Which data centre network topology's best? Depends on what you want to break

Which data centre topology is better is an arcane, vexed and vital question: after all, as any cloud users knows while they're thumping the table/monitor/keyboard/whatever, we've gone long beyond a world where outages can be regarded as trivial.

Researchers from France and Brazil reckon one important question is where you expect failures – in switches, or in data links.

In this paper at Arxiv, the researchers led by Rodrigo de Souza Couto of the Universidade Federal do Rio de Janeiro compared how traditional three-layer, Fat-Tree, DCell and BCube architectures behave when things go wrong (compared to the traditional three-layer edge/aggregation/core model).

The paper explains that these three topologies were chosen because they have this in common: they're designed for the modern data centre design that combines modular infrastructure and low-cost equipment.

Their conclusion is that if a link fails, BCube recovers better, but if a whole switch goes dark, DCell is better.

For link failures, BCube maintains “at least 84 per cent of its servers when 40 per cent of the links are down, against 74 per cent in DCell … because, as a server-centric network, BCube employs redundant server interfaces.”

On the other hand, “DCell presents the best performance under switch failures, being able to achieve an MTTF up to 12 times higher than Bcube”, because it's servers rather than switches that control reachability.

DCell architecture: good if switches fail

The three-layer and Fat-Tree models didn't make the cut because they only provide a single path from server to edge: if either the link or the switch fails, the server becomes unreachable.

The BCube model is the most survivable, the authors write, but needs the most wiring and the most switches:

Lots of links keeps BCube running if one fails

The full paper is here. ®