Dimension-Extended Dragonfly: A More Flexible Interconnection Network Aimed at Reducing Hardware Cost
Yaodong Wang, Yamin Li · 2023
Dragonfly topology is a commonly utilized design for interconnection networks in parallel and distributed systems. A classical dragonfly can be denoted as dragonfly(k, $m, l$), where m is the number of routers in a group, l is the number of links per router connected to other groups, and k is the number of links per router connected to compute nodes. Each router has other $m-1$ links fully connected to other $m-1$ routers within a group. Each group has $ml$ links connected to other groups. The groups are also fully connected, therefore there are $ml+1$ groups in total. The router radix in a dragonfly(k, $m, l$) is $l+k+m-1$. Building a large dragonfly system requires a large number of high-radix routers, increasing hardware costs. To reduce hardware costs, this paper proposes a more flexible topology called Dimension-Extended Dragonfly (DED). Instead of fully connected routers within a group, each router within a group is arranged in an n-dimensional matrix and routers in the same dimension are fully connected. We use n to denote the dimension such that each group in the DED has $m^{n}$ routers. We evaluate the cost, performance, and packet latency of DED. The results show that the DED network has a lower hardware cost for $n\geq 3$ compared to the classical dragonfly and cascade (an alternative to dragonfly). Additionally, DED is more flexible than classical dragonfly and cascade. It provides more options for choosing the system’s scale with different diameter and radix combinations. The simulation results also show that the packet latency of DED is shorter than dragonfly and cascade.