Building Data Centers With Optically Connected Memory

Abstract

Abstract-Future data centers will require novel, scalable memory architectures capable of sustaining high bandwidths while still achieving low memory access latencies. Electronic interconnects cannot meet the challenges presented by the need for multi-terabit off-chip memory data paths. In this work, the electronic bus between main memory and its host processor is replaced with a circuit-switched optical interconnection network. We investigate the impact of our optically connected memory system on large-scale architectures and experimentally validate the protocol using field-programmable gate array based processor nodes and a custom-designed memory controller. The processor communicates all-optically with multiple synchronous dynamic random access memory nodes using 4×2.5-Gb/s wavelength-striped payloads, operating error free with bit-error rates less than 10 −12 .

Blocking probability in optical interconnects in data center networks

“…A high-end system requires high memory capacity and bandwidth while simultaneously demanding low latency and energy efficiency of interconnects. Making the interface between processors and memory devices optically connected significantly increases the amount of data available to the data center server processors, so that they operate at their maximum speeds [13].…”

A photonic interconnection network for hardware accelerator enabled utility computing

“…An ideal architecture for hardware accelerators would be radically data parallel and enjoy direct access to main memory [2]. An optical network is capable of this data parallelism and has previously been shown to interface well with memory [3]. In order for this architecture to succeed, the network must be able to actively switch and multicast.…”

“…Programming hardware accelerators is an additional hurdle, but with the development of languages like the Open Computing Language (OpenCL), which enables programs to be written in a way so that they will work across CPUs and GPUs, and IBM's Liquid Metal, a comprehensive compiler and run-time system that enables the use of a single language to program heterogeneous computing platforms, seamless co-execution of resultant programs on CPUs and accelerators is now possible [3,4].…”

“…However, in order for this system to be successful, current electronic networks are inadequate. Optical interconnects offer the bandwidth, latency, and energy specifications necessary to support delocalized accelerators [3]. We propose a novel system architecture with an optically-connected bank of hardware accelerators.…”

Resilient Optically Connected Memory Systems Using Dynamic Bit-Steering [Invited]

“…Optically connected memory (OCM) has been proposed as a solution to scaling memory systems within large-scale computers [10]. The high-bandwidth density of optical interconnects can alleviate pin-count constraints by allowing a single optical fiber to deliver terabits-per-second of memory bandwidth using wavelength-division multiplexing (WDM).…”