Apostolos Siokis

Laying out interconnects on optical printed circuit boards

Short distance optical interconnections, on-printed circuit boards, on-backplanes, and even on-chip, are a promising solution for replacing copper interconnections in future Data Center and HPC systems. Since photonic technology introduces new network building blocks, topology design for all the packaging levels should be reconsidered. This paper focuses on the on-board level of the packaging hierarchy, and proposes lay-out strategies for optical interconnection networks on optical printed circuit boards (OPCBs), based on direct topology families (tori, meshes and fully connected networks). We also describe a methodology for designing OPCBs given a set of input parameters, including building blocks specifications as well as traffic demands. The onboard topology design methodology generates all the feasible designs within the topology families examined, following our proposed OPCB lay-out approach, and selects the optimal designs based on specific optimization criteria.

Performance analysis and layout design of optical blades for HPCs using the OptoBoard-Sim simulator

We demonstrate the Optical Board Simulator platform for optical PCB layout design and performance evaluation. Performance of two optical Blades is compared to CRAY-XK7 Blade for the FFTW benchmark, revealing significant throughput and latency improvements.

Electro-optic switches based on space switching of multiplexed WDM signals: Blocking vs non-blocking design trade-offs

Abstract Short distance optical interconnects are a promising solution for tackling the bandwidth and low energy consumption requirements of next generation Data Centers (DC) and High Performance Computing (HPC) systems. The realization of optical switching should offer scalability, allowing the interconnection of multiple racks and/or servers/compute nodes, and quick reconfiguration times. To this end, fast small-radix MicroRing Resonator (MRR)- and Mach-Zehnder Interferometer (MZI)-based space switching devices, capable of supporting multiple optical signals multiplexed through Wavelength Division Multiplexing (WDM) have been reported. Using such devices as building blocks we evaluate the performance of a number of simple electro-optic switch architectures based on successive wavelength selection, WDM multiplexing and space switching, attempting to achieve scalable switching fabrics with good throughput performance on average using little additional hardware and few switching stages, thus lower total insertion losses as well as lower power consumption. The price paid for such architectural simplicity is that it introduces additional constraints on the feasible permutation matrices of such switching fabrics, affecting performance for some traffic patterns. We discuss the trade-offs between performance and hardware requirements and based, on our findings, we propose alternative architectures that overcome these limitations.

Efficient bandwidth allocation in the NEPHELE optical/electrical datacenter interconnect

The NEPHELE data center interconnection network relies on hybrid electro-optical top-of-rack switches to interconnect servers over multi-wavelength optical rings. The bandwidth of the rings is shared, and an efficient utilization of the infrastructure calls for coordination in the time, space, and wavelength domains. To this end, we present offline and incremental dynamic resource assignment algorithms. The algorithms are suitable for implementation in a software defined network control plane, achieving efficient, collision-free, and on demand capacity use. Our simulation results indicate that the proposed algorithms can achieve high utilization and low latency in a variety of traffic scenarios that include hot spots and/or rapidly changing traffic.

Routing algorithm with smart energy management on VCSEL interconnected networks

Energy consumption and the associated costs constitute a crucial issue concerning the design and operation of data networks and data centers. Energy-awareness is required in all levels, ranging from physical layer to algorithms, protocols and applications. Architecture-wise, a promising solution for tackling the increasing energy requirements is the deployment of optics at both long and shorter distances, including within data centers. Vertical Cavity Surface Emitting Lasers (VCSEL) constitute a popular photonic transmitter technology used in numerous short-range applications, providing also the ability to reduce energy consumption by scaling down the transmission bit rate. In this study we focus on the algorithmic aspects of energy management by proposing an OptiMal EnerGy Aware (OMEGA) routing algorithm to operate in optical networks utilizing VCSEL-based opto-electronic links. The algorithm leverages the capability of VCSELs to adapt the energy dissipation with respect to the transmission bit rate. Simulation results, under various traffic patterns, show that OMEGA balances efficiently the traffic load over the networks links, resulting in high throughput and low energy consumption.

The role of optical interconnects in the design of data center architectures

Abstract Optics have already found their way inside the data center (DC) by replacing electrical links for rack-to-rack interconnections. To cope with both the energy and bandwidth requirements, and to overcome the limitations of the electrical interconnects, DCs will have to deploy optical technologies not only in rack-to-rack but also in smaller packaging modules and lower levels of their architectures, namely board-to-board, on-board, and even on-chip. In this chapter the architectures proposed for these packaging levels are briefly surveyed. Implementing and laying out complex topologies via optical waveguides on the on-board (Optical Printed Circuit Boards) level presents a number of issues that have to be considered when designing architectures for DCs. To address these issues, we outline appropriate and general lay-out strategies for both point-to-point and multi-point topologies for OPCBs.

Application-Oriented On-Board Optical Technologies for HPCs

The increased communication bandwidth demands of high performance computing (HPC) systems calling at the same time for reduced latency and increased power efficiency have designated optical interconnects as the key technology in order to achieve the target of exascale performance. In this realm, technology advances have to be accompanied by the development of corresponding design and simulation tools that support end-to-end system modeling in order to evaluate the performance benefits offered by optical components at system scale. In this paper, we present recent advances on electro-optical printed circuit boards (EOPCB) technology development pursued within the European FP7 PhoxTroT research program and directed toward system-scale performance benefits in real HPC workload applications. We report on high-density and multilayered EOPCBs together with all necessary building blocks for enabling true optical blade technology, including multimode polymer-based single- and dual-layer EOPCBs, a board-compatible optically interfaced router chip, and passive board-level connectors. We also demonstrate a complete optical blade design and evaluation software simulation framework called OptoHPC that tailors optical blade technology development toward optimized performance at HPC system scale, allowing for its validation with synthetic workload benchmark traffic profiles and for reliable comparison with existing HPC platforms. The OptoHPC simulator is finally utilized for evaluating and comparing a 384-node HPC system relying on optically enabled blades with the state-of-the-art Cray XK7 HPC network when performing with a range of synthetic workload traffic profiles, revealing the significant throughput and delay improvements that can be released through application-oriented optical blade technology.

Energy-optimal routing on VCSEL-based interconnected networks

Optical interconnection networks are being used in systems on chip, supercomputers, and datacenters, fueling exascale computing, big data, and artificial intelligence applications. The vertical cavity surface emitting laser (VCSEL) is a popular, mature, and cost-effective photonic transmitter technology that enables energy proportionality by allowing the links’ data rate and the associated power consumption to be adjusted. Our work assumes VCSEL-based optical interconnects and presents intelligent centralized and distributed mechanisms to jointly and optimally select the routes, the flow sizes, and the transmission powers needed to serve a given input traffic load, minimize the consumed energy, and optimize performance. For this purpose, we use a detailed VCSEL energy model and formulate the energy minimization problem as a constrained nonlinear multicommodity optimization problem, which is solved optimally with the proposed approaches. The simulation results, carried out under a variety of scenarios, show the efficiency of these methods in terms of throughput and energy consumption.

Collisions free scheduling in the NEPHELE hybrid electrical/optical datacenter interconnect

The NEPHELE datacenter network is divided into pods/clusters of racks and relies on hybrid electro-optical top-of rack switches that access an all-optical network consisting of WDM rings. To enable dynamic and efficient sharing of the optical resources and a collision-free network operation, the NEPHELE network is designed to operate in a slotted manner with a software-defined-network (SDN) based control plane. We describe the NEPHELE resource allocation problem, consider the wavelength conflicts on the shared WDM rings, translate them into resource allocation constraints and investigate the effect of these constraints on network performance. To do so, we define the worst-case traffic pattern and perform simulations to evaluate performance for the average traffic. Finally, we propose a variation to the NEPHELE architecture that reduces the effects of these wavelength conflict constraints on performance.

Multipoint architectures for on-board optical interconnects

Optical technology offers a high-bandwidth, energy-efficient solution for the increased communication requirements of data center and high-performance computing environments and is expected to be gradually deployed at all levels of the packaging hierarchy: from board-toboard, to on-board, andeven on-chipcommunication. In this work we focus on the on-board architecture level, outlining layout strategies for multipoint networks, such as single buses as well as a mesh of buses (MB) on optical printed circuit boards (OPCBs). Driven by that, we discuss how related point-to-point topologies, such as ameshof fully connected networks (MFCN), can be realized using wavelength division multiplexing (WDM) and multipoint layouts. We also provide closed-form formulas for the network capacity and the average internodal distance of these two mesh-like topology families,MB and MFCN, and demonstrate how the proposed techniques and formulas can be used for designing reconfigurable mesh-like architectures on OPCBs.