Jul 2026· International Conference on Optical Communications and Networks· pp. 1-3· 0 citations· 16 references
Abstract
Collective communications (CC) and distributed artificial intelligence (AI) training workloads involve substantial one-to-many chunked data transfers. Enabling optical multicast in next-generation optical circuit switching (OCS)-based data centers could streamline such traffic and improve CC efficiency. This paper presents our investigation of when and how to exploit optical multicast for CC and AI training. We discuss the topological property of OCS with multicast capability, showing reduced port-count overhead for the same connectivity degree as its point-to-point (P2P) counterpart. Further, by formulating and addressing the grouped-multicast CC schedule optimization (GMS), we study and compare the completion time of two key types of CC operations in GMS against the P2P schedule under identical port use conditions. Simulations with representative topologies reveal the potential of multicast OCS to enhance CC via a co-design of multicast-based topology and scheduling.
Multicast transmission is a fundamental mechanism used to deliver content to multiple receivers/users and is still an essential method in television broadcasting services and data distribution networks. Efficient management of multicast traffic has a direct impact on bandwidth consumption and helps to prevent network c...
A. Shehu, Fatmir Basholli, Lea Likaj et al.· international journal of eng...· 0 citations
Flux is presented, a scheduler that optimally schedules optical circuit switches based on the structure of the entire workload and reduces training iteration time and peak NIC buffer requirements by more than three orders of magnitude compared to traditional periodic schedulers.
Arno Troch, Seyyidahmed Lahmer, A. Nada et al.· 0 citations
The rapid growth of large-scale AI models has driven the emergence of AI data centers (AIDCs), where distributed training produces massive communication demands under multi-job concurrent execution. All-optical data center networks have emerged as a promising solution due to their high bandwidth and low latency. Howeve...
Huitao Zhou, Haoyang Chen, Bojun Zhang et al.· Journal of Optical Communica...· 0 citations
This paper presents the first approximation algorithm with an \(\mathcal {O}(\tau _D)\) performance guarantee for single-coflow scheduling in an N × N hybrid-switched DCN under the not-all-stop setting, and extends to the multiple-coflow setting and proposes a heuristic algorithm that converts specific pure-EPS schedul...
Xin Wang, Hong Shen, Hui Tian· Proceedings of the Internati...· 0 citations
ReTri, a bidirectional All-to-All schedule for ORNs based on the Trivance algorithm is presented, a bidirectional All-to-All schedule for ORNs based on the Trivance algorithm that improves completion time and improves reconfigurable Bruck by up to 2.1×.
Anton Juerss, Stefan Schmid· Conference on Applications,...· 0 citations
QPS-ToR is proposed, which replaces NegotiaToR's scheduling logic with SW-QPS, a sliding-window algorithm originally proposed for crossbar scheduling that achieves around 90% throughput with a single low-complexity iteration.
Dong-Zhao Song, Qian-Ru Yu, Jun Xu· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
