HyperShare-native implementations - i.e. torus and torus-like network topologies - operate in non-coherent mode, which is not penalized by the software complexity and performance scalability issues plaguing coherent cluster network architectures. The non coherent approach, coupled with a full 64-bit node addressing capability give HyperShare-based networks a practically unrestricted cluster node scalability.
Although in HyperShare-over-Ethernet (HSoE) and HyperShare-over-InfiniBand (HSoE) cluster node scalability is restricted by the physical scaling limits of the applicable Ethernet and InfiniBand network switching hardware, Native HyperShare implementations do not suffer such limitations thanks to its:
Non-Coherent Global Shared Memory Architecture
Does not suffer the scaling constraints of memory-coherent arcjhitectures.
64-bit Cluster Node Addressing Capability
Enables cluster nodes addressability to scaling to levels that are only limited by geography,
physical, cost and power consumption factors.
Torus Topology Network Architecture
Unlike Fat Tree network architectures, Torus topologies networks can scale linearly, granularly,
with a high degree of modularity to any cluster size. Also, incremental network expansions
do not require the whole cluster to be powered-down and rewired, but remain localized within
just a portion of the cluster network's periphery
The two slides below illustrate with technical and business considerations the scalability strengths of Torus networks.
(Click on Arrows to Scroll Through or on Image to Open in Full Size)