Optical flow switching (OFS) has been proposed as a simple and cost-effective transport technology for users with large transactions (>1 second). In previous studies, a fast wavelength reservation method was deployed for flow transmission in OFS-based networks. However, reserving a single wavelength for users with small transactions encounters a very common problem: inefficient wavelength utilization. In this paper, a flow transmission cycle is introduced to each wavelength, and each cycle consists of multiple slots, so that flows of different transactions can be multiplexed onto a single wavelength. It is assumed that inter-metropolitan area network (MAN) traffic is transported over wide area network (WAN). A global time-shift scheduling methodology taking into account propagation delays in MAN is designed to avoid potential contentions occurring among different flows which are carried by the same wavelength in WAN. The contributions of this paper are, first it provides a new OFS network architecture which can achieve better throughput and average wavelength utilization performance without losing the feature of simple transport structure provided by OFS; second, it is the first time that issues of how OFS networks are managed and controlled are addressed from a system point of view.
© 2011 OSA
The popularity of Internet TV and smart phones accelerates the trend shift from the text-based applications to the streaming contents dominant ones . Most of the streaming-based applications are delay-sensitive and consume high bandwidth, which will stress the current network infrastructure performing electrical packet switching, because electrical IP/MPLS routers with faster processing capability become necessary to maintain a high degree of customer satisfaction (good user experience). To relieve the stress of IP/MPLS router layer, one possible solution is to offload the traffic from the IP/MPLS router layer to the optical transport layer, in that the traffic can bypass the intermediate electrical IP/MPLS routers when possible. In addition, since optical RAM is still not practicable to packet switching in optical transport layer, the traffic can be handled in terms of flows, in that better wavelength resource utilization can be obtained [2,3].
Optical flow switching (OFS) has been proposed as a simple optical transport technology for providing quality of service (QoS) guaranteed end-to-end service by eliminating both buffer at core nodes of WAN and electrical traffic aggregation at MAN areas and thus allowing large transactions to bypass intermediate electronic routers . OFS features quick optical path setup (100ms), however, it is clearly distinct from optical circuit switching (OCS) because it adopts statistical multiplexing of flows from many end users. OFS is also differentiated from optical burst switching (OBS), in that the transmissions of flows are pre-coordinated on an optical path so as to avoid the contention. On the contrary, in OBS the burst data have to be discarded if there is no wavelength available at the intermediate node for the next hop. OFS does rely on multiplexing; however, no buffer is used at intermediate nodes unlike optical packet switching (OPS).
The scalability of OFS was often questioned because the network state information has to be updated per flow basis, and this heavily bears the scheduling task to the control plane. Recently a solution has been proposed to address the scaling issue, which adopts the network sampled entropy function to simplify the control plane messaging . Previous studies have implemented the scheduling algorithm of flow transmission using queuing theory based on M/G/S model for the case with the significant multiplexing of flows . Obviously, this is a cost-effective when end users demand large bandwidth. However, a problem remains that the reservation of single wavelength for flow transmission between two end-users encounters low wavelength utilization when the required bandwidth becomes small.
In this paper, in order to improve the wavelength utilization efficiency as well as the energy saving without complicating the node architecture, a flow transmission cycle is introduced to each wavelength and each cycle consists of multiple slots, so that flows of different transactions can be accommodated onto a single wavelength. It is assumed that inter-MAN traffic is transported over WAN. A global time-shift scheduling methodology taking into account propagation delays in MAN is designed to avoid the potential contentions occurring among different flows which are carried by the same wavelength over WAN. Numerical simulation results show that our proposed time-shift scheduling is effective to maintain the throughput performance even when the bandwidth required by end users is small. Fewer wavelengths are needed to be kept active with our method, which results in less power consumption of transceiver.
2. Optical flow switched network architecture by flow multiplexing
2.1 Optical flow switched network architecture
Figure 1 illustrates our OFS network architecture which consists of three parts: OFS-based optical transport layer, scheduler deciding how to multiplex flows from different transactions, and out-of-band control plane for exchanging network state and control information.
The OFS-based optical transport layer is composed of distribution networks (DNs), MANs, and WAN. DN and MAN can be connected by transponders, while MAN and WAN are connected by gateway OFS switches. DN serves as the interface between the network and its end-users, through which traffics are loaded to MAN. For optical signal transmission, wavelength-division multiplexing (WDM) technology is applied in the MAN and WAN areas. Optical flow switches and wavelength cross-connects (WXCs) are applied in MAN and WAN areas, respectively. According to whether going across WAN or not, traffics can be classified into intra-MAN traffic and inter-MAN traffic . In this paper, we only concentrate on inter-MAN traffic because of the following reason: transferring inter-MAN traffic consumes wavelength resources in WAN, thus, improvement of wavelength resource utilization efficiency in WAN is beneficial to helping network operators to reduce capital expenditure (CAPEX).
In OFS-based networks, scheduler located at the gateway OFS node plays an important role in assigning bandwidth to transactions. In our proposal, the scheduler will calculate a global flow-assignment table (hereafter, called FA-table) indicating how to assign bandwidth to different DNs and switching tables (hereafter, called S-table) indicating how to configure the OFS switches. In Section 3, we will explain what the scheduler is and its operation in details.
The control plane takes responsibility of dealing with transaction setup and termination requests, network state information distribution for notifying bandwidth assigned to DNs and communications among schedulers for obtaining global available bandwidth information. As we will explain in Section 3, each DN keeps the above-mentioned FA-table indicating the timing of a flow from DN can be loaded to MAN, while each OFS node keeps the above S-table indicating the local forwarding decision for the incoming flows. These two tables need to be updated by the schedulers. In this paper, we assume the control information is transmitted by a separated IP network instead of the OFS-based optical transport layer (out-of-band manner).
2.2 Flow multiplexing network model for inter-MAN traffic
Multiplexing flows of inter-MAN traffics onto a single wavelength can provide better wavelength utilization efficiency, comparing to the existing wavelength path-based OFS network architecture . In our proposed flow multiplexing, first no buffer is needed to avoid flow contentions; second, the flow length is flexible to be adjusted by a network planning phase. These features make our proposal clearly distinct from the statistical multiplexing commonly used in packet switching and TDM-like multiplexing.
Prior to perform flow multiplexing, tree topology rooting from gateway OFS node needed to be created for each MAN in the network planning phase beforehand. Tree-based MAN topology has the advantage of accommodating many end-users cost-effectively . In WAN, quasi-static wavelength paths are set up to link up the gateway OFS nodes (roots of the tree topologies), so that the individual tree topologies can be connected to transmit the inter-MAN traffics. In general, the wavelength paths in WAN can be also dynamically re-configured, but network operators do not expect frequent re-configuration for their optical transport layer because such operations have to face complicated optical engineering problems, and eventually, increase the operational expenditure (OPEX). Therefore, we assume the wavelength paths in WAN are quasi-static and not changed when the network is running. Figure 2 illustrates that two tree topologies are linked by a quasi-static wavelength path. The inter-MAN traffics from DNs are first loaded to its adjacent MAN, followed by multiplexing and de-multiplexing operations carried out by gateway OFS nodes at the source MAN and destination MAN, respectively.
Figure 3 illustrates an example of flow multiplexing for inter-MAN traffics. The basic concept of flow multiplexing is to slice the bandwidth of a wavelength into N flow slots with a given time period (hereafter, called cycle). A transaction is assigned some flow slots and its data is sent with the same slots in each cycle. By this method, flows of individual transactions become possible to be multiplexed onto a single wavelength. On the other hand, optical flow switching in terms of flow slot is performed periodically to forward the flows to their desired directions at every OFS node. To millisecond-order optical switch deployed to switch flows, we consider the switching time cannot be ignored, thus, a guard band Tg is inserted into two consecutive flows of different transactions. Note that, the maximum length of the flows that can be loaded to the network must not exceed Tf–Tg. In performance evaluation section, we will observe that per flow length, guarding time Tg and number of flows per cycle have impacts on the network throughput performance in the numerical simulations.
3. Time-shift scheduling methodology
3.1 Flow scheduling with link propagation delay considerations
Inter-MAN traffic from the source MAN traverses through the gateway OFS node (root of the tree topology) to the destination MAN. Because there exists only one wavelength path, and the destination MAN has a tree topology, there is no contention in WAN and the destination MAN. Therefore, the only problem we have to resolve is the contentions in the source MAN. In addition, we have to line up the inter-MAN flows efficiently so as to avoid a waste of bandwidth in time domain.
To resolve the above problems, we propose a time-shift scheduling methodology. For an example, the source MAN topology with link propagation delay index Dx is shown in Fig. 4(a) . Figure 4(b) illustrates the timing chart of contention-free transmission to the gateway OFS node, in which the timing of transmitting flows from DN(A), DN(B) and DN(C) is shifted by some delay. From Fig. 4(b), it can be observed that, to avoid the contention with the flows from DN(A), the timing of sending flows out from DN(B) and DN(C) has to be shifted by DB-DA + Tf and DC-DA + 2 × Tf, respectively. It is reminded that Tf denotes the time duration of flow and its guard band.
Assume that the flows from individual DNs are coordinated at the gateway OFS node at t0, the timing of sending flows by a DN can be given by,
3.2 Flow assignment methodology
We explain how to configure the global FA-table. First, the number of flow slots scheduled for a DN should be calculated and it is given by,Figure 5 shows the FA-table of a wavelength for the case with Fofs = 10 and Bw = 10 Gb/s. From Eq. (3), Fofs =2, 5 and 3 flow slots are assigned to DNA, DNB and DNC, respectively.
3.3 Working principle of scheduler
Figure 6 illustrates the scheduler structure which consists of a scheduling processor, a first-in-first-out (FIFO) request queue, and an FA-table and a switching table (S-table). Transaction requests that cannot be processed immediately are pushed to the tail of the queue and the one to be processed is popped from the top of the queue. Such a queue is beneficial for providing more candidates to be served by the network so as to achieve a better multiplexing efficiency. Moreover, the requests will be discarded if the given queuing time is out. The scheduling processor calculates the time needed to be shifted and the flow assignment mentioned above. According to the calculated results, the FA-table and S-table at the scheduler are updated and distributed to the DNs and OFS nodes.
We consider two types of scheduling methods: on-demand and in-cycle methods. On-demand method conducts the scheduling and distributes the updated FA-table/S-table each time the connection request arrives. In contrast, in in-cycle method, the scheduling and distribution are performed periodically with a fixed interval (hereafter, called cycle). Before the next scheduling cycle arrives, all received requests reside in the FIFO queue. The on-demand method can response to requests immediately, considering to achieve better wavelength utilization performance, but this makes the control plane bear heavy processing burden; on the other hand, the in-cycle method cannot response to the request at once, which may be considered to increase the latency of serving a connection and be a reason of deteriorating wavelength utilization efficiency, but it helps relaxing the computation load of the control plane. In both methods, the scalability of updating FA-tables and S-tables in the control plane is affected by the following two factors: first, the number of flows per wavelength and the number of wavelengths needed to be updated, which decide the amount of information sent to DN or OFS nodes; second, the number of DN and OFS nodes needed to be configured, because a central scheduler is assumed and it has to establish different communication sessions to the nodes.
The global time-shift scheduling procedures are summarized as follows (Fig. 7 ):
- 1. Upon arrival of flow setup or termination request, flow scheduling is performed and FA-table is configured;
- 2. FA-table is forwarded to Destination-MAN by Source-MAN scheduler;
- 3. S-tables in S- and D-MANs are updated according to the FA-tables;
- 4. FA-table and S-table are distributed to DNs and OFS switches in MAN, respectively;
- 5. In case of flow setup request, flow transmission is initiated according to the FA-table and forwarded through OFS switches using S-table.
4. Performance evaluation
Computer simulation is conducted to evaluate the performance of our proposed OFS-based network architecture, which only handles inter-MAN traffics.
4.1 Simulation model
Figure 8 illustrates a tree-based logical topology consisting of a source MAN and a destination MAN. Parameters used in the simulation are configured as follows:
- · the network includes 18 DNs, 20 OFS nodes and 2 WXCs,
- · link propagation delay of access link, links in MAN and WAN is 0.02 ms, 0.5 ms and 2.25 ms, respectively (each link is considered bi-directional),
- · 10 wavelengths are assumed to be available in WAN,
- · the maximum queuing time is 1000 ms,
- · flow transaction arrival is Poisson distributed (arrival rate is λ), average flow service duration (1/μ) is exponential distributed and bandwidth demanded by a flow transaction is quasi-normal distribution with a range of 1 ~10 Gb/s
In the simulation, network performances are evaluated with the following equations:
4.2 Throughput performance of the proposal
Figure 9 provides some insights to throughput performance by varying flow length, number of flow slots per cycle and FA-table/S-table update interval. In the illustrated cases, the average arrival rate λ is 20, and the average demanded bandwidth is 5.5 Gb/s with a variation of 2 Gb/s.
In Fig. 9(a), we assume there are 30 flow slots per cycle and the update interval is set 600 ms. We can observe that the throughput performance degrades when the flow length increases. This is due to the resource waste of the last slot used for flow transmission. In addition, the on-demand method outperforms the in-cycle one because the FA-table and S-table can only be updated with the in-cycle method when the next update cycle comes, and this leads to better bandwidth utilization efficiency. An increase of update interval also results in similar performance behavior which is illustrated in Fig. 9(b).
Figure 9(c) illustrates how the network performance varies with the number of flow slots per cycle, in which flow length and update interval is set 10 ms and 660 ms, respectively. From the results, we can observe that lower throughput performance is obtained with fewer flow slots per cycle. This is caused by the coarse bandwidth granularity which results in poor wavelength utilization efficiency.
4.3 Performance comparison: time-shift scheduling method vs. wavelength path method
In this subsection, we compare our proposal time-shift scheduling method with the existing wavelength path method by investigating the impacts of average demanded bandwidth and switching speed.
Impact of average demanded bandwidth
Figure 10 illustrates the impacts of average demanded bandwidth on the above-mentioned three methods: on-demand, in-cycle and wavelength path, with respects of service latency (delay), blocking rate, network throughput, average served bandwidth and average used wavelengths. Simulation parameters are configured as follows: average arrival rate λ = 20, average service time 1/u = 500 ms, optical switching time = 1 ms, flow slot length = 10 ms, number of flow slots per cycle = 40, and FA-table update interval = 660 ms.
Figure 10(a) shows that service latency increases in on-demand and in-cycle methods while decreases in the existing wavelength path method when demanded bandwidth increases. It indicates that transactions demanding large bandwidth have to wait longer in the FIFO queue until available wavelength is found in on-demand and in-cycle methods. However, in the wavelength path method, it takes shorter to finish data transmission with larger demanded bandwidth than small one, therefore, transactions can be finished faster.
Results in Fig. 10(b) indicate that the blocking performance owns similar behavior to the service latency. In this paper, we consider that transaction requests are discarded if the queuing time is out. Therefore, longer service latency may result in more discards. The blocking performance has a direct impact on network throughput: lower blocking rate, higher network throughput.
Figure 10(c) shows the network throughput performance, whose behavior is opposite to the blocking performance. On-demand and in-cycle methods outperform the wavelength path method with a bandwidth of 2 ~7.5 Gb/s and 2 ~6.5 Gb/s, respectively. With transactions requesting low bandwidth, a good multiplexing effect is achieved in on-demand and in-cycle methods, whereas the wavelength path method can only adopt one transaction regardless of demanded bandwidth. This multiplexing effect is deteriorated with an increase of demanded bandwidth, but the network throughput tends to have a similar behavior as the wavelength path method with bandwidth larger than 7.5 Gb/s and 6.5 Gb/s, respectively. Moreover, we can notice that on-demand and in-cycle methods reach an upper bound of 0.8 and 0.7, respectively, which are considered to be determined by the flow slot length, the number of flow slots per cycle and the overhead due to guarding time.
Another advantage of our on-demand and in-cycle methods is that, fewer wavelengths are needed to accommodate the traffic regardless of the demanded bandwidth (Fig. 10(d)). From energy-saving point of view, our time-shift scheduling methodology is superior to the conventional wavelength path method because more wavelengths in the WDM system can go in dark as well as more transceivers can.
Impact of optical switching speed
In time-shift scheduling OFS network, switching is performed on a flow basis, and we call the switching time called guard time, while in wavelength path-enabled OFS networks, it is performed on a wavelength basis, and we call it reconfiguration time. Simulation configuration is as follows: average arrival rate λ = 20, average service time 1/μ = 500 ms, number of flows per cycle = 40. Ts is taken for switching time, and the flow length is 10·Ts. From Fig. 11 , we can observe that, our proposal time-shift scheduling method suffers evident throughput performance deterioration with an increase of switching time. The reason is that, switching time is treated as part of flow slot, and large switching time results in large overhead which in turn decreases wavelength utilization efficiency. On the other hand, switching time has only a minor affect on wavelength path method.
To tackle the inefficient wavelength utilization problem in the existing OFS network architecture, a global time-shift scheduling enabled OFS network architecture was proposed and a corresponding network system. Numerical simulations have demonstrated that proposed on-demand and in-cycle methods outperform a conventional wavelength path method in terms of the throughput and the wavelength count regardless of the average bandwidth per flow with scarifying some service speed. Hence, from the energy-saving viewpoint, it can be claimed that proposed time-shift scheduling algorithm is effective because more wavelengths can go in dark as well as more transceivers can.
This project was financially supported by NEC Corporation, Japan.
References and links
1. Cisco Systems, Inc., “Approaching the Zettabyte Era,” White Paper, June 16, 2008.
2. L. G. Roberts, “The Internet is broken,” IEEE Spectr. 46(7), 36–39 (2009).
3. M. Zukerman, “Back to the future,” IEEE Commun. Mag. 47(11), 36–38 (2009). [CrossRef]
4. G. Weichenberg, V. W. Chan, and M. Médard, “Design and Analysis of Optical Flow-Switched Networks,” J. Opt. Commun. Netw. 1(3), B81–B97 (2009). [CrossRef]
5. V. W. S. Chan and Z. Lei, “Scalable control plane architecture for optical flow switched networks,” in Optical Fiber Communication Conference, OSA Technical Digest Series (Optical Society of America, 2011), paper OWP4.