Fair Sharing Using Service-Level Agreements (SLAs) for Open Access in EPON Amitabha Banerjee , Glen Kramer , and Biswanath Mukherjee Dept. of Computer Science, University of California, Davis, CA 95616, USA Teknovus Inc., Petaluma, CA 94954, USA
[email protected],
[email protected]
[email protected]
Abstract— Ethernet Passive Optical Networks (EPONs) are an attractive solution for meeting the broadband access requirements of residential users. Open access is a regulatory requirement in many countries which mandates that the access infrastructure must be open to various service providers for free competition. Thus, different service providers may cater services to same or different users using the same shared access channel. Open access demands fairness in the use of network resources among sharing entities, namely service providers and end users. In this study, we investigate scheduling algorithms for fair bandwidth sharing in the context of open access for EPONs.
I. I NTRODUCTION Ethernet Passive Optical Networks (EPONs) are point-tomultipoint optical access networks. An Optical Line Terminal (OLT) at the Central Office (CO) is connected to many Optical Network Units (ONUs) at remote ends using optical fiber and passive splitters. [1]. EPON promises a high bandwidth to end users. Ethernet technology is inexpensive and ubiquitous, and hence compared to other solutions such as BPON and GPON, EPON seems most promising today. Any broadband access infrastructure to residential customers should be carrier neutral. The ability of the user to freely choose their service providers (SPs) and of the SPs to connect to the transport network and solicit customers is known as open access [2], [3]. In recent years, the term open access has been expanded to include not only Internet SPs (ISPs) but also voice-over-IP SPs (VoIP SPs), video SPs, and application SPs (ASPs). Open access is a regulatory requirement for residential access networks in many regions, such as some municipalities in the United States. It has also been shown that open access is a sustainable and profitable economic model for an access solution such as EPON [4]. An EPON employing open access is shown in Fig. 1. In this scenario, an end user could be a home, a multi-dwelling unit (MDU), an apartment, or an interface to a local-area network (LAN). When demand is high and bandwidth is insufficient for everyone, different end users will compete for the EPON bandwidth. Similarly SPs compete for the access bandwidth. Our motivation in this work is to ensure fairness in terms of bandwidth allocation in the shared access channel, amongst both end users and SPs simultaneously. Accomplishing both of the above objectives is challenging, because the end users and SPs are connected to opposite ends of the access channel.
Fig. 1.
An EPON employing open access.
II. FAIR Q UEUING
IN
EPON
We term a connection from a SP to a user as a flow. In this work, we consider the downstream operation of EPON, in which packets are broadcast from the OLT to the ONUs. Since traffic may be regulated by the OLT, scheduling in downstream operation is much easier than scheduling in the upstream operation (ONUs to OLT) in which queues are located at remote ONUs. We shall consider the upstream operation in future work. For illustration, we consider an example with five users and two SPs sharing the access network. A total of 60 units of bandwidth is available. Users and are accessing services simultaneously from one SP, . Two users and are accessing services from . The queue sizes1 for all the queues, , , , , and are 15 units each. Since the aggregate queue size is greater than the available capacity, we cannot satisfy all the queues. The first approach shown in Fig. 2(a) aims at fair sharing of the channel capacity, considering all the flows independently. A number of algorithms have been proposed to achieve fair queuing across different independent network flows in the literature, e.g., Deficit Round Robin (DRR) [6], Core-Stateless Fair Queuing (CSFQ), and other packet implementations of 1 The terms bandwidth, queue size, and time slot will be used interchangeably in this discussion. An EPON channel is run in a time-divisionmultiplexing (TDM) fashion so, if its cycle time is 60 units, we say that the EPON bandwidth is 60 units or it consists of 60 time slots. Backlogged traffic in queues at the OLT will need to use these time slots to travel to the ONUs. So, if a queue size is 15, it means that the OLT requires a bandwidth of 15 units, or 15 time slots, to transmit this information on the EPON channel.
(a) Independent flows
(b) SP SLA
(c) User SLA
(d) Dual SLA Fig. 2.
Bandwidth allocation using different schemes.
Generalized Processor Sharing (GPS) [5]. The disadvantage of using such a flow-based fair queuing model in an access network such as EPON is that we wish to be fair to SPs and users, and not just to flows. In Fig. 2(a), we observe that is allocated twice the bandwidth than other users, because it has two active flows. Similarly, is allocated twice the bandwidth than , because it has four active flows compared to two for . Such a scheme is not suitable for an access network because a user subscribing to large number of services may deprive other users of bandwidth. Similarly, a SP having low volume of customers may be denied fair competition from a SP carrying a high volume of customers. In Fig. 2(b), we show a solution in which the bandwidth is
shared equally amongst the SPs. This solution doesn’t achieve good fairness across users, e.g., now, is allocated three times more bandwidth than , and . In Fig. 2(c), we show a solution in which the bandwidth is shared equally amongst the users. In this case, while users get uniform bandwidth, SPs do not. From the above observations and discussions, we are motivated to investigate Dual Service Level Agreements (SLAs), i.e.,, SLAs defined for both SPs and users in one system. A solution based on such a scheme with a chosen User SLA of 10 units per user and SP SLA of 25 units per SP is shown in Fig. 2(d), and the corresponding algorithm is described in detail in Section IV. We observe that all the users as well as the SPs are granted at least the bandwidth specified by the Dual SLA. We should ensure that we do not oversubscribe the User SLAs or the SP SLAs, i.e., the sum of the User SLAs must be less than the channel capacity; similarly, the sum of the SP SLAs must be less than the channel capacity as well. However, the sum of User and SP SLAs combined together may exceed the channel capacity. Hence, it may not be always possible to meet both sets of SLAs. Therefore, we define two levels of SLA. The primary SLA is defined to be the one, whose specified minimum guarantees must be given the highest priority to be met. After the primary SLA has been met, the next priority is to meet the secondary SLA. III. A M ATHEMATICAL M ODEL FOR D UAL SLA S We define a mathematical model for fair queuing in an EPON system based on Dual SLAs. For a Dual-SLA system, we consider the User SLA as the primary SLA, and the SP SLA as the secondary SLA, although this can easily be reversed. Thus, we can state the problem as follows. Given: 1) : Rate of the EPON system. 2) : Number of SPs in the system. 3) : Number of users in the system. 4) !% #"$ : Minimum bandwidth guaranteed to SP & by the corresponding SLA, &') %/. (**+*, . - %/. % #"0$ 1 , so that the We require that: SLAs are not oversubscribed. 5) 24 3"$ : Minimum bandwidth guaranteed to User 5 by the4,. corresponding SLA, 56'7(8***, . 6)
:
- ,4 . $ 94 #"0$ 1
: The maximum scheduling time-slot. The scheduler at the OLT may schedule bandwidth for a maximum duration of : in one iteration of scheduling. This time is limited by the maximum tolerable delay, as shall be explained later. Define: 1) ; : Total capacity available in a time-slot of duration : . 2) 3)
;= @?: AB%3"$ : Minimum guaranteed bandwidth by the SLA to SP i in a time-slot of duration : . A % #"$
