Advances in data-driven applications, particularly artificial intelligence and deep learning, are driving the explosive growth of computation and communication in today’s data centers and high-performance computing (HPC) systems. Increasingly, system performance is not constrained by the compute speed at individual nodes, but by the data movement between them. This calls for innovative architectures, smart connectivity, and extreme bandwidth densities in interconnect designs. Silicon photonics technology leverages mature complementary metal-oxide-semiconductor (CMOS) manufacturing infrastructure and is promising for low cost, high-bandwidth, and reconfigurable interconnects. Flexible and high-performance photonic switched architectures are capable of improving the system performance. The work in this dissertation explores various photonic interconnected systems and the associated optical switching functionalities, hardware platforms, and novel architectures. It demonstrates the capabilities of silicon photonics to enable efficient deep learning training.
We first present field programmable gate array (FPGA) based open-loop and closed-loop control for optical spectral-and-spatial switching of silicon photonic cascaded micro-ring resonator (MRR) switches. Our control achieves wavelength locking at the user-defined resonance of the MRR for optical unicast, multicast, and multiwavelength-select functionalities. Digital-to-analog converters (DACs) and analog-to-digital converters (ADCs) are necessary for the control of the switch. We experimentally demonstrate the optical switching functionalities using an FPGA-based switch controller through both traditional multi-bit DAC/ADC and novel single-wired DAC/ADC circuits. For system-level integration, interfaces to the switch controller in a network control plane are developed. The successful control and the switching functionalitiesachieved are essential for system-level architectural innovations as presented in the following sections.
Next, this thesis presents two novel photonic switched architectures using the MRR-based switches. First, a photonic switched memory system architecture was designed to address memory challenges in deep learning. The reconfigurable photonic interconnects provide scalable solutions and enable efficient use of disaggregated memory resources for deep learning training. An experimental testbed was built with a processing system and two remote memory nodes using silicon photonic switch fabrics and system performance improvements were demonstrated. The collective results and existing high-bandwidth optical I/Os show the potential of integrating the photonic switched memory to state-of-the-art processing systems. Second, the scaling trends of deep learning models and distributed training workloads are challenging network capacities in today’s data centers and HPCs. A system architecture that leverages SiP switch-enabled server regrouping is proposed to tackle the challenges and accelerate distributed deep learning training. An experimental testbed with a SiP switch-enabled reconfigurable fat tree topology was built to evaluate the network performance of distributed ring all-reduce and parameter server workloads. We also present system-scale simulations. Server regrouping and bandwidth steering were performed on a large-scale tapered fat tree with 1024 compute nodes to show the benefits of using photonic switched architectures in systems at scale.
Finally, this dissertation explores high-bandwidth photonic interconnect designs for disaggregated systems. We first introduce and discuss two disaggregated architectures leveraging extreme high bandwidth interconnects with optically interconnected computing resources. We present the concept of rack-scale graphics processing unit (GPU) disaggregation with optical circuit switches and electrical aggregator switches. The architecture can leverage the flexibility of high bandwidth optical switches to increase hardware utilization and reduce application runtimes. A testbed was built to demonstrate resource disaggregation and defragmentation. In addition, we also present an extreme high-bandwidth optical interconnect accelerated low-latency communication architecture for deep learning training. The disaggregated architecture utilizes comb laser sources and MRR-based cross-bar switching fabrics to enable an all-to-all high bandwidth communication with a constant latency cost for distributed deep learning training. We discuss emerging technologies in the silicon photonics platform, including light source, transceivers, and switch architectures, to accommodate extreme high bandwidth requirements in HPC and data center environments. A prototype hardware innovation - Optical Network Interface Cards (comprised of FPGA, photonic integrated circuits (PIC), electronic integrated circuits (EIC), interposer, and high-speed printed circuit board (PCB)) is presented to show the path toward fast lanes for expedited execution at 10 terabits.
Taken together, the work in this dissertation demonstrates the capabilities of high-bandwidth silicon photonic interconnects and innovative architectural designs to accelerate deep learning training in optically connected data center and HPC systems.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/pj5h-n838 |
Date | January 2022 |
Creators | Zhu, Ziyi |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.0016 seconds