Equalizer State Caching for Fast Data Recovery in Optically-Switched Data Center Networks

Abstract

Optical switching offers the potential to significantly scale the capacity of data center networks (DCN) with a simultaneous reduction in switching time and power consumption. Previous research has shown that end-to-end switching time, which is the sum of the switch configuration time and the clock and data recovery (CDR) locking time, should be kept within a few nanoseconds for high network throughput. This challenge of low switching time has motivated research into fast optical switches, ultra-fast clock and amplitude recovery techniques. Concurrently, the data rate between server-to-server and server-to-switch interconnect is increasing drastically from the current 100 Gb/s (4×25 Gb/s) to 400 Gb/s and beyond, motivating the use of high order formats such as 50-GBaud four-level pulse-amplitude modulation (PAM-4) for signalling. Since PAM-4 is more sensitive to noise and distortion, digital equalizers are generally needed to compensate for impairments such as transceiver frequency rolloff, dispersion and optical filtering, adding additional time for equalizer adaptation and power consumption that are undesired for fast optical switching systems. Here we propose and investigate an equalizer state caching technique that reduces equalizer adaptation time and computation power consumption for fast optical switching systems, underpinning optically-switched DCNs using high baud rate and impairment-sensitive formats. Through a proof-of-concept experiment, we study the performance of the proposed equalizer state caching scheme in a three-node optical switching system using 56 GBaud PAM-4. Our experimental results show that the proposed scheme can tolerate up to 0.8-nm (100-GHz) instantaneous wavelength change with an adaptation delay of only 0.36 ns. Practical considerations such as clock phase misalignment, temperature-induced wavelength drift, and equalizer precision are also studied.

Jitter-Based Authentication for Automotive Wireline Networks

“…The implementation complexity of the proposed source-ID driven equalization is low, with reconfiguration settling times akin to that of the Rapid ON/OFF links [37] or burst-mode DFEs [38]- [40] and equalizer state caching similar to [41]. The eye margin of the receiver assuming no equalization, single-stage CTLE only equalization or 3-tap DFE only equalization is shown in Fig.…”

Clock-Synchronized Clock and Data Recovery to Enable Sub-Nanosecond Optically-Switched Networks

“…>50 GBd) and higher order modulation formats (e.g. PAM4) [12]. One possible approach to mitigate this impact is to disaggregate the light sources from the switches and use tuneable lasers and wavelength demultiplex/multiplex for switching [13].…”

“…However, the thermal cross talk between tuneable lasers on the same photonic circuit may cause unwanted wavelength shifts that degrade system performance [14]. In addition, multiplexer/multiplexers may perform poorly in an unpredictable thermal environment, causing degraded performance or link failure due to the thermal dependent wavelength drift [12,13].…”