Architecture and Performance Studies of 3D-Hyper-FleX-LION for Reconfigurable All-to-All HPC Networks

Liu, Gengchen; Proietti, Roberto; Fariborz, Marjan; Fotouhi, Pouya; Xiao, Xian; Yoo, S. J. B.

doi:10.1109/sc41405.2020.00030

Cited by 11 publications

(9 citation statements)

References 56 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In summary, although the literature on reconfigurable networks has grown considerably over the past two decades [14][15][16]32,35,38,[41][42][43][44][49][50][51][52][53][54][55], there is very little understanding of the trade-offs of different reconfigurable network architectures. This is due to (1) prior proposals having all been point designs and (2) the lack of commercial adoption, with no recorded system having deployed a reconfigurable network.…”

Section: B Dynamic Reconfigurable Networkmentioning

confidence: 99%

“…At present, the literature focusing on transport protocol design for reconfigurable networks remains underdeveloped. Past work has relied on standard TCP/IB protocols [16,35,49], modified TCP stacks [58], or novel transport protocols specific to the underlying network architecture [14,40].…”

Section: Flow Controlmentioning

confidence: 99%

“…To address these challenges, researchers have made large strides in improving networks with better routing schemes [5][6][7][8][9], task placement [10], and topology designs [11][12][13]. The integration of optical circuit switching in the context of high performance systems has been one such effort, and is an intense area of recent research [14][15][16]. In addition to bringing significant benefits to bandwidth density and power, optical circuit switching enables topological reconfiguration, which is a key enabler for traffic-aware topology designs at runtime.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Performance trade-offs in reconfigurable networks for HPC

Teh

Glick

et al. 2022

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

Designing efficient interconnects to support high-bandwidth and low-latency communication is critical toward realizing high performance computing (HPC) and data center (DC) systems in the exascale era. At extreme computing scales, providing the requisite bandwidth through overprovisioning becomes impractical. These challenges have motivated studies exploring reconfigurable network architectures that can adapt to traffic patterns at runtime using optical circuit switching. Despite the plethora of proposed architectures, surprisingly little is known about the relative performances and trade-offs among different reconfigurable network designs. We aim to bridge this gap by tackling two key issues in reconfigurable network design. First, we study how cost, power consumption, network performance, and scalability vary based on optical circuit switch (OCS) placement in the physical topology. Specifically, we consider two classes of reconfigurable architectures: one that places OCSs between top-of-rack (ToR) switches—ToR-reconfigurable networks (TRNs)—and one that places OCSs between pods of racks—pod-reconfigurable networks (PRNs). Second, we tackle the effects of reconfiguration frequency on network performance. Our results, based on network simulations driven by real HPC and DC workloads, show that while TRNs are optimized for low fan-out communication patterns, they are less suited for carrying high fan-out workloads. PRNs exhibit better overall trade-off, capable of performing comparably to a fully non-blocking fat tree for low fan-out workloads, and significantly outperform TRNs for high fan-out communication patterns.

show abstract

Section: B Dynamic Reconfigurable Networkmentioning

confidence: 99%

Section: Flow Controlmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Performance trade-offs in reconfigurable networks for HPC

Teh

Glick

et al. 2022

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

show abstract

“…workloads, etc.). Thanks to recent advances in scalable and manufacturable silicon-photonic technologies [12], flat and disruptive optical interconnect architectures scaling up to tens of thousands of terminals by enabling direct (thus, low-diameter) wavelength switching between racks [3,12,15] have been recognized as promising solutions for meeting the above challenges.…”

mentioning

confidence: 99%

“…Each column (or row) is a cluster (often referred to as portable data centers, PODs) with ToR switches per POD. The intra-POD connectivity in each dimension is augmented with wavelength-space selective optical switches [12] that allow reconfiguring the topology and bandwidth (number of links between ToR pairs) based on certain algorithms and policies (which will be discussed in Section 3). Fig.…”

mentioning

confidence: 99%

Reconfigurable Optical Datacom Networks by Self-supervised Learning

Liu

Chen²,

Proietti

et al. 2021

Proceedings of the ACM SIGCOMM 2021 Workshop on Optical Systems

Self Cite

View full text Add to dashboard Cite

This paper presents a self-supervised machine learning approach for cognitive reconfiguration in a Hyper-X-like flexible-bandwidth optical interconnect architecture. The proposed approach makes use of a clustering algorithm to learn the traffic patterns from historical traces. A heuristic algorithm is developed for optimizing the connectivity graph for each identified traffic pattern. Further, to mitigate the scalability issue induced by frequent clustering operations, we parameterize the learned traffic patterns by a deep neural network classifier. The classifier is trained offline by supervised learning to enable classification of traffic matrices during online operations, thereby facilitating cognitive reconfiguration decision making. Simulation results show that compared with a static all-toall interconnection, the proposed approach can improve throughput by up to 1.76× while reducing end-to-end packet latency and flow completion time by up to 2.8× and 25×, respectively.

show abstract