The road to optical networking: IOWN & distributed infrastructures

Infrastructure reset

The infrastructure enabling the world’s computing is outdated. 

For years, much of the planet’s technological expansion (especially in AI training) has hinged on the use of hyperscale datacentres – vast facilities housing compute, storage and networking all concentrated in a single location. 

As AI workloads and the associated demand for compute power continue to grow, it is becoming infeasible in most geographies to continue to rely on the traditional approach of clustering monolithic datacentres physically close together.

Instead, the industry is poised to move toward architectures that link multiple geographically dispersed datacentres together as a single, unified virtual system, sharing compute and accelerator resources and synchronising storage across locations.

Networking for the modern age

For that to work, however, a new kind of network is required.

This new notion of a network is one that can tie these distributed datacentres together with ultra-high throughput and deterministic, ultra-low latency, thereby allowing workloads to perform as expected as they scale beyond the physical limits of a single data centre site. That’s the challenge being tackled by the Innovative Optical and Wireless Network (IOWN) initiative spearheaded by NTT and the more than 170 member companies in the IOWN Global Forum.

The network in question is the All-Photonics Network (APN), an architecture focused on open, end-to-end optical technologies rather than electronic packet switching.

Latency, jitter & packet loss

Standard optical/photonics networks, while effective for transmission, introduce latency, jitter, packet loss, and retransmissions, as well as increase power demands through repeated signal conversions.

This happens as optical signals are converted to electrical signals for routing, switching, grooming, regeneration, etc., and then back to light for transmission to the ‘next hop.’

The APN, by contrast, is designed to keep traffic in the optical domain end-to-end. By eliminating or at least dramatically minimising optical-electronic-optical conversions, it improves energy efficiency and maximises performance in terms of throughput, latency, jitter, packet loss, and retransmissions, which are the crucial factors in supporting large-scale distributed computing across long distances.

NTT’s Lawrence: The All Photonics Network is designed to keep traffic in the optical domain end-to-end, eliminating optical-electronic-optical conversions which lead to latency, jitter, packet loss and retransmissions.

One primary focus for the APN that’s already being tested by the IOWN Global Forum is the establishment of flexible, dedicated, ultra-high-capacity optical wavelength paths between facilities.

Notably, these optical resources must be programmable and able to be configured to allocate bandwidth according to demand. Optical pathways should not simply be a static transport method between distributed computing environments; instead, individual locations should be able to treat the network as an extension of the datacentre fabric.

Auto optical wavelength-path switching

For example, network disruption of any kind poses a severe threat to distributed infrastructure, necessitating the implementation of countermeasures to ensure network continuity in the event of an unforeseen adverse event. To address this, NTT and Japan’s National Institute of Informatics recently collaborated to demonstrate automated optical wavelength-path switching based on real-time network conditions. In the case of a natural disaster affecting multiple fiber routes, the system can, by coordinating the IP network controller alongside an APN optical controller, automatically determine the optimal wavelength and signal rates and then establish a new backup wavelength path in under ten minutes

It has become clear that the siloed design paradigm of current data centre infrastructure will not be able to keep pace with the explosive growth of the workloads being demanded of them.

Geographically distributed datacentres paired with maximum-performance, deterministic APN interconnections enable effective and efficient pooling of dispersed resources to keep up with the demands of emerging workloads.

Additionally, the next phases of IOWN will progressively see the superior qualities of optical data transmission leveraged inside the datacentre, inside the compute platforms, and ultimately even inside the silicon packages themselves.

Digital twins, digital twins

Complementary technologies, like digital twins, will also be essential for synchronising the tremendous amounts of data generated by many modern use cases for AI, such as real-world sensor data for smart cities or deterministic network performance for autonomous infrastructure.

As the industry continues to develop new and incredibly demanding workloads at breakneck speed, how infrastructure can support them holistically will not be determined by how many compute and storage resources can be installed in one location. Rather, the solution will lie in the network and data transmission technologies that tie together all the distributed compute and storage resources into massive, efficient, performant virtual systems.

Image credit: NTT Inc.