Results for “accelerated computing” — Page 60

DOE OSTI2021

Performance Improvements for the Griffin Transport Solvers

Griffin is a Multiphysics Object-Oriented Simulation Environment based reactor multiphysics analysis application jointly developed by Idaho National Laboratory and Argonne National Laboratory. Griffin includes a variety of deterministic radiation transport solvers for fixed source, k-eigenvalue, adjoint, and subcritical multiplication, as well as transient solvers for point-kinetics, improved quasi-static, and spatial dynamics. A code assessment performed in FY-20 identified two significant issues with the transport solvers in Griffin: first, the primary heterogeneous SN (discrete ordinates) transport solver based on continuous finite element methods required significant mesh refinement and higher memory usage compared to solvers based on the method of characteristic for equivalent accuracy. Second, the homogeneous PN (spherical harmonics expansion) transport solver did not adequately support polynomial refinement, which is a feature usually required for problems with spatial homogenization and pronounced streaming, typical in fast or gas-cooled reactor systems. To address the first issue, the development effort focused on the more promising discontinuous finite element method (DFEM)-based SN transport solver in Griffin. The addition of an asynchronous parallel transport sweeper and coarse mesh finite difference (CMFD) acceleration have rendered a superior heterogeneous SN transport capability for multiphysics problems that requires far less computing resources in terms of both CPU time and memory usage. This is demonstrated with typical thermal- and fast-spectrum reactor benchmark problems, including 2D Transient Reactor Test, 3D Advanced Burner Test Reactor (ABTR), and 2D and 3D Empire microreactor. For the second issue, the development effort focused on a new transport solver based on the hybrid finite element PN method (HFEM-PN), equivalent to the variational nodal method, as well as a new diffusion solver based on HFEM-Diffusion. This solver is intended for homogenized domains with multiphysics coupling (i.e., supports mesh displacement, seamless temperature feedback, etc.). Initial calculations with the HFEM-Diffusion implementation show very good parallel efficiency for the residual evaluations with the 2D ABTR benchmark. A future development effort will be centered on further improvements to the CMFD, HFEM-PN, and DFEM diffusion solvers to ensure Griffin meets performance and software quality assurance requirements for advanced reactor design and analysis.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

DOE OSTI2025

Radiation damage effects in beryllium for next generation neutrino beam targetry (Final Technical Report)

Current and future high-power accelerators put severe requirements on materials used for target and beam windows and target facilities have been recognized as a critical challenge in development of future particle accelerators. In accelerators, window and target materials are exposed to extreme conditions, which include bombardment with very high energy protons (1- 100 GeV) and thermomechanical shock waves. Radiation can cause direct damage in the material, and it leads to production of transmutation products (especially helium), both phenomena having a potential adverse effect on the stability and durability of the target/window material. At high enough temperatures, He can aggregate to form gas bubbles, which in turn cause significant dimensional changes (swelling), enable easy crack propagation, and eventually cause failure by fracture. On the other hand, if the temperature is too low, radiation damage accumulates in the form of internal defects (e.g., dislocations), leading to hardening and a decreased ductility of the material. In this project, we will focus on beryllium since it is considered to be one of the candidate materials for beam windows and targets in the next-generation proton accelerators, e.g., the Long Baseline Neutrino Facility (LBNF). Radiation effects in Be have been studied in the context of nuclear fusion reactor applications. However, key differences exist between reactor and accelerator conditions, including neutron vs. proton irradiation, continuous vs. pulsed beam flux, much higher energies of bombarding particles in accelerators, and higher operating temperatures for typical reactors. For example, the impact of beam pulsing on the radiation damage and the He bubble kinetics is largely unknown. While results obtained on Be from fusion research might not be directly transferrable to understanding target materials, there is an opportunity to bring state-of-the-art tools from materials research in nuclear reactors to aid design of target and beam window materials in high-power accelerators. To this end, the overarching goal of this project are to develop an experimentally-validated computational framework capable of predicting radiation damage evolution in beryllium relevant to beam window and target conditions, focusing on He bubble formation and growth as a function of irradiation temperature. Our model will be based on the cluster dynamics formalism, where size distribution of defects and He bubbles is simulated as a function of time, temperature, and radiation dose. Parameters for the model will be taken from published experiments and from high-fidelity atomistic simulations proposed in this project. In addition, we will carry out a series of targeted ex-situ and in-situ dual-beam experiments using low-energy protons to provide critical data for validation of the model on the effects of radiation on He clustering, He bubble distribution, and dislocation loop density/size in proton irradiated Be.

36 MATERIALS SCIENCE↗

DOE OSTI2025

SYCL for Performance Portability: Application Experience with Coupled Cluster Formalism in Quantum Chemistry on Exascale Systems

The exascale computing has brought unprecedented heterogeneity in node architectures, with systems such as Frontier and Aurora featuring diverse GPU accelerators, network connectivity among others. Ensuring performance portability across these platforms is a key challenge. To address this, we employ the SYCL programming model to develop portable, high-performance quantum chemistry workloads. As a representative application, we focus on the non-iterative Triples component of the coupled-cluster CCSD(T) method, a key driver in quantum chemistry. In this work, we report on our experience deploying SYCL-based implementations using both DPC++ and AdaptiveCPP across two flagship exascale platforms: OLCF Frontier with AMD MI250X GPUs and ALCF Aurora with Intel GPUs. Our results demonstrate that SYCL enables efficient, single-source implementations that scale to thousands of nodes, delivering performance on par with vendor-optimized HIP solutions. We highlight key insights into runtime behavior, kernel portability, and scaling characteristics, showing that SYCL offers a viable path for performance-portable computing.

Bagusetty, Abhishek [Argonne National Laboratory (↗

DOE OSTI2023

A Mobile Edge Computing Framework for Traffic Optimization At Urban Intersections Through Cyber-Physical Integration

The stop-and-go traffic pattern on urban roads often results in excessive energy consumption because of unnecessary vehicle braking, idling, and accelerations. With the widespread and increased use of automobiles, this traffic pattern creates many negative impacts (e.g., delayed travel time, air pollution, and additional carbon emission) on the sustainability of our cities. Taking advantage of the recent emerging Internet of Things (IoT) and edge computing paradigms, we propose a mobile edge computing framework that integrates the capability of real-time vehicle-to-infrastructure communication and intelligent speed optimization algorithms into a mobile app to optimize individual vehicles' driving speed at signalized intersections. The optimization aims to mitigate the stop-and-go traffic pattern and its undesirable consequences in urban transportation systems. The framework consists of (1) a cyberinfrastructure-enabled dynamic messaging system for retrieving and delivering real-time traffic and signal phase and timing information from IoT-connected signal controllers and sensors, (2) a real-time speed optimization algorithm for generating intelligent speed advisory using vehicle's information (e.g., GPS and driving directions from mobile sensing) and corresponding signal and traffic information, and (3) an ad-hoc mobile computing environment that converts drivers' smartphones into edge devices to host the speed optimization algorithms for enabling intelligent advisory on the vehicle's driving speed within signalized corridors. The paper presents the design and implementation of the proposed framework. Finally, we demonstrate the feasibility, usefulness, and energy-saving benefits of our proposed framework and its prototyping mobile app on urban transportation systems through traffic simulation, real-vehicle laboratory experiments, an evaluative survey, and field communication tests. The simulation-based energy evaluation results show that the 100% usage of the mobile app can achieve 24% energy savings in the transportation system.

33 ADVANCED PROPULSION SYSTEMS↗

DOE OSTI2024

ICED: An Integrated CGRA Framework Enabling DFVS-Aware Acceleration

oarse-grained reconfigurable arrays (CGRAs) are a promising solution to enable energy-efficient acceleration of applications from different domains. By leveraging reconfiguration at the functional level, they can adapt to significantly different computational patterns. Existing CGRA mapping approaches extract instruction-level parallelism, exploit loop-pipelining opportunities, guarantee the data dependency, and target high throughput of a given loop. However, the recurrence data-dependency in the DFG and the mismatch between required and available computing/communication resources complicate the mapping, and might lead to significant unbalances in the utilization of the CGRA's tiles. This results in wasted power for tiles with low utilization. Applying dynamic voltage and frequency scaling (DVFS) can potentially solve this challenge and improve energy efficiency by adjusting voltage and frequency of different tiles independently. CGRAs have also been successful in accelerating data-dependent streaming applications. However, in these applications, the execution time of each kernel in the pipeline might dynamically vary depending on the characteristics of the input. This also leads to under-utilization of resources for the dynamically changing kernels that do not limit the application throughput. DVFS can also improve energy efficiency for these applications by dynamically changing the voltage and frequency levels of tiles that host non performance-constraining kernels. This paper proposes ICEDTEA -- an integrated DVFS-aware framework to map applications on CGRAs that support power islands. ICEDTEA proposes a CGRA architecture supporting DVFS islands at varying granularity (from a single tile to a group of tiles) and the related DVFS-aware compilation and mapping toolchain. ICEDTEA is the first work that introduces DVFS support for spatio-temporal CGRAs at power-island levels. The experimental evaluation shows that ICEDTEA improves average utilization by 2.3$\times$ and energy-efficiency by 1.32$\times$ over a conventional CGRA. With streaming applications, ICEDTEA improves energy efficiency by 1.12$\times$ over a state-of-the-art CGRA that introduces partial dynamic reconfiguration to adapt to variations in kernels' throughput.

Tan, Cheng↗

DOE OSTI2020

Accelerating geostatistical modeling using geostatistics-informed machine Learning

Ordinary Kriging (OK) is a popular geostatistical algorithm for spatial interpolation and estimation. The computational complexity of OK changes quadratically and cubically for memory and speed, respectively, given the number of data. Therefore, it is computationally intensive and also challenging to process a large set of data, especially in three-dimensional (3D) cases. This paper develops a geostatistics-informed machine learning (GIML) model to improve the efficiency of OK by reducing the number of points required to be estimated using OK. Specifically, only a very few of the unknown points are estimated by OK to get the weights and estimations, which are used as the training dataset. Moreover, the governing equations of OK are used to guide our proposed machine learning to better reproduce the spatial distributions. Our results show that the proposed GIML can reduce the computational time of OK by at least one order of magnitude. The effectiveness of the GIML is evaluated and compared using a 2D case. Furthermore, we demonstrate its efficiency and robustness by considering a different number of training samples on various 3D simulation grids.

58 GEOSCIENCES↗

DOE OSTI2022

Denoising Convolutional Networks to Accelerate Detector Simulation

The high accuracy of detector simulation is crucial for modern particle physics experiments. However, this accuracy comes with a high computational cost, which will be exacerbated by the large datasets and complex detector upgrades associated with next-generation facilities such as the High Luminosity LHC. We explore the viability of regression-based machine learning (ML) approaches using convolutional neural networks (CNNs) to "denoise" faster, lower-quality detector simulations, augmenting them to produce a higher-quality final result with a reduced computational burden. The denoising CNN works in concert with classical detector simulation software rather than replacing it entirely, increasing its reliability compared to other ML approaches to simulation. We obtain promising results from a prototype based on photon showers in the CMS electromagnetic calorimeter. Future directions are also discussed.

Banerjee, Sunanda↗

DOE OSTI2021

Fast ion transport by sawtooth instability in the presence of ICRF-NBI synergy in JET plasmas

JET experiments have shown that the three-ion scenarios using waves in the ion cyclotron range of frequencies (ICRF) is an efficient way to build fast ion population through beam ion acceleration by radio frequency (RF) waves. Here, such a heating scheme is applied to plasmas with at least two thermal ion species. Analysis of mixed discharges with complex heating schemes requires a workflow that allows to model thermal and fast ion transport consistently. This paper is dedicated to modelling of a mixed plasma discharge with significant fraction of fast ions and contributes to development of fast ion transport models. For interpretive analysis with the TRANSP code a JET hydrogen-deuterium (HD) plasma discharge with neutral beam injection (NBI) and ICRF heating has been chosen. The task is complicated by NBI-ICRF synergy and plasma magnetohydrodynamic activity, like sawtooth crashes. D beam ions accelerated by RF waves form a high energy tail in fast ion distribution. Significant difference between the neutron rate computed by TRANSP and measured one is observed if the same diffusivity for electrons and ions is assumed. Sensitivity studies show that uncertainties in input plasma parameters and thermal ion transport models are crucial for modelling mixed plasma discharges and increased D transport is required to reach the plasma composition consistent with diagnostic measurements at the plasma edge. Fast ion redistribution by a sawtooth instability is characterised by non-resonant transport due to reconnection of magnetic field lines and resonant transport caused by resonance interaction between the instability and fast ions. With ORBIT simulations it has been shown that resonant interaction strongly affects fast ions of high energies, like beam ions accelerated by RF waves and fusion products. For the considered case, fast ion profiles simulated by ORBIT remain peaked after the sawtooth crashes.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

DOE OSTI2023

The VTK-m Users' Guide (V.2.0)

High-performance computing relies on ever finer threading. Advances in processor technology include ever greater numbers of cores, hyperthreading, accelerators with integrated blocks of cores, and special vectorized instructions, all of which require more software parallelism to achieve peak performance. Traditional visualization solutions cannot support this extreme level of concurrency. Extreme scale systems require a new programming model and a fundamental change in how we design algorithms. To address these issues we created VTK-m: the visualization toolkit for multi-/many-core architectures. VTK-m supports a number of algorithms and the ability to design further algorithms through a top-down design with an emphasis on extreme parallelism. VTK-m also provides support for finding and building links across topologies, making it possible to perform operations that determine manifold surfaces, interpolate generated values, and find adjacencies. Although VTK-m provides a simplified high-level interface for programming, its template-based code removes the overhead of abstraction. VTK-m simplifies the development of parallel scientific visualization algorithms by providing a framework of supporting functionality that allows developers to focus on visualization operations. Consider the listings in Figure 1.1 that compares the size of the implementation for the Marching Cubes algorithm in VTK-m with the equivalent reference implementation in the CUDA software development kit. Because VTK-m internally manages the parallel distribution of work and data, the VTK-m implementation is shorter and easier to maintain. Additionally, VTK-m provides data abstractions not provided by other libraries that make code written in VTK-m more versatile.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Noisy-Intermediate-Scale Quantum Electromagnetic Transients Program

Quantum-empowered electromagnetic transients program (QEMTP) is a promising paradigm for tackling EMTP's computational burdens. Nevertheless, no existing studies truly achieve a practical and scalable QEMTP operable on today's noisy-intermediate-scale quantum (NISQ) computers. The strong reliance on noise-free and fault-tolerant quantum devices--which appears to be decades away--hinder practical applications of current QEMTP methods. Here, we devise a NISQ-QEMTP methodology which for the first time transitions the QEMTP operations from ideal, noise-free quantum simulators to real, noisy quantum computers. The main contributions lie in: (1) a shallow-depth QEMTP quantum circuit for mitigating noises on NISQ quantum devices; (2) practical QEMTP linear solvers incorporating executable quantum state preparation and measurements for nodal voltage computations; (3) a noise-resilient QEMTP algorithm leveraging quantum resources logarithmically scaled with power system dimension; (4) a quantum shifted frequency analysis (QSFA) for accelerating QEMTP by exploiting dynamic phasor simulations with larger time steps; (5) a systematical analysis on QEMTPs performance under various noisy quantum environments. Extensive experiments systematically verify the accuracy, efficacy, universality and noise-resilience of QEMTP on both noise-free simulators and IBM real quantum computers.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗

DOE OSTI2021

HiPACE++ v0.1

Hipace++ is the new implementation of the DESY-LBNL code Hipace (not distributed nor licensed) for plasma acceleration, relying on the quasi-static particle-in-cell (PIC) method, enabling fast and accurate simulations of plasma accelerators for problems satisfying the quasi-static approximation. Plasma acceleration is a priority in the ATAP division, and such problems represent a wide range of parameters. Hipace++ is an HPC (high-performance computing) code running on supercomputers and it provides the first efficient implementation of the quasi-static PIC method on Graphic Processing Units (GPU), showing 10x speedup and using much less resources than CPU implementations for production simulations. It currently runs on Nvidia GPUs and will ultimately be portable to more accelerated HPC platforms. The porting strategy is novel, and can be adapted to other flavors of quasi-static PIC. Besides, Hipace++ aims at using mesh refinement for higher efficiency, thus enabling simulations unfeasible today. For this purpose, Hipace++ depends on the open-source library AMReX (Adaptive Mesh Refinement at EXascale) developed at LBNL.

Vay, Jean-Luc↗

DOE OSTI2021

VTK-m User's' Guide (V.1.7)

High-performance computing relies on ever finer threading. Advances in processor technology include ever greater numbers of cores, hyperthreading, accelerators with integrated blocks of cores, and special vectorized instructions, all of which require more software parallelism to achieve peak performance. Traditional visualization solutions cannot support this extreme level of concurrency. Extreme scale systems require a new programming model and a fundamental change in how we design algorithms. To address these issues we created VTK-m: the visualization toolkit for multi-/many-core architectures. VTK-m supports a number of algorithms and the ability to design further algorithms through a top-down design with an emphasis on extreme parallelism. VTK-m also provides support for finding and building links across topologies, making it possible to perform operations that determine manifold surfaces, interpolate generated values, and find adjacencies. Although VTK-m provides a simplified high-level interface for programming, its template-based code removes the overhead of abstraction. VTK-m simplifies the development of parallel scientific visualization algorithms by providing a framework of supporting functionality that allows developers to focus on visualization operations. Consider the listings in Figure 1.1 that compares the size of the implementation for the Marching Cubes algorithm in VTK-m with the equivalent reference implementation in the CUDA software development kit. Because VTK-m internally manages the parallel distribution of work and data, the VTK-m implementation is shorter and easier to maintain. Additionally, VTK-m provides data abstractions not provided by other libraries that make code written in VTK-m more versatile.This book includes contributions from the VTK-m community including the VTK-m development team and the user community.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Portability for GPU-accelerated molecular docking applications for cloud and HPC: can portable compiler directives provide performance across all platforms?

High-throughput structure-based screening of drug-like molecules has become a common tool in biomedical research. Recently, acceleration with graphics processing units (GPUs) has provided a large performance boost for molecular docking programs. Both cloud and high-performance computing (HPC) resources have been used for large screens with molecular docking programs; while NVIDIA GPUs have dominated cloud and HPC resources, new vendors such as AMD and Intel are now entering the field, creating the problem of software portability across different GPUs. Ideally, software productivity could be maximized with portable programming models that are able to maintain high performance across architectures. While in many cases compiler directives have been used as an easy way to offload parallel regions of a CPU-based program to a GPU accelerator, they may also be an attractive programming model for providing portability across different GPU vendors, in which case the porting process may proceed in the reverse direction: from low-level, architecture-specific code to higher-level directive-based abstractions. MiniMDock is a new mini-application (miniapp) designed to capture the essential computational kernels found in molecular docking calculations, such as are used in phar-maceutical drug discovery efforts, in order to test different solutions for porting across GPU architectures. Here we extend MiniMDock to GPU offloading with OpenMP directives, and compare to performance of kernels using CUDA and HIP on NVIDIA and AMD GPUs, respectively, as well as across different compilers, exploring performance bottlenecks. We document this reverse-porting process, from highly optimized device code to a higher-level version using directives, compare code structure, and describe barriers that were overcome in this effort.

Thavappiragasam, Mathialakan↗

DOE OSTI2022

The VTK-m Users' Guide (V.1.9)

High-performance computing relies on ever finer threading. Advances in processor technology include ever greater numbers of cores, hyperthreading, accelerators with integrated blocks of cores, and special vectorized instructions, all of which require more software parallelism to achieve peak performance. Traditional visualization solutions cannot support this extreme level of concurrency. Extreme scale systems require a new programming model and a fundamental change in how we design algorithms. To address these issues we created VTK-m: the visualization toolkit for multi-/many-core architectures. VTK-m supports a number of algorithms and the ability to design further algorithms through a top-down design with an emphasis on extreme parallelism. VTK-m also provides support for finding and building links across topologies, making it possible to perform operations that determine manifold surfaces, interpolate generated values, and find adjacencies. Although VTK-m provides a simplified high-level interface for programming, its template-based code removes the overhead of abstraction. VTK-m simplifies the development of parallel scientific visualization algorithms by providing a framework of supporting functionality that allows developers to focus on visualization operations. Consider the listings in Figure 1.1 that compares the size of the implementation for the Marching Cubes algorithm in VTK-m with the equivalent reference implementation in the CUDA software development kit. Because VTK-m internally manages the parallel distribution of work and data, the VTK-m implementation is shorter and easier to maintain. Additionally, VTK-m provides data abstractions not provided by other libraries that make code written in VTK-m more versatile.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2021

Position Papers for the ASCR Workshop on Reimagining Codesign

On behalf of the Advanced Scientific Computing Research (ASCR) program in the US Department of Energy (DOE) Office of Science, we are organizing a Workshop on Reimagining Codesign (ReCoDe). Codesign is the process of jointly designing interoperating components of a computing system—in particular: applications, algorithms, system software, programming models, and the hardware on which they run. The goal is to maximize the overall performance, efficiency, and other desirable qualities of the system as a whole. Codesign is a standard methodology in the embedded-systems community, where space, power, and cost constraints are commonly pitted against execution speed for a tightly constrained feature set. Over the last decade, the DOE has invested in codesign efforts to foster the development of exascale computing systems for broad classes of scientific and engineering applications. The ReCoDe workshop hopes to explore how scientific applications of interest to the DOE can be accelerated through close interactions with hardware designers and software-stack developers, in which all components adapt to each other’s requirements and constraints. We want to answer the question of what are the key tools and methodologies for accomplishing codesign in today’s computing landscape, and what will be the highest impact targets for meeting DOE’s emerging mission requirements. This workshop aims to bring together DOE, industry, and academia to identify opportunities to build on past codesign successes and identify new areas that are either emerging or that may need reimagining for the future. We want to continue to find opportunities that can be pursued as a joint effort and continue to break down the traditional customer/vendor dichotomy with true partnerships. From this work, DOE will benefit from increased application performance relative to what stock hardware or existing general-purpose roadmaps can provide, and vendors will benefit from expanding their hardware’s capabilities to address needs they might have not otherwise anticipated and thereby create more widespread interest in their products. The workshop will be structured around a set of breakout sessions, with every attendee expected to participate actively in the discussions. Afterward, workshop attendees—from DOE, industry, and academia—will produce a report for ASCR that summarizes the findings made during the workshop.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

Accelerating Surface Radiation Content to Investigate the Impact of Radon Progeny on Superconducting Qubits

Ionizing radiation in the form of $α$, $β$, $γ$, and additional high-energy particles can induce decoherence via phonon and quasiparticle poisoning in superconducting qubits. Recent studies have explored this effect using cosmic rays or controlled radioactive sources held in the proximity of a qubit package, and have concluded that reductions in such ``external'' environmental radiation may benefit stable operation of qubit devices. However, the effect of long-lived, unstable daughters of $^{222}$Rn that ``plate out'' directly on device and packaging surfaces has not been as extensively explored. This plate-out process, well-known to the dark matter direct detection field, occurs throughout the fabrication and testing lifecycle of a device and (separately) its packaging, and produces a local source of $α$-decays which can remain active for decades. As this scales with chip area, understanding and managing this source of ionizing radiation is relevant for successfully scaling quantum computing architectures to larger numbers of qubits in a radiation-robust way. We present a setup capable of accelerating and enhancing radon daughter plateout by a factor of $7\times10^4$ over ambient, in order to study, \textit{in situ}, the impact of these events on superconducting qubits. We also provide outlook on the potential impact of this source of ionizing radiation on current and future qubit arrays.

Poudel, Sagar S. [South Dakota Sch. Mines Tech.]↗

DOE OSTI2021

Scaling Inference Using Triton to Accelerate Particle Physics at the LHC and DUNE

DUNE and the LHC experiments consist of unique and cutting-edge particle detectors that create massive, complex, and rich datasets with billions of events. They require sophisticated algorithms to reconstruct and interpret the data. Modern machine learning algorithms provide a powerful toolset to detect and classify particles, from familiar image processing convolutional neural networks to newer graph neural network architectures. A full reconstruction of these particle collisions requires novel approaches to handle the computing challenge of processing so much raw data. In a series of studies, physicists from Fermilab, CERN, and university groups explored how to accelerate their data processing using the Triton Inference Server.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

DOE OSTI2022

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↗