For the last 20 years, L2 tree-based network topologies meant that the only practical design methodology was to buy large, vertically scaled switch chassis for the core of the data centre. This limitation was largely due to the tree-structure forced on LAN networking by Spanning Tree Protocol. For every new device at point Access/1 we needed bigger boxes in the Core/2 or Distribution/3.
This is not a good use of resources. A L2-Tree Data Centre Network design is always focussed buying as much switching capacity as possible in the core switch to support unknown and unpredictable growth at the edge of the DC LAN. More edge ports need bigger distribution and core switches. It’s impractical to have many core switches in an STP design.
New Design Models
There are new design models for non-SDN networks based around L2 or L3 ECMP protocols like TRILL at L2 or IP loadbalancing at L3. In these methods, the upfront architecture is designed for 10 year plan but implemented in parts and steadily growing over time. This removes the requirements to overcapitalise on early purchases like switch chassis and have empty slots that may never be used.
The money spent on buying this capacity is a a viable investment when options don’t exist but it is bad business practice.
The diagram shows fought outline of a small L3/L2 ECMP network that uses 1 RU switches and can scale horizontally over time.
Capital and Cash Outflows
There are two parts to controlling cash outflows from the business. The most important is expenditure of capital becoming more difficult to justify and spending millions on data centre networks that are 10% used on day one is definitely no longer acceptable.
The second, and more useful, argument is that networking equipment is expected to get cheaper over the next five years. The cost of silicon and increase in the manufacturing base through the arrival of whitebox Ethernet and merchant silicon means that costs of switch hardware are set to fall dramatically. Predictions vary but 30-70% savings over 5 years is a practical guess. Therefore, it is much better to spend less capital and take advantage of cheaper products in the years ahead. I’m suggesting that a $1MM dollar hardware purchase today is likely to $500K in 3 to 5 years. (Note that you probably won’t save $500K, it will be spent on licenses for networking orchestration).
A third consideration is that vendor maintenance on big switch chassis starts on day one, even when you use just one port. Over 5 years, vendor maintenance will exceed the cost of the purchase.
An L2-Tree based network design does a brilliant job of transferring wealth from your company to a network vendor. The upfront capitalisation means you transfer your company profits to a supplier ahead of revenue or value extraction.
You Can Do Both
And don’t forget that you can keep your existing tree-based DC LAN and start growing an ECMP structure.
The EtherealMind View
You don’t need to replace the entire data centre network to start growing an ECMP network design that gives you a 10-year plan for your core network. There is no requirement to replace the existing L2 Tree network that connects your legacy compute infrastructure. Operate both networks until it makes sense to converge them.
For most companies it is time to stop overspending and overcapitalising on network products until there is a clear demand and justification for the investment. Change your architecture to scale up designs based on ECMP / CLOS to radically change the capex/opex spend in your LANs.