Performance Analysis of Optical Burst Switching with Fast Optical Switches for Data Center Networks Muhammad Imran, Pascal Landais, Martin Collier, Kostas Katrinis*, The Rince Institute, School of Electronic Engineering, Dublin City University, Dublin, Ireland * IBM—Research, Dublin, Ireland Tel: (+353) 1700 5805, e-mail:
[email protected]
© IEEE (2015). This is an authors’ copy of the work. It is posted here by permission of IEEE for your personal use. Not for redistribution. The definitive version is published in the Proceedings of the 17th International Conference on Transparent Optical Network (ICTON 2015), July 2015, Budapest, Hungary, DOI 10.1109/ICTON.2015.7193596.
Performance Analysis of Optical Burst Switching with Fast Optical Switches for Data Center Networks Muhammad Imran, Pascal Landais, Martin Collier, Kostas Katrinis*, The Rince Institute, School of Electronic Engineering, Dublin City University, Dublin, Ireland * IBM—Research, Dublin, Ireland Tel: (+353) 1700 5805, e-mail:
[email protected] ABSTRACT In this paper, we evaluate the performance of methods of optical burst switching (OBS) designed for data center networks (DCNs) by using network-level simulation. We use fast optical switches in a single hop topology with a centralized optical control plane and OBS with a two-way reservation protocol that results in zero burst loss. We use different workloads with various burst assembly parameters to identify a suitable set of parameters. Our results show that, with suitable choices for the OBS system parameters, excellent delay performance with zero burst loss can be obtained, even in the presence of high load or high values of the topological degree of communication. Keywords: Optical interconnects, optical switches, data center networking, optical burst switching. 1. INTRODUCTION Optical networks for DCNs have gained significant attention over the last few years due to the potential and benefits of using optical components. The performance of optical network is directly related to the type of the optical switching technique used. These switching techniques are Optical Circuit Switching (OCS), Optical Packet Switching (OPS) and OBS. The OCS is a connection-oriented technique in which a connection is established before actual data transmission on a pre-established dedicated path from the source to the destination [1]. Long connection establishment time and bandwidth underutilizations in the case of low traffic load are the major limitations of the OCS. The microelectromechanical system (MEMS) optical cross connect (OXC) is also called OCS switch and has been used in the backbone optical network in recent years. The OCS in DCNs has been proposed with a hybrid approach in which traditional packet switching is used with electrical switches and OCS with MEMS switches [2], or simply OCS with MEMS switches by using multi-hopping technique [3]. In the OPS, a packet consists of a data and a header portion which are in the optical domain. When the packet arrives at the node, the header is removed from the packet and is converted into the electrical domain for processing. During this processing time, the data in the packet has to be buffered in the node. Fiber delay lines (FDLs) are used for this purpose which can provide limited buffering by routing a light to the specified fibers. The packet is dropped if the switch is not configured within this time. The OPS drawback is a lack of optical buffers and output port contention. The OPS for DCNs has been described recently in LIONS [4]. OBS [5] is different from other techniques and is considered as a compromise between OCS and OPS. It has a separate control and data plane similar to OCS. Packets are aggregated into bursts. A control packet is then transmitted on a dedicated control channel to reserve resources on all intermediate nodes from the source to the destination. The burst is sent at a particular time after sending the control packet which is called the offset time. During the offset time, these bursts are temporarily stored at edge node before transmission. During this time, the switch controller at the core node processes the control information and sets up the switching matrix for the incoming burst. Burst loss due to output port contention is the major limitation of the OBS network. Output port contention can occur due to unavailability of a wavelength at the desired output port for incoming burst. Several techniques exist in the literature to avoid contention such as FDLs, Deflection Routing, Wavelength Conversion and Segmentation based dropping but none of them can guarantee zero burst loss. OBS with two-way reservation ensures zero burst loss in which a control packet reserves resources in all nodes from the source to the destination and is sent back to the source as an acknowledgement. The control packet has a high round trip time (RTT) for a large optical network. In this paper, we evaluate the performance of OBS in the context of DCN by using network-level simulation [6]. We implement OBS with a two-way reservation protocol to ensure zero burst loss. The two-way reservation is not suitable for long haul backbone optical networks due to the high RTT of the control packet but for our optical interconnect for the DCN, this RTT is not high for several reasons: 1) negligible propagation delay, 2) usage of fast optical switches at the core, 3) fast optical control plane, 4) quick processing of the control packet and 5) single hop topology. We use various workloads with different burst assembly parameters to explore a suitable configuration that shows good performance for all types of workloads. 2. FAST OPTICAL SWITCHES MEMS switches are not feasible for OBS due to high switching time, and so we selected fast optical switches for our analysis. We categorize fast optical switches into three types: 1) Arrayed waveguide grating routers
(AWGRs), 2) 1×N Space Switches and 3) Semiconductor optical amplifiers (SOAs). These are shown in Figure 1. 2.1 AWGR Switches Arrayed waveguide gratings (AWGs) are passive devices and are used mostly as optical multiplexers/ demultiplexers in WDM systems. They can multiplex a large number of wavelengths into a single optical fiber on the sender side and can demultiplex signals from single fiber into a large number of wavelengths on the receiver side. AWGs can also be used as a static WDM router called AWGR. AWGR is a combination N (1×M) AWGs on the sender side and N (M×1) AWGs on the receiving side arranged in a cyclic way. It can provide strictly non-blocking switching, if it is used with tunable wavelength converters (TWCs) at the sender side and fixed wavelength converters (FWCs) at the receiver side and is shown in Figure 1(a). TWC selects wavelength according to the desired output port. TWC provides a switching time of the order of few nanoseconds. An AWGR switch design with (512×512) ports has been reported by using 512 channels with 10GHz channel spacing [7]. TWC1
1
FWC1
1xN
Nx1
1
Receivers
Senders 1xN Coupler 1
SOA SOA N (SOAs)
Nx1 Coupler
1
Nx1 Coupler
2
Nx1 Coupler
N-1
Nx1 Coupler
N
SOA SOA
2 TWC2
1xN
Nx1
2
1xN Coupler 2
SOA SOA
1xN Coupler N-1
SOA SOA
1xN Coupler N
SOA SOA
FWC2
N-1
1xN
Nx1
N
1xN
Nx1
N-1
FWCN
TWCN
NxN AWGR
(a) AWGR Switches
N
(b) 1×N Space Switches
N (SOAs) SOA SOA
N (SOAs) SOA SOA
N (SOAs) SOA SOA
(c) SOA-Switches
Figure 1. Design of different types of fast optical switches 2.2 1×N Space Switches Photonic switches with (1×N) configurations have been built by using Polarized Lead Zirconium Titanate (PLZT) waveguide technology [8]. These switches are made by using (1×2) Mach-Zehnder with 3dB couplers arranged in multiple stages to realize a (1×N) switch architecture. These switches provide a switching speed in the range of few nanoseconds. Another kind of (1×N) switches has been presented [9] by using phased array switching technology. These switches use a single stage of phased array modulators between star couplers instead of multiple stages of Mach-Zehnder modulators. An (N×N) strictly non-blocking switch fabric can be achieved by arranging these switches into two stages. An example of an (N×N) switch configuration is shown in Figure 1(b). The number of (1×N) switches required for an (N×N) implementation is 2N, half of them arranged in a (1×N) configuration and the other half in an (N×1) configuration. A very large switch fabric can be achieved by arranging these switches in multiple stages in conjunction with SOAs in order to make good the losses introduced by insertion losses. 2.3 SOA-Based Switches An SOA is an active device that amplifies an optical signal in the optical domain. It is based on the use of semiconductor materials to provide gain. It is also used for wavelength conversion and as a gate element of switching systems based on on/off switching. SOA-based switches are commonly realized using a broadcastand-select architecture arranged in 3 stages as shown in Figure 1(c). First, the input signal is broadcasted using a (1×N) coupler. Each output of the (1×N) coupler is attached to one of N SOAs per output port. The SOA is used as a gate element to let the light pass through or not and also provides optical gain in order to make up for the losses introduced by insertion and coupling. It has been shown [10] that SOAs are required after every (1×32) coupling factor, in order to provide gain to overcome losses. A large scale strictly non-blocking switch comprising (1024×1024) ports can be realized if two stages of SOA gates are used i.e. after the first stage of SOAs, light is broadcasted again by using (1×32) coupler to another stage of SOAs. 3. PERFORMANCE ANALYSIS To assess the performance of OBS for DCNs, we developed simulations models in the OMNeT++ simulation framework [6] by modifying OBS network models for OMNeT++[13]. Important simulation parameters are presented in table 1. We consider a one hop topology which consists of top of the rack (ToR) switches at the edge and an array of fast optical switches at the core similar to the hybrid design of our recent work [11]-this is shown in Figure 2. The biggest advantage of this topology is that it is single hop and so can be scaled up and scaled out efficiently. The OBS with two-way reservation can be easily implemented in a single hop topology
because the controller has to configure only one switch per request. It has separate data and control planes. The control plane is realized by using a centralized controller. Routing, scheduling and switch configuration are the main functions of the controller. It processes control packets from all ToR switches, finds appropriate path to the destination ToR switch through optical switches, assigns timeslots to the control packets, and configures optical switches with respect to the timeslots allocated. The data plane is realized by using fast optical switches, performing data forwarding on pre-established lightpaths configured by the controller. Fast Optical Switches (NxN)
Switch Configuration Routing
Scheduling
1
Control Plane
Management Network
2
X
Data Plane
T1 TCtrl
T2
Tx
ToR1 :L
T1 TCtrl
T2
Tx
ToR2 :L
T1 TCtrl
T2
Tx
ToRN :L
Figure 2. A single hop network topology In order to realize burst assembly function, each ToR switch at the edge consists of (N-1) virtual output queues (VOQs) where N is the number of ToR switches in the network. There is a VOQ for each target ToR switch in the DCN. Packets for the same destination ToR are assembled into the same VOQ. Burst assembly can be timer based, length based or a mix of both. We consider a mixed approach in which either a timer expires or the burst length exceeds a threshold. We consider six cases for burst assembly by considering 50µs timer with {100,200,300}KiB bursts and 100µs with {100,200,300}KiB bursts. The goal is to identify a suitable burst assembly configuration which is appropriate for all traffic scenarios. Table 1. Simulation Parameters. Parameter Name Symbol Value Unit Racks/ToR Switches N 40 – Servers per rack H 40 – Control packet processing time Tproc 1 µs Overhead Toverhead 1 µs Switching Time 1 µs Burst Aggregation Time Ta {50,100} µs Burst Length L {100,200,300} KiB ON Period Exponential(900) µs OFF Period Exponential(100) µs Our simulation model consists of N = 40 ToR switches. Each ToR switch has H = 40 servers connected to it. The controller and ToR switches are connected to the management network through an electrical switch. We use one fast switch which is interfaced to the management network. We use a bursty traffic model in our simulator to reflect the bursty nature of data center traffic [12]. We use a Markov Chain Process model for bursty traffic with an ON period of 900µs and an OFF period of 100µs. We consider various inter-arrival rates of packets during the ON period to investigate traffic at different loads. We use a value of 1µs for the switching time of the optical switches because this is a conservative choice, although in some types of fast optical switch this value can be as low as few nanoseconds [4,7,9]. The RTT of the control packet includes its processing time at the controller (Tproc) and the overhead time (Toverhead). The overhead time comprises propagation delay (5ns for 1m optical fiber), the processing delay of the control packet at the electronic switch, and the optical-electrical-optical (O-EO) conversion delay. The aggregate value of Toverhead is conservatively set to 1µs although all these delays are negligible (at most a few nanoseconds [3]). We choose a value of 1µs for Tproc because hardware controllers are efficient enough to process a packet in a few nanoseconds [4]. The 2:1 oversubscription ratio is selected because this is common for DCNs [3]. We examine end-to-end delay by considering six cases of burst assembly and three cases of topological degree of communication (TDC)[11] and the results are shown in Figure 3. The TDC represents rack level flows i.e. TDC=1 means that servers in a rack communicate with servers in only 1 destination rack. If TDC=1 there is low traffic diversity while TDC=10 and TDC=20 imply medium and high traffic diversity respectively. Figure 3 shows the simulation results obtained for these three values of TDC across a range of values of load. Our results show that when the TDC is high, the delay is independent of L. Combining a value of burst aggregation time (Ta) of 100µs and a burst length (L) of 100KiB results in a low value of delay that is insensitive to load. Hence, operating a live network configured to use these parameters will result in excellent delay performance.
(a) TDC = 1
(b) TDC = 10
(c) TDC = 20
Figure 3. Load vs End-to-End Delay with respect to different TDC values. 4. CONCLUSIONS We investigated the performance of OBS in DCNs by using network-level simulation. Burst loss is the major limitation of traditional OBS but we implement it with two-way reservation to get zero burst loss similar to OCS. Two-way reservation is not appropriate for traditional backbone optical networks due to the high RTT of the control packet but in a DCN, this RTT is not high. We examine high and low diversity traffic workloads with different choices of burst assembly parameters to identify a suitable configuration for the OBS parameters. Our results reveal that the burst assembly algorithm with appropriately chosen parameters produces excellent delay performance even with high load. ACKNOWLEDGEMENTS The work was supported by the Irish Research Council and IBM through the Enterprise Partnership Scheme. REFERENCES [1] K. J. Barker, et al.: On the feasibility of optical circuit switching for high performance computing systems, in Proceedings of the 2005 ACM/IEEE conference on Supercomputing, 2005. [2] N. Farrington, et al.: Helios: a hybrid electrical/optical switch architecture for modular data centers, ACM SIGCOMM Computer Communication Review, vol. 41, no. 4, pp. 339-350, 2011. [3] K. Chen, et al.: OSA: An Optical Switching Architecture for Data Center Networks With Unprecedented Flexibility, IEEE/ACM Transactions on Networking, vol. 22, no. 2, pp. 498-511, 2014. [4] Y. Yin, et al.: LIONS: An AWGR-based low-latency optical switch for high-performance computing and data centers, IEEE J. Sel. Topics Quantum Electron, vol. 19, no. 2, pp. 3600409-3600409, 2013. [5] C. Yang, et al.: Optical burst switching: a new area in optical networking research, Network, IEEE, vol. 18, no. 3, pp. 16-23, 2004. [6] OMNeT Simulation Framework, [Online]. Available: http://www.omnetpp.org/. [7] K. Takada, et al.: Low-crosstalk 10-GHz-spaced 512-channel arrayed-waveguide grating multi / demultiplexer fabricated on a 4-in wafer, IEEE Photon. Technol. Lett, vol. 13, no. 11, pp. 1182-1184, 2001. [8] K. Nashimoto, et al.: High-speed switching and filtering using PLZT waveguide devices, in OptoeElectronics and Communications Conference (OECC), 15th. IEEE, 2010. [9] I. Murat, et al.: Monolithic InP 100-port photonic switch, in 36th European Conf. and Exhibition on Optical Communication (ECOC), Torino, 2010. [10] O. Liboiron-Ladouceur, et al.: Energy-efficient design of a scalable optical multiplane interconnection architecture, Selected Topics in Quantum Electronics, IEEE Journal of, vol. 17, no. 2, pp. 377-383, 2011. [11] M. Imran, et al.: HOSA: Hybrid Optical Switch Architecture for Data Center Networks, in ACM International Conference on Computing Frontiers, Ischia, Italy, 2015. [12] T. Benson, et al.: Network traffic characteristics of data centers in the wild, in Proceedings of the 10th ACM SIGCOMM conference on Internet measurement, 2010. [13] F. Espina, et al.: OBS network model for OMNeT++: a performance evaluation, in Proceedings of the 3rd International ICST Conference on Simulation Tools and Techniques, 2010.
