OMUX: Optical Multicast and Unicast-capable interconnection network for data centers

https://doi.org/10.1016/j.osn.2019.01.002Get rights and content

Abstract

Exponential growth of traffic and bandwidth demands in current data center networks requires low-latency high-throughput interconnection networks, considering power consumption. By considering growth of both multicast and unicast applications, power efficient communication becomes one of the main design challenges in today's data center networks. Addressing these demands, optical networks suggest several benefits as well as circumventing most disadvantages of electrical networks. In this paper, we propose an all-optical scalable architecture, named as OMUX, for communicating intra-data centers. This architecture utilizes passive optical devices and enables optical circuit switching without any switch configuration or path reservation. In addition, it is customized for both unicast and multicast transmissions. It can replace aggregation and core switches in the network, to reduce power consumptions. Simulation results verify its lower latency, higher throughput, and reduced power consumption, compared to conventional electrical or alternative optical solutions.

Introduction

Increasing traffic and bandwidth demands are among current challenges of efficient Data Center Network (DCN) design, which should, along with the consideration of power consumption, enhance throughput and reduce latency. Data center communication growth has been exponential in recent years and is predicted to increase more [1]. Data center communications are mostly divided into Unicast and Multicast transmissions. Multicast traffic is a kind of packet transmission in which same data is sent to multiple destinations. Multicast data transmission is usually carried out as multiple unicast packets separately sent to different destinations.

Unicast is quite familiar in DCN but according to Fig. 1, common case of multicast transmission can be observed in video streaming which comprises 83% of transactions to end-users and 10% of intra-data center communications in 2015. However, for other traffic patterns intra-data centers, such as those initiating by data mining and web browsing applications, internal software updating, and data center or VM migrations, certain algorithms are executed producing significant amount of multicast packets [2], such as MapReduce [3], parallel database join operation [4], Latent Dirichlet Allocation algorithm for text mining [5], and Alternating Least Squares algorithm for prediction [6].

As electrical solutions for multicast, protocols such as Media Transfer Protocol (MTP), Protocol-Independent Multicast (PIM), and PIM Sparse Mode (PIM-SM) were formerly used in small networks for multicast, but are not applicable in large networks like DCN due to their defects [7]. In this manner, IP Multicast protocol based on IP addressing subnets was introduced to facilitate multicast in DCN [8]. However, IP Multicast requires complex configurations in routers, whose scaling in multilayer networks is a serious problem which has been addressed in various studies [[9], [10], [11]], and still lack desirable efficiency in terms of latency, scalability and reducing multiple sending [12,13].

Capabilities of optical communication equipment and its progress in recent years have proposed an optical network as an auxiliary, or even, alternative solution for electrical DCNs. An optical network offers various benefits, such as high bandwidth through WDM technique, low propagation delay due to capacitance and inductance load free optical fiber, and low power consumption [[14], [15], [16]].

Several outstanding works proposing optical infrastructure for unicast data transmission have been presented and verified through small prototypes in Refs. [[17], [18], [19], [20]]. However, considering related studies, in this paper, we review those architectures capable of data multicasting, alongside data unicasting, through the optical networks. For example [21], proposes a star topology in which all nodes are interconnected through an optical switch, while one of its output ports is connected to an optical splitter to transmit multicast packets. In the case of optical multicasting, corresponding optical data is sent from the source node to the multicast port, at where it is copied and transmitted to multiple destination nodes, simultaneously. As a collision avoidance policy, packets might be buffered electrically to ensure two or more optical streams with the same modulation wavelength do not pass through an optical fiber at the time. In this manner, electrical controllers required for wavelength management, as well as, wavelength scheduling, lead to performance degradation and power overhead. On the other hand, since each multicast packet is transmitted to all output ports of the splitter, overcrowded network may lead to further power dissipation.

Ref. [22] uses ring topology and connects data center pod switches to each other by replacing them with optical MEMS switches, while, local optical switches are utilized for intra-pod communication. Furthermore, a control unit manages optical switches, as well as processes incoming requests using external signals. This unit allocates wavelength channels to source nodes and configures switches in order to guarantee collision-free network. Internal splitters of these local switches can enable multicast, as well as unicast, data transmission. In current architecture, multiple path-switching reduces network performance due to low switching speed of MEMS. On the other hand, space switching advantages cannot be easily obtained in ring topology, due to lack of path diversity.

Star topology in Ref. [23] uses multicast-capable cyclic Arrayed Waveguide Grating Router (AWGR) as a central switch to interconnect nodes. AWGR is a passive optical switch that directs input optical stream modulated on a specific wavelength channel to a specific output [24]. Moreover, multicast-capable AWGR can also direct each input stream to multiple output ports by power splitting. As a result, multicast packets can be distributed by these AWGRs. Nevertheless, limited number of wavelength channels available in optical infrastructures restricts unicast and multicast transmission, and hence, the Star topology suffers from scalability problem even for a few hundred of nodes.

Refs. [12,13,15] propose an optional auxiliary optical network facilitating multicast transmission in electrical DCN. Despite differences in routing algorithms optical architectures are built upon similar physical infrastructure, shown in Fig. 2. As shown in this figure, input of each optical splitter is connected directly, or through a controller, to specific ToRs, while its outputs are connected to intra-cluster ToRs or some other ToRs in the optical network. Multicast packets are delivered to splitter's input through the electrical network, and are optically broadcasted to the corresponding output ports. This auxiliary network improves latency, power, and throughput by preventing multiple packet transmission in the electrical network. Although fixed multicast groups eliminates group allocation and scheduling overheads, it necessitates periodic updates of network routing tables Moreover, since splitters impose optical power loss, their number of output ports do not exceeds 16, and hence, limit the size of local clusters. Finally, scheduling incoming multicast packets at the input port of splitters imposes additional latency overhead.

In this paper, we introduce a scalable optical architecture for interconnecting ToRs within a data center. As an all-optical and low latency network, the proposed architecture eliminates both electrical buffers and external controllers. Taking advantages of passive optical devices and by eliminating aggregation and core switches, it results in less power consumption for both unicast and multicast transmission. Finally, our simulation results indicate that the proposed optical DCN outperforms counterpart optical designs in terms of optical power loss, data transmission latency, and network throughput. As follows, in the subsequent section, the proposed design is explained in detail. In the last two sections, results are compared to related works and conclusion is made.

Section snippets

OMUX: Optical Multicast and Unicast-capable interconnection network

In this section, topology and routing algorithms of the proposed architecture for intra data center communication, denominated as Optical Multicast and Unicast-capable interconnection network (OMUX), is presented. OMUX is built upon a Mesh-like topology for connecting ToRs in rows and columns, while it is a single layer (i.e. flat) and a fully connected network. As a key advantage, this all-optical architecture eliminates opto-electrical conversion, wavelength conversion, and the central

Simulation results

In this section, we use OMNET++ [31], as an event-driven simulation environment, to evaluate the performance of OMUX architecture against two alternative architecture; (1) FatTree network [32] as a pure electrical solution, which is mostly used in current data centers, and (2) FatTree network facilitated with an auxiliary optical network (FT with OMC) similar [12,13,15] for multicast data transmission.

The assumptions for simulating the aforementioned networks are presented in Table 1. For this

Conclusion

In this paper, a novel all-optical architecture, named as OMUX, is proposed enabling both unicast and multicast data communications intra data centers. This architecture eliminates opto-electrical conversions, wavelength conversions, path reservation, and central controller requirement. In OMUX architecture, the optical switches are distributed around the network to increase scalability and reduce interconnecting complexity. By employing passive optical devices and performing transmissions

References (36)

  • T. Barnett et al.

    Cisco Global Cloud Index 2015-2020, Tech. Rep.

    (Nov 2016)
  • R.N. Calheiros et al.

    Virtual machine provisioning based on analytical performance and qos in cloud computing environments

  • J. Dean et al.

    Mapreduce: simplified data processing on large clusters

    Commun. ACM

    (2008)
  • W. Mach et al.

    Parallel database join operations in heterogeneous grids

  • D.M. Blei et al.

    Latent dirichlet allocation

    J. Mach. Learn. Res.

    (2003)
  • M. Chowdhury et al.

    Managing data transfers in computer clusters with orchestra

  • R. Perlman et al.

    Simple Multicast: A Design for Simple, Low-Overhead Multicast, Internet Engineering Task Force

    (Oct 1999)
  • R. Rümmler et al.

    Fundamentals of IP Multicast

    (2009)
  • O. Komolafe

    Ip multicast in virtualized data centers: challenges and opportunities

  • X. Li et al.

    Scaling ip multicast on datacenter topologies

  • T. Kernen et al.

    Optimised ip multicast architectures for real-time digital workflows

  • Y. Xia et al.

    Blast: accelerating high-performance data analytics applications by optical multicast

  • P. Samadi et al.

    Optical multicast system for data center networks

    Optic Express

    (2015)
  • H.J.S. Dorren et al.

    Challenges for optically enabled high-radix switches for data center networks

    J. Lightwave Technol.

    (2015)
  • D. Wu et al.

    Hyperoptics: a high throughput and low latency multicast architecture for datacenters

  • W. Liang

    Constructing multiple light multicast trees in wdm optical networks

  • A. Shacham et al.

    Building ultralow-latency interconnection networks using photonic integration

    IEEE Micro

    (2007)
  • R. Luijten et al.

    Optical interconnection networks: the osmosis project

  • Cited by (0)

    View full text