Search NASA⌕ Search

Engineering topics

Bergman, Keren

Publications and source records attributed to Bergman, Keren.

At least 19 records

Scaling comb-driven resonator-based DWDM silicon photonic links to multi-Tb/s in the multi-FSR regime

The use of chip-based micro-resonator Kerr frequency combs in conjunction with dense wavelength-division multiplexing (DWDM) enables massively parallel intensity-modulated direct-detection data transmission with low energy consumption. Resonator-based modulators and filters used in such systems can limit the number of usable wavelength channels due to practical constraints on the maximum achievable free spectral range (FSR). In this work, we introduce the design of multi-Tb/s comb-driven resonator-based silicon photonic links by leveraging the multi-FSR regime. We demonstrate the viability of the link architecture with yield estimates that are supported by extensive wafer-scale measurements of 704 micro-resonators fabricated in a commercial complementary metal–oxide–semiconductor foundry. We show that a 2.80 Tb/s link is realizable with a ≥6 σ yield (∼99.999%), and that aggregate bandwidths of 3.76 Tb/s and 4.72 Tb/s are possible if yield targets are relaxed (3 σ and 1 σ , respectively). All designs represent a 1.94−3.28× boost to aggregate link bandwidth while maintaining BER≤10 −10 performance, with a theoretical bandwidth of 10.51 Tb/s being possible for sufficiently robust resonators. We use high-speed BER measurements to inform co-optimization of data rate and aggressor spacing ( λ ag ), limiting any additional loss-based power penalties to off-resonance insertion loss (IL) and routing loss. This work demonstrates that, through the multi-FSR regime, there is a clear path toward Kerr comb-driven ultra-broadband, high bandwidth silicon photonic links that can support next-generation data centers and high-performance computers.

James, Aneek (ORCID:0000000262527807)↗

Fabrication-robust silicon photonic devices in standard sub-micron silicon-on-insulator processes

Perturbations to the effective refractive index from nanometer-scale fabrication variations in waveguide geometry plague high index-contrast photonic platforms; this includes the ubiquitous sub-micron silicon-on-insulator (SOI) process. Such variations are particularly troublesome for phase-sensitive devices, such as interferometers and resonators, which exhibit drastic changes in performance as a result of these fabrication-induced phase errors. In this Letter, we propose and experimentally demonstrate a design methodology for dramatically reducing device sensitivity to silicon width variations. We apply this methodology to a highly phase-sensitive device, the ring-assisted Mach–Zehnder interferometer (RAMZI), and show comparable performance and footprint to state-of-the-art devices, while substantially reducing stochastic phase errors from etch variations. This decrease in sensitivity is directly realized as energy savings by significantly reducing the required corrective thermal tuning power, providing a promising path toward ultra-energy-efficient large-scale silicon photonic circuits.

Rizzo, Anthony (ORCID:000000034752797X)↗

Performance trade-offs in reconfigurable networks for HPC

Designing efficient interconnects to support high-bandwidth and low-latency communication is critical toward realizing high performance computing (HPC) and data center (DC) systems in the exascale era. At extreme computing scales, providing the requisite bandwidth through overprovisioning becomes impractical. These challenges have motivated studies exploring reconfigurable network architectures that can adapt to traffic patterns at runtime using optical circuit switching. Despite the plethora of proposed architectures, surprisingly little is known about the relative performances and trade-offs among different reconfigurable network designs. We aim to bridge this gap by tackling two key issues in reconfigurable network design. First, we study how cost, power consumption, network performance, and scalability vary based on optical circuit switch (OCS) placement in the physical topology. Specifically, we consider two classes of reconfigurable architectures: one that places OCSs between top-of-rack (ToR) switches—ToR-reconfigurable networks (TRNs)—and one that places OCSs between pods of racks—pod-reconfigurable networks (PRNs). Second, we tackle the effects of reconfiguration frequency on network performance. Our results, based on network simulations driven by real HPC and DC workloads, show that while TRNs are optimized for low fan-out communication patterns, they are less suited for carrying high fan-out workloads. PRNs exhibit better overall trade-off, capable of performing comparably to a fully non-blocking fat tree for low fan-out workloads, and significantly outperform TRNs for high fan-out communication patterns.

Teh, Min Yee↗

A Case For Intra-rack Resource Disaggregation in HPC

The expected halt of traditional technology scaling is motivating increased heterogeneity in high-performance computing (HPC) systems with the emergence of numerous specialized accelerators. As heterogeneity increases, so does the risk of underutilizing expensive hardware resources if we preserve today’s rigid node configuration and reservation strategies. This has sparked interest in resource disaggregation to enable finer-grain allocation of hardware resources to applications. However, there is currently no data-driven study of what range of disaggregation is appropriate in HPC. To that end, we perform a detailed analysis of key metrics sampled in NERSC’s Cori, a production HPC system that executes a diverse open-science HPC workload. In addition, we profile a variety of deep-learning applications to represent an emerging workload. We show that for a rack (cabinet) configuration and applications similar to Cori, a central processing unit with intra-rack disaggregation has a 99.5% probability to find all resources it requires inside its rack. In addition, ideal intra-rack resource disaggregation in Cori could reduce memory and NIC resources by 5.36% to 69.01% and still satisfy the worst-case average rack utilization.

97 MATHEMATICS AND COMPUTING↗

Distributed deep learning training using silicon photonic switched architectures

The scaling trends of deep learning models and distributed training workloads are challenging network capacities in today’s datacenters and high-performance computing (HPC) systems. We propose a system architecture that leverages silicon photonic (SiP) switch-enabled server regrouping using bandwidth steering to tackle the challenges and accelerate distributed deep learning training. In addition, our proposed system architecture utilizes a highly integrated operating system-based SiP switch control scheme to reduce implementation complexity. To demonstrate the feasibility of our proposal, we built an experimental testbed with a SiP switch-enabled reconfigurable fat tree topology and evaluated the network performance of distributed ring all-reduce and parameter server workloads. The experimental results show up to 3.6× improvements over the static non-reconfigurable fat tree. Our large-scale simulation results show that server regrouping can deliver up to 2.3× flow throughput improvement for a 2× tapered fat tree and a further 11% improvement when higher-layer bandwidth steering is employed. The collective results show the potential of integrating SiP switches into datacenters and HPC systems to accelerate distributed deep learning training.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Optically connected memory for disaggregated data centers

Recent advances in integrated photonics enable the implementation of reconfigurable, high-bandwidth, and low energy-per-bit interconnects in next-generation data centers. We propose and evaluate an Optically Connected Memory (OCM) architecture that disaggregates the main memory from the computation nodes in data centers. OCM is based on micro-ring resonators (MRRs), and it does not require any modification to the DRAM memory modules. We calculate energy consumption from real photonic devices and integrate them into a system simulator to evaluate performance. Here, our results show that (1) OCM is capable of interconnecting four DDR4 memory channels to a computing node using two fibers with 1.02 pJ energy-per-bit consumption and (2) OCM performs up to 5.5× faster than a disaggregated memory with 40G PCIe NIC connectors to computing nodes.

97 MATHEMATICS AND COMPUTING↗

PINE: Photonic Integrated Networked Energy efficient datacenters

We review the motivation, goals, and achievements of the Photonic Integrated Networked Energy efficient datacenter (PINE) project, which is part of the Advanced Research Projects Agency–Energy (ARPA-E) ENergy-efficient Light-wave Integrated Technology Enabling Networks that Enhance Dataprocessing (ENLITENED) program. Here, the PINE program leverages the unique features of photonic technologies to enable alternative mega-datacenters and high-performance computing (HPC) system architectures that deliver more substantial energy efficiency improvements than can be achieved through link energy efficiency alone. In phase 1 of the program, the PINE system architecture demonstrated an average factor of 2.2× improvement in transactions/joule across a diverse set of HPC and datacenter applications. In phase 2, PINE will demonstrate an aggressive 1.0 pJ/bit total link budget with high-bandwidth-density dense wavelength-division multiplexing (DWDM) links to enable additional 2.5× or more efficiency gains through deep resource disaggregation.

42 ENGINEERING↗

TURBO: Terabits/s Using Reconfigurable Bandwidth Optics (Final Report)

Large scale, big science applications are generating petabytes of data that need to be shared among multiple locations. Moving a petabyte of data at 100 Gb/s takes 22 hours. Reducing this to single digit hours or below is clearly attractive and with future applications anticipating exabyte scale transfers, multi-terabit per second capacities are needed. This scale of capacity is available using optical systems. Core optical systems have aggregate per fiber capacities on the order of 10 Tb/s and more than 100 Tb/s system experiments have been realized in the lab. But large optical capacities are not just reserved for the core as similar system capacities are available in metro networks right up to the enterprise, campus, or data center. In general, the same technology is used today in the metro area as in the long-haul networks—reconfigurable optical add drop multiplexing (ROADM) node-based wavelength division multiplexing (WDM) systems. Data centers already support massive internal capacities on the scale of 100’s of Tb/s. As new photonic integrated technologies mature, for example through initiatives such as the Integrated Photonics Innovative Manufacturing Institute (IP-IMI), the cost of optical interfaces is expected to decrease, and higher capacity and higher performance interfaces will become more affordable. This set of circumstances creates the potential that Tb/s capacities will be available at the enterprise and campus level to support large scale science network applications.

42 ENGINEERING↗

Push—pull microring-assisted space-and-wavelength selective switch

We introduce a novel design of a space-and-wavelength selective switch based on microring-assisted Mach–Zehnder interferometers. Multiple pairs of overcoupled microring resonators are incorporated as efficient and narrowband phase shifters and are driven in push–pull scheme. We design and demonstrate a 2 × <#comment/> 2 × <#comment/> 2 λ <#comment/> elementary switch block with full spatial and spectral switching capabilities. The switching device’s cross talk suppression and extinction ratio both exceed 21 dB. We measure over 75 GHz optical bandwidth per channel and less than 1.5 dB power penalty at 10 − <#comment/> 9 BER when two 32 Gbps on–off keying signals are loaded simultaneously. This new class of switching elements can further enable compact and high-performance space-and-wavelength selective switch fabrics without the need for wavelength (de)multiplexers or parallel switching planes.

Huang, Yishen (ORCID:0000000278710462)↗

Flexspander: augmenting expander networks in high-performance systems with optical bandwidth steering

Communication efficiency is one of the deciding factors in determining many of today’s high-performance computing (HPC) applications. Traditionally, HPC systems have been on static network topologies, making them inflexible to the variety of skewed traffic demands that may arise due to the spatial locality inherent in many applications. To handle traffic locality, researchers have proposed integrating optical circuit switches (OCSs) into the network architecture, which reconfigures the network topology to alter and dynamically adapt to the predicted traffic. In this paper, we present a novel reconfigurable network topology called Flexspander. Beyond offering a flexible interconnect, Flexspander also offers full flexibility in terms of construction and can be built with any arbitrary combination of commercial electrical packet switches and OCSs. We evaluate Flexspander performance through extensive simulations with multiple network traces, and our results show improved performance for Flexspander over currently proposed static and reconfigurable topologies in terms of the flow completion time.

Teh, Min Yee↗

Future High Performance Computing Capabilities: Summary Report of the Advanced Scientific Computing Advisory Committee (ASCAC) Subcommittee

The ASCAC Subcommittee on Future High Performance Computing (HPC) Capabilities has reviewed opportunities and challenges related to the most promising technologies that are currently planned for the post-exascale (2020's) and post-Moore (2030's and beyond) timeframes. We briefly summarize in this report the key findings and recommendations from this review, from the perspective of planning and research directions that need to be given priority to prepare for the very significant challenges that await us in the post-Moore computing era. An overarching concern that emerged from the subcommittee's deliberations is that DOE has lost considerable momentum in funding and sustaining a research pipeline in the applied math and computer science areas that should have been the "seed corn" for preparing for these future challenges, and it is therefore critical to correct this gap as soon as possible. While the subcommittee understands the paramount importance of DOE's commitment to deliver exascale capabilities, we believe that it is essential for DOE ASCR to fund research and development that looks beyond the Exascale Computing Project (ECP) horizon so as to ensure our nation's continued leadership in HPC and we were glad to learn about the ASCAC exascale transition subcommittee in this regard.

97 MATHEMATICS AND COMPUTING↗