Abstractions for Reconfigurable Hybrid Network Update and A Consistent Update Approach

Abstract

Reconfigurable Hybrid (electrical/optical) Network (RHN) [1-4, 6, 8, 10, 11, 13-19] for modern datacenter architectures has gained significant momentum during the last decade. The primary advantage of such RHN architectures is the dynamic topological reconfigurability enabled by optical circuit switches (OCS). On one hand, RHN can benefit throughput-intensive applications by providing on-demand high-bandwidth links between the hosts (CPU/GPU/TPU), such as distributed deep neural network training and recommendation systems, etc. On the other hand, RHN can reduce the hop-count between the host pairs, improving the performance for latency-sensitive applications such as real-time customer interactions with in-memory file system. However, previous works mostly focused on finding a suitable topology to efficiently handle a given traffic demand. Performing such topology update together with SDN policy update in a holistic manner while maintaining per-packet consistency and other network invariants is still an open issue. Existing network maintenance and policy update solutions define the notion of per-packet consistency assuming a pure SDN network where the physical network topology is static. This assumption does not hold for RHN because dynamic topology reconfiguration is inherent to RHN. In this paper, first, we define an extended notion of per-packet consistency and discuss the other critical requirements for RHN updates. Next, we provide an abstraction of RHN update and propose Transtate, a general method to perform such RHN update while satisfying the critical requirements. We believe such innovations remove one of the key obstacles towards reconfigurable-hybrid SDN.

Performance trade-offs in reconfigurable networks for HPC

“…Actuating a topology reconfiguration requires applying a sequence of atomic updates on the link and network layers, all of which can be coordinated using software-defined networking (SDN). The reason reconfigurations must be applied sequentially in incremental steps-as opposed to updating everything in one fell swoop-is to ensure routing consistency as the network carries live traffic [45]. It is important to maintain routing consistency in a network with dynamic topology to prevent unnecessary packet drops or routing "blackholes" due to route changes as a result of circuit switching.…”