Congestion control is required to achieve more stringent equilibrium and transient properties with Remote Direct Memory Access (RDMA) deployed in modern disaggregated data center networks. However, existing congestion control algorithms still suffer from throughput loss or queueing buildup during congestion events. With congestion signals getting more fine-grained, outdated information lagged by the feedback delay could lead to overreaction or underreaction. As another factor, rate or congestion window is insufficient for simultaneously achieving both throughput and latency properties. To overcome these intrinsic limitations, we propose the methodologies of the ideal congestion response, adjusting both the sending time and sending rate. The sending time should be set so that subsequent packets arrive at the queue right after the packets accumulated from the congestion event have been drained. Low latency is achieved by pausing before the sending time, which provides the fastest way of draining, and feedback delay should be considered when setting the sending time. For high throughput after draining the queue, the sending rate is set to match in-flight packets with the network capacity. Guided by the methodologies, we design RICE, a delay-based congestion control algorithm with a novel congestion response mechanism, which provides significant performance gain without support from switches or receivers. Evaluation results show that RICE simultaneously achieves high throughput, low latency and fairness across diverse scenarios. Compared with prior work, the 99.9th-percentile FCT slowdown is up to 60% lower for short flows and 34% lower for long flows.
Decentralized learning enhances privacy, scalability, and fault tolerance by distributing data and computation across nodes, making it a promising alternative to centralized approaches. Federated Learning, a widely adopted method, relies on a central aggregator but suffers from vulnerabilities such as single points of failure, limited scalability, and privacy risks. Gossip Learning and Epidemic Learning offer fully decentralized solutions through peer-to-peer exchanges of model updates, ensuring robustness and privacy but at the cost of slower convergence. In this work, we present HEAL, an innovative decentralized learning framework based on a self-organizing and self-healing Peer-to-Peer overlay. HEAL leverages emergent multi-hub networks, combining the randomness of graph structures with the efficiency of star networks. This architecture enables the organic formation of prominent hubs that enhance fault tolerance, reduce network diameter, and facilitate robust decentralized aggregation. HEAL is the first framework that integrates the strengths of Federated Learning, Gossip Learning, and Epidemic Learning, utilizing the Elevator algorithm to dynamically select aggregators. Our simulations demonstrate that HEAL outperforms Gossip Learning and Epidemic learning, and achieves performance comparable to Federated Learning in stable environments. These performances are maintained in scenarios with high node churn or network instability. This framework offers a scalable, privacy-preserving, and fault-tolerant solution for decentralized learning applications.
Distributed Quantum Computing (DQC) enables the execution of quantum circuits across multiple interconnected quantum processing units (QPUs) but requiring efficient qubit allocation and network topology design to optimize computational performance. Proper qubit allocation minimizes entanglement costs across QPUs, balances computational workload, and ensures efficient execution of quantum computing tasks. Meanwhile, network topology plays a crucial role in reducing entanglement routing complexity and communication overhead for remote quantum gate operations. Therefore, in this paper, we propose a joint optimization framework for network topology design and qubit allocation in DQC to minimize communication overhead. We formulate the problem as a tractable integer nonlinear programming model that explicitly incorporates entanglement routing, thereby ensuring a more tractable optimization process. To improve computational efficiency, we present a partially linearized version of the problem, making it solvable using any classical optimization solver. To further address the scheduling of large-scale quantum circuits, we develop heuristics from edge contraction in the qubit interaction graph to pre-assign groups of qubits to QPUs, ensuring that our method remains effective at scale. Extensive simulations on real-world quantum circuits over various quantum clusters or distributed quantum networks validate the effectiveness of our proposed approach, demonstrating its capability to handle complex quantum circuits while reducing communication costs in DQC.
We study an opinion optimization problem based on an extended DeGroot model for a social network, where stubborn leaders hold fixed opinions of 0 or 1. Given a set of leaders with opinion 0, our goal is to select a small number of leaders with opinion 1 from the remaining nodes to maximize the average opinion of the network. Previous studies mainly focus on undirected graphs, but real-world social networks are often directed. Therefore, it is essential to address the problem in directed graphs. To solve the problem, we first establish a novel interpretation of equilibrium opinions using random rooted spanning forests. Building on this interpretation, we propose a random spanning forest sampling greedy algorithm that leverages an extension of Wilson’s algorithm and a sample update technique, which is efficient and provides an error guarantee. Our algorithm applies to both undirected and directed graphs. Extensive experiments on real-world networks demonstrate that for undirected graphs our algorithm outperforms the state-of-the-art method—achieving accuracy improvements of up to 42% and speedups of up to 80%. Moreover, for directed graphs our algorithm achieves comparable effectiveness to the deterministic greedy algorithm and is scalable to large directed graphs with millions of nodes, where the deterministic greedy algorithm is not scalable.
We investigate the sym/asyn/homo/global multichannel rendezvous problem (MRP), in which two identical users equipped with multiple radios must discover a common channel among N licensed channels under arbitrary clock offsets. Let k denote the number of radios per user. Building on relaxed difference sets (RDS), we introduce a multiple-radio emulation viewpoint that unifies DS-ACH and role-based constructions. Within this framework, we propose two systematic designs: a hierarchical DS-ACH construction and a direct composition method based on coprime periods. The hierarchical construction recursively emulates virtual radios and guarantees rendezvous under arbitrary clock offsets with explicit bounds on period and maximum time-to-rendezvous (MTTR) for general N and k. When coprimality conditions are satisfied, the direct composition achieves strictly improved MTTR while preserving the maximum rendezvous diversity (MRD) and even channel use (ECU) properties. By integrating recursive hierarchy with channel partitioning and starting-time staggering, we obtain order-optimal (N,k,L) MACH sequences with period L=(N2/k2) and MTTR=(N/k2) under mild conditions. Numerical verification confirms the tightness of the analytical bounds and demonstrates substantial delay reduction over existing multi-radio schemes.
