgement:

The work presented in this paper was partially funded by the EU ACTS project AC327 ”AROMA”.

References

[1] IETF, Differentiated Services Working Group, RFC 2475 “Architecture for Differentiated Services”, December 1998

[2][2] Van Jacobson, Kathleen Nichols, Kedarnath Poduri, Internet Draft, draft-ietf-diffserv-phb-ef-02.txt, “An Expedited Forwarding PHB”, February, 1999

[3]Juha Heinanen, Fred Baker, John Wroclawski, Internet Draft, draft-ietf-diffserv-af-06.txt “Assured Forwarding PHB Group”, February, 1999