Over this two-part blog, we reflect on some of the design decisions faced by data centre designers as fibre optic technology evolves to meet the challenges of a Big Data world.

In part one we look at how demand is driving a shift in data centre architecture and how optical transceiver technology developments have become a key enabler for hyper-scale data centres. In Part 2 we consider various aspects of fibre links and how they are changing: from the fibre cable itself, the optical transceiver modules, transmission speeds, associated standards, distances and how they are relevant to new build and legacy installations for all data centres.

The demand for cloud scale data centres shows little sign of slowdown and global data centre IP traffic is predicted to grow from 4.4 to 15.3 Zettabytes per year during the period 2015 to 20202, at 27% CAGR. In particular, trends towards Hyperscale data centres i.e. very large facilities with a single massively scalable compute architecture, are expected to grow to represent 53% of all DC traffic by 2020 from 34% in 2016. There are estimated to be around 300 of these hyper-scale facilities worldwide of which 45% of are located in the US with just 5% in the UK equating to 15 facilities.

Most larger data centres have been 3 tier switch fabric architectures up until recently but this is shifting to a flatter 2 tier architecture (e.g. Spine/Leaf) which typically reduces points of failure, power, latency and deployment time. It enables workloads to be shared across hundreds or even thousands of virtual machines. This creates substantially more ‘east-west’ server to server traffic and it is this that is accounting for a significant proportion of machine to machine (M2M) traffic  - 44% CAGR 2015-2020 (Cisco - Global Cloud Index 2016). In this architecture, each server port should be capable of connecting to any other server port at full bandwidth capacity i.e. non-blocking.  This is helping to drive data centre connections to move from 10G/40G to 25G/100G and the cost/ benefit in terms of $ per Gb improves with higher speeds. Due to the physical size and architecture of hyper-scale data centres, 100G Spine connections are desirable and started being deployed in 2016.

Optical Shift Several technologies are facilitating this data centre evolution but optical technology is perhaps the most crucial. Developments in optics for both short wave (SR) and long range (LR), has meant that the choices facing a data centre operator planning for future expansion are getting increasingly complex.  Previously the decision on whether to deploy lower cost, lower bandwidth SR optics and multimode fibre i.e. OM3 and subsequently OM4 was relatively straightforward with single-mode / LR optics being restricted to mainly core & aggregation switches.  With the flatter Spine/leave architectures in larger cloud-scale data centres, single mode is becoming more dominant with an eye to the next step in evolution which will be 100G/400G or 200G/400G where standards are being developed for a short reach(500m) over Singlemode e.g. IEEE P802.3cd (100GBASE-DR). That is to say 100Gbit/s transmission over one wavelength (2 fibres) and IEEE P802.3bs (200GBASE-DR4/400GBASE-DR4) using 4 pairs of single mode fibres. These will use PAM-4 encoding to modulate the signal and double the data in the same amount of time on a serial channel compared to NRZ (Non-return to zero) encoding.