THESIS
2019
xiv, 122 pages : illustrations ; 30 cm
Abstract
The rapid growth of big-data applications, cloud services and scientific applications is resulting
in tremendous data traffic in high-performance computing (HPC) systems and data centers.
Conventional electrical switches are currently facing challenges to meet the performance requirement
under tight energy and thermal constraints.
To accomplish high-bandwidth and low-latency communications among hundreds of nodes
with low energy consumption, we propose DRAGON and FODON, two scalable integrated
high-radix optical switch fabrics. Their topology and routing algorithm are discussed, and a
formal proof for the strictly non-blocking property of DRAGON is presented. Analyses suggest
DRAGON can achieve lower loss and crosstalk compared to other strictly non-blocking switch
fabrics. The...[
Read more ]
The rapid growth of big-data applications, cloud services and scientific applications is resulting
in tremendous data traffic in high-performance computing (HPC) systems and data centers.
Conventional electrical switches are currently facing challenges to meet the performance requirement
under tight energy and thermal constraints.
To accomplish high-bandwidth and low-latency communications among hundreds of nodes
with low energy consumption, we propose DRAGON and FODON, two scalable integrated
high-radix optical switch fabrics. Their topology and routing algorithm are discussed, and a
formal proof for the strictly non-blocking property of DRAGON is presented. Analyses suggest
DRAGON can achieve lower loss and crosstalk compared to other strictly non-blocking switch
fabrics. The crossing analyses and loss analyses also reveal that FODON can reduce a considerable
number of waveguide crossings and improve the scalability significantly compared to other
optical switch fabrics based on multi-stage indirect networks.
We also present a cross-layer optimization framework, CLOSO, to investigate optical switch
fabrics based on physical device models together with switch fabric models. It helps efficiently
optimize the loss of optical switch fabrics and obtain the corresponding loss parameters of
optical switch devices with optimal device design parameters.
Furthermore, we perform comparative analyses of DRAGON, FODON, Benes and arrayed
waveguide grating router (AWGR) at the system level, regarding the latency, throughput and energy consumption. The comparisons show the advantage of DRAGON in low latency and high
throughput, especially for cases where large packets dominate. FODON, even as a blocking
space switch, performs comparably to the rearrangeable non-blocking switch yet with much
smaller energy consumption. Further design explorations on the packet length, virtual output
queues (VOQs), synchronization overhead and wavelength division multiplexing (WDM)
channels also pave the way to the design optimization of chip/rack-level optical interconnection
systems.
Post a Comment