Scalable and Low Latency Optical Packet Switching Architectures for High Performance Data Center Networks

As the scale of data centers and cloud computing applications increases, data center networks play a critical role in meeting the huge communication bandwidth requirement of such applications. The scalability of conventional electronic data center networks is limited by wiring complexity and reaching distance of links under fixed power budget. To overcome this problem, in this paper we propose a nonblocking hybrid switch architecture, called NEO (Nonblocking Electronic and Optical), which is able to provide nonblocking interconnections for as many as 1,000,000 servers in a data center. NEO maintains electronic interconnections for intra-pod networks, while providing interpod interconnections by optical core switches, which not only increases the scalability of the switch architecture, but also lowers the switch cost and power consumption compared to other existing optical switch architectures. We also design a packet scheduler for NEO, which adopts a credit flow control mechanism and a parallel scheduling algorithm to avoid packet loss, and provide low communication latency. Our simulation results demonstrate that NEO achieves very low average packet delay compared to other existing optical switching architectures under various traffic patterns.

The explosive growth of data center (DC) traffic imposes unprecedented challenges on the current electrical switch-based data center networks (DCNs) with the bottleneck of limited bandwidth and high latency. Benefitting from transparency to the data rate and format, optical switching with theoretically infinite bandwidth could overcome the bandwidth bottleneck of electrical switching DCNs. However, DCNs normally deploy multitenant applications with a variety of DC traffic. It is hard to reconfigure the optical interconnections in real-time to provide adaptable bandwidth to the traffic with heterogeneous characteristics. Moreover, to improve the bandwidth utilization, statistical multiplexing is generally deployed in optical DCNs to forward the traffic flow in a time slot, which requires the network time to be precisely synchronized to all network nodes. For fast optical switches with a nanosecond switching configuration time, overall end node times must be synchronized at subnanosecond magnitude. In this paper, we propose and experimentally investigate a reconfigurable and picosecond-synchronized optical DCN (ReSAW) based on an arrayed waveguide grating router (AWGR) and the White Rabbit (WR) protocol. A scheduler based on a distributed field-programmable gate array is implemented in the proposed ReSAW to realize flexible wavelength configuration by controlling the fast laser array based on semiconductor optical amplifiers (SOAs) according to time slot and traffic priority. Moreover, the WR protocol is implemented in optical DCNs for what we believe is the first time to synchronize the time of the distributed top of racks (ToRs). The experimental demonstration validates that ReSAW achieves an average end-to-end latency of 317.44 ns and a precise synchronized time with an average 386 ps skew. When the load is 0.4, the packet loss after ReSAW reconfiguration is less than 1.83 × 10 − 6 , and the network latency is less than 1.73 µs. Based on the experimental parameters and results, the OMNeT++ simulation model is built to further verify the reconfigurability and scalability of the ReSAW network. Results show that the packet loss rate and latency performance increase 8.24% and 12.47%, respectively, at a load of 0.6 as the ReSAW network scales from 2560 to 40,960 servers compared to before the reconfiguration.

As data centers enter the exascale computing era, the traffic exchanged between internal network nodes, increases exponentially. Optical networking is an attractive solution to deliver the high capacity, low latency, and scalable interconnection needed. Among other switching methods, packet switching is particularly interesting as it can be widely deployed in the network to handle rapidly-changing traffic of arbitrary size. Nanosecond-reconfigurable photonic integrated switch fabrics, built as multi-stage architectures such as the Clos network, are key enablers to scalable packet switching. However, the accompanying control plane needs to also operate on packet timescales. Designing a central scheduler, to control an optical packet switch in nanoseconds, presents a challenge especially as the switch size increases. To this end, we present a highly-parallel, modular scheduler design for Clos switches along with a proposed routing scheme to enable nanosecond scalable scheduling. We synthesize our scheduler as an application-specific integrated circuit (ASIC) and demonstrate scaling to a 256 × 256 size with an ultra-low scheduling delay of only 6.0 ns. In a cycle-accurate rack-scale network emulation, for this switch size, we show a minimum end-to-end latency of 30.8 ns and maintain nanosecond average latency up to 80% of input traffic load. We achieve zero packet loss and short-tailed packet latency distributions for all traffic loads and switch sizes. Our work is compared to state-of-the-art optical switches, in terms of scheduling delay, packet latency, and switch throughput.

We propose an AWGR and SOA based scalable optical packet switch architecture with modular structure and highly distributed control scheme for data center networks. Numerical results show this architecture can achieve high throughput with low blocking probability and latency.

We highlight the advantages of dynamic wavelength allocation and path rerouting in hybrid optical flow-switched data center networks compared with static electronic packet-switched data center networks. Current electronic packet-switched networks perform routing based on lookup tables and shortest path routing over high speed optical connections. However, load balancing and path rerouting are not designed in coordination with the optical data plane. In this work, we design and analyze the performance of load balancing and resource allocation in hybrid flow data center architectures. In the hybrid flow architecture, the transport layer and control plane are responsible for load balancing, congestion control and reliable data delivery in wavelength division multiplexed networks. The transport layer abstracts the underlying network and data link layers from user applications. Applications are unaware of whether flows are transmitted via the electronic data plane, optical data plane, or both, where the relevant metric for applications is transaction delay. We show that the data center architecture that minimizes delay in the presence of elephant flows and dynamic loads is the hybrid flow architecture with dynamic resource allocation.

Data center networks (DCNs) are at the forefront of a radical transformation, fueled by the rapid rise of artificial intelligence (AI), cloud-native adoption, and enterprise digitalization, which together place unprecedented demands on their design and performance. From healthcare AI and intelligent transportation to business analytics, modern DCNs are expected to support a diverse and evolving mix of bandwidth-demanding applications, each with distinct quality of service (QoS) requirements. Yet, most state-of-the-art DCN solutions rely on oversimplified DCN data traffic models, failing to address the required application-level differentiation in terms of end-to-end (E2E) latency demands. In response, we propose ALAS-DCN, a novel, to our knowledge, application-aware, low-latency, scalable DCN architecture, based on optical packet switching in the intra-rack and inter-rack domains that is both designed and rigorously evaluated using commercial QoS specifications and realistic traffic patterns of different applications coexisting within the same DCN. ALAS-DCN integrates three hierarchical optical sub-networks: intra-rack, inter-rack, and bridge, each governed by a custom wavelength division multiplexing (WDM) medium access control (MAC) protocol that dynamically prioritizes critical traffic packets, according to the specific application E2E latency demands. Unlike prior approaches, our evaluation leverages realistic workload traces generated by established traffic generators, capturing the application diversity within modern data center (DC) environments. Extensive experimental simulations across five representative DC categories, each hosting a different mix of modern applications, demonstrate that ALAS-DCN consistently achieves up to 95% throughput efficiency and sub-µs latency for high priority traffic across a variety of applications, consistently meeting their differentiated latency constraints. Scalable up to 20,000 servers, ALAS-DCN redefines the DCN as not just a transport layer but an intelligent enabler of next-generation application performance.

Maintaining high-performance operation under dynamic and nonuniform network traffic has been a technical challenge in current data center networks (DCNs). With the aim to provide better quality of service (QoS) for diverse applications, this work presents a dynamic and adaptive DCN reconfiguration framework based on deep reinforcement learning (DCR2L). The proposed framework is integrated into the SDN control plane of the DCN, implementing real-time and automatic DCN reconfiguration. Performance of the DCR2L framework is experimentally demonstrated in our DCN lab, including 4 racks and 16 servers. Experimental results show a network latency improvement of 6.9% based on the DCR2L at an average network bandwidth of 2.3 Gb/s. Based on measured traffic and physical parameters in the experiment, performance of the DCR2L framework is numerically assessed with the realistic traffic of diverse QoS requirements for both electrical and optical DCNs and for different data center scales. Leaf–spine electrical and OPSquare optical networks are set up in an OMNeT++ simulator. For a data plane network consisting of 16 racks and 320 servers, results indicate that the DCR2L framework improves the network latency of up to 16.4% under leaf–spine and up to 24.6% under OPSquare for the overall traffic with respect to the classical heuristic method. For a DCN scale of 10,240 servers, the DCR2L framework provides up to 12.5% lower latency for the leaf–spine electrical network and up to 17.8% latency improvement for the OPSquare optical network.

We present a field-programmable gate array (FPGA)-based label processor for in-band optical labels with a processing time independent of the number of label bits. This allows for implementing an optical packet switching architecture that scales to a large port count without compromising the latency. As a proof of concept, we have employed an FPGA board with 100 MHz clock to validate the operation of the label processor in a 160 Gb/s optical packet switching system. Experimental results show successful three label bits processing and 160 Gb/s packets switching with 1 dB power penalty and 470 ns of latency. Projections on the label processor performance by using more powerful FPGAs indicate that 60 label bits ( optical addresses) can be processed within 31 ns.

Optical switching, which benefits from high bandwidth and low latency, promises to revolutionize data center networks that suffer from a bandwidth bottleneck and hierarchical structure. However, the lack of adaptable and fast optical switch control schemes, the lack of optical buffers for contention resolution, and the high cost of clock and data recovery (CDR) impede the practical deployment of optical switches in data centers. Networking communities have tried to address these challenges from different perspectives. However, the proposed solutions may be interdependent and thus introduce unexpected issues. In this article, a synergistic optical switch control mechanism is proposed to simultaneously solve the aforementioned challenges. The developed label control technique allows for controlling and configuring packet switching within 43.4 ns. The implementation of clock distribution achieves a fast CDR of 3.1 ns at low cost, and the design of an optical flow control protocol prevents packet loss caused by packet contention.

When optical switching is deployed in Data Center Networks (DCNs), the reconfiguration of the optical switching matrix leads to substantially longer overheads, posing a significant impact on the system performance. Despite the extensive studies on the scheduling algorithms based on demand matrix decomposition (DMD), the stateful and irregular nature of the scheduling processes hinders the development of quantitative models, thereby limiting our understanding of resource scheduling in optical switching DCNs based on DMD. In this paper, we model the DMD based resource scheduling process under a bursty and skewed traffic pattern and derive closed-form equations for the burst completion time. Our study shows that an increased reconfiguration delay will lead to an approximate linear increase in the burst completion time. Our study also demonstrates that the size of the slot and the maximum allowed duration of one match are approximately inversely proportional to the burst completion time, with diminishing marginal returns.

Optical switching technology arises as the dominant technology for data center networks (DCNs), grace to its inherent capability to serve time-sensitive traffic with low power consumption. In this study, we introduce an optical DCN that groups the servers into computing clusters. An effective MAC algorithm is adopted for the communication within a cluster, based on wavelength and time division multiple access (WDMA and TDMA) techniques. The proposed MAC algorithm avoids the collisions over the channels and achieves high effectiveness, while it serves different traffic types. We investigate the proposed MAC protocol performance under diverse servers’ population per cluster to determine its scalability and performance limits. The simulation study proves that the proposed DCN architecture with the MAC protocol reaches high performance, i.e. high throughput, low delay, low dropping rate and almost 100% bandwidth exploitation, for diverse servers’ population. Finally, comparative performance and power consumption study is given.

In this talk, optical switching technologies will be reviewed for possible deployments in data center networks. Recent research in intra/inter DCN, such as DCN architecture and experimental demonstrations, will be summarized. Due to its low power consumption, optical switching technology will be one of the key enabling technologies for future large-scale data center networks. In this talk, recent research in data center architectures is reviewed and discussed to explore the potential deployment of optical switching technology in optical data center networks. Optical switching technologies include time division multiplexing (TDM), wavelength division multiplexing (WDM) and space division multiplexing (SDM) will be discussed. In addition, some latest research in inter-DCN communication technologies is reported.

Data center networks generate high volumes of traffic. In order to reduce packet latency, packet transmissions are often centrally scheduled. Such approaches have been proposed for both packet-switched and hybrid optical-packet switched networks. This paper investigates algorithm design choices for transmission scheduling in a tightly synchronized hybrid optical packet data center network. This problem is studied in two cases: with precedence where the requests are scheduled in the order of arrival, and without precedence, where the requests can be reordered in time. It is shown that the problem without any precedence constraints is NP-complete. For scheduling with precedence constraints, a greedy algorithm is proposed and shown to be optimal. Theoretical approximation for the performance of scheduling with the greedy algorithm is presented. Simulation experiments were performed on a two-tier network with 1024 servers and 64 wavelengths. Parallel implementation aspects of the scheduling algorithm are also discussed.

Optical switching based small-world data center network

We investigate the performance of a novel flat data center network (DCN) architecture based on distributed buffer-less optical switches and fast flow control. Scalability of the novel DCN is assessed by numerical simulation of medium and large size data centers reporting an end-to-end latency of less than 8μs (including retransmission), a packet loss 9Tb/s average network throughput under load of 0.4.