It is observed that the bottleneck in distributed training has shifted from computation to communication due to contention in concurrent transmissions and substantial redundant traffic. In the Parameter Server (PS) architecture, the server aggregates gradients from multiple workers and then distributes updated model parameters back to the workers in a one-to-many manner. Currently, model parameters are distributed via unicast, sending multiple identical copies of the data, which leads to significant bandwidth waste. Although multicast can save bandwidth, current approaches have two main drawbacks: on one hand, many protocols require maintaining excessive multicast state inside the network; on the other hand, the lack of coordination among multiple multicast trees can still lead to path conflicts. In this work, we propose path-aware multicast, which includes innetwork multicast tree reservation and per-hop control multicast. Specifically, before each round of model parameter distribution, the server queries the network for a multicast tree that satisfies the bandwidth requirement. The calculated multicast tree is then returned with bandwidth reserved at its tree nodes. Next, model parameters are forwarded with hop-by-hop control along the multicast tree. After the multicast is completed, the reserved network resources are released. Our evaluation shows that in an$8 \times 8$spine-leaf topology, path-aware multicast improves link load balancing by 32.6 % compared to random multicast and accelerates model parameter distribution by up to nearly$N \times$compared to unicast, where$N$is the number of workers.

Low Earth Orbit (LEO) satellite constellations are becoming a viable means for Internet access. However, their topology changes as satellites move towards or away from orbital intersection points, leading to constant link down or up. This may cause routing table entry updates and thus path changes between satellites. A path change during transmission may lead to out-of-order packet delivery and invalidate the current TCP congestion window. While some path changes are inevitable, some are avoidable. Dijkstra's algorithm is a popular choice among the routing protocols proposed for LEO satellite networks. We observe that many next-hop route updates by Dijkstra's algorithm are avoidable. Motivated by this, we propose StableR-oute, which stabilizes routing paths from different perspectives. StableRoute Local (SR_L) leverages equal-cost shortest paths and stays with the current one if it is still valid. StableRoute K-Short (SR_K) allows a path longer than the shortest path. StableRoute Global (SR_G) leverages the predictable satellite trajectories and topology variations, and thus works out a next-hop route selection sequence that minimizes the number of route updates over a time period. The evaluation shows that SR_L, SR_K and SR_G outperform Dijkstra's algorithm, substantially reducing the number of route updates in changing topologies.

Named Data Networking (NDN) enables efficient content distribution through in-network caching. However, the additional states of network intermediary nodes make NDN forwarding more burdensome, and the unpredictability of cache hits during forwarding leads to uncertain content retrieval latency. To overcome performance bottlenecks at the router's data plane and enhance network determinism, we propose the decoupled content router with remote content store (D-Router). This novel architecture decouples the local content store (CS) from routers and introduces the remote CS device for pooling important content. When Interest packets arrive at a router whose CS is overloaded, we ensure determinism by forwarding them to the remote CS for processing if the requested content is cached there, preventing blocking before the local CS of routers and potential random cache hits along the forwarding path. The dual-path bypass forwarding is supported through the design of routers and a dual-path routing protocol. D-Router is compatible with traditional NDN. Experiments show notable enhancements in data plane performance, including a 30% reduction in round-trip time (RTT), a 25% increase in throughput, improved determinism, and reduced network jitter. Additionally, the decoupling of CS makes it easier for network administrators to deploy network upgrades.

Due to the high costs of low Earth orbit (LEO) satellite manufacturing and launch, as well as the complexity of in-orbit network protocol debugging, simulating and verifying satellite network protocols on the ground before satellite launch holds significant importance. Compared to the expensive emulation with one-to-one replication, simulation (e.g., using ns-3) can achieve discrete event processing at a relatively lower cost by extending the wall clock time. However, very few studies have used ns-3 for LEO satellite network simulation, facing challenges such as faithfully simulating the on/off state switching of inter-satellite links (ISLs) and achieving simulation performance scalability for high-density satellite constellations. In this work, we propose a system to accelerate mega-scale LEO satellite network simulation in ns-3 via MPI-based parallelization. Specifically, we simulate ISLs based on ns-3's P2P channels/P2P remote channels and achieve runtime link connection/disconnection by implementing stateful traffic dropping inside the network interface. Then, we conduct concurrent simulation with ns-3's parallel and distributed simulation capability and partition the satellite constellation into multiple simulation processes through a hierarchical clustering algorithm and automated scripts, considering satellite locality and inter-process workload balance. Our evaluation shows significant speed improvements via parallelization, e.g., a 373% speedup with 12 processes for LEO-192, and a 156% speedup with 3 processes for LEO-3072.

For mega-scale low Earth orbit (LEO) satellite constellations, the relative position changes between satellites and ground terminals pose challenges for end-to-end TCP session maintenance, which serves as the substrate for many Internet services. Mobile IP resolves the session maintenance issue by in-troducing a binding mechanism between the care-of address and home address; however, this also leads to inefficient triangular routing. Our recently proposed LISP-LEO, through partition-satellite mapping, routes traffic to the service satellite above the destination terminal's partition, addressing the triangular routing problem. However, due to the corner case of partition-satellite mapping, LISP-LEO introduces the issue of the last-hop route selection, as well as the associated per-terminal registration states on the satellite, making the solution non-scalable. In this work, we propose Gaia, a ground station-centric mobility management scheme for LEO satellite networks. Gaia maintains a precise mapping of each ground terminal's IP and its geographical location. When receiving traffic from a source terminal, the access satellite can, based on the geographical location of the destination terminal carried by the traffic, directly locate the satellite above the destination terminal and tunnel the traffic to it. In addition, to reduce the satellite's burden, we add a DNS-like querying mechanism by offloading the mapping of terminal IPs and geographical locations to the ground station. The evaluation shows that Gaia outperforms Mobile IP and LISP-LEO in both end-to-end latency and on-board resource consumption.

With the low latency and high coverage, the low earth orbit (LEO) satellite systems are attracting more and more venture capitals as well as research attentions. Due to their highly dynamic constellation topologies, routing protocols on the ground have to be tailored to efficiently adapt to the regular topology changes. However, for irregular topology changes caused by exceptional link failure/recovery, network-wide route flooding is still necessary for route convergence. But, this will cause significant traffic flooding redundancy due to the high density of constellation topologies. For larger-scale constellations, the redundancy issue will be exacerbated. To lessen the redundancy, this work proposes a lightweight route flooding mechanism by generating a sparse flooding topology that prunes the original full-mesh topology, and only flooding the route information on the sparse topology. By considering the maximum flooding hop as well as the robustness of the flooding topology, we design an algorithm to calculate the optimal topology instead of just applying the minimum spanning tree. The evaluation shows that, for the LEO-96 constellation, the new flooding topology has a 28.1% reduction in the inter-satellite links (ISLs) compared with the original topology, and the new flooding mechanism has a 37.52% reduction in the traffic flooded and a 10.03% reduction in the route convergence time compared with OSPF. Such improvements will be amplified on larger-scale constellations.