MiKANT: A Mirrored K-Ary N-Tree for Reducing Hardware Cost and Packet Latency of Fat-Tree and Clos Networks

Interconnection networks based on the fat-tree topology are widely used in high-performance parallel supercomputers. In a classical fat-tree, the radix of root switches is less than that of other switches. Fat-tree is a folded version of a Clos network. A Clos network uses the same radix switches in all stages. However, fat-tree or Clos network has a high switch cost and great packet latency. This paper proposes a variant of the fat-tree, named Mirrored k-ary n-tree (MiKANT), that doubles the number of compute nodes of the fat-tree by adding a few switches and making all the switches have a same radix. Compared to the classical fat-tree and Clos network, MiKANT not only reduces the numbers of switches and links so that it can be implemented at lower hardware cost, but also makes the network average distance shorter for achieving higher communication performance. We describe the structure of MiKANT, examine its topological properties, give a minimal per-hop deterministic routing algorithm, and evaluate the cost performance. Our results show that MiKANT reduces the average distance by about 0.5, saves 6.3% to 25.0% links and 12.5% to 50.0% switches, and improves performance of 9.1% to 41.4%, compared to the classical fat-tree, when n is in the range of 2 and 8. Our simulation results also show that MiKANT achieves much lower average packet latencies than the Clos network.

[1]  Charles Clos,et al.  A study of non-blocking switching networks , 1953 .

[2]  Wei Zhang,et al.  Floorplan Optimization of Fat-Tree-Based Networks-on-Chip for Chip Multiprocessors , 2014, IEEE Transactions on Computers.

[3]  John Kim,et al.  High Performance Datacenter Networks: Architectures, Algorithms, and Opportunities , 2011, High Performance Datacenter Networks: Architectures, Algorithms, and Opportunities.

[4]  Charles E. Leiserson,et al.  Fat-trees: Universal networks for hardware-efficient supercomputing , 1985, IEEE Transactions on Computers.

[5]  J. Duato,et al.  Butterfly vs . Unidirectional Fat-Trees for Networks-on-Chip : not a Mere Permutation of Outputs ∗ , 2008 .

[6]  Kwan Lawrence Yeung,et al.  An Efficient Routing Algorithm in Fat-Tree Data Center Networks , 2016, 2016 IEEE Global Communications Conference (GLOBECOM).

[7]  Pedro López,et al.  A Family of Fault-Tolerant Efficient Indirect Topologies , 2016, IEEE Transactions on Parallel and Distributed Systems.

[8]  Javier Navaridas,et al.  Reducing complexity in tree-like computer interconnection networks , 2010, Parallel Comput..

[9]  José Duato,et al.  RUFT: Simplifying the Fat-Tree Topology , 2008, 2008 14th IEEE International Conference on Parallel and Distributed Systems.

[10]  Amin Vahdat,et al.  Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google's Datacenter Network , 2015, Comput. Commun. Rev..

[11]  Fabrizio Petrini,et al.  k-ary n-trees: high performance networks for massively parallel architectures , 1997, Proceedings 11th International Parallel Processing Symposium.