Learning causal structures from temporal data is fundamental to many practical tasks, such as physical laws discovery and root causes localization. Real-world systems often exhibit long-term carry-over effects, where the value of a variable at the current time can be influenced by distant past values of other variables. These effects, due to their large temporal span, are challenging to observe or model. Existing methods typically consider finite lag orders, which may lead to confounding from early historical data. Moreover, incorporating historical information often results in computational scalability issues. In this paper, we establish a theoretical framework for causal discovery in complex temporal scenarios where observational data exhibit long-term carry-over effect, and propose LEVER, a theoretically guaranteed novel causal discovery method for incomplete temporal data. Specifically, based on the \textit{Limited-history Causal Identifiability Theorem}, we refine the variable values at each time step with data at a few preceding steps to mitigate long-term historical influences. Furthermore, we establish a theoretical connection between QR decomposition and causal discovery, and design an efficient reinforcement learning process to determine the optimal variable ordering. Finally, we recover the causal structure from the R matrix. We evaluate LEVER on both synthetic and real-world datasets. In static cases, LEVER reduces SHD by 17.29\%-40.00\% and improves the F1-score by 5.30\%-8.79\% compared to the best baseline. In temporal cases, it achieves a 64\% reduction in SHD and a 45\% improvement in F1-score. Additionally, LEVER demonstrates significantly higher precision on real-world data compared to baseline methods.

Modern cloud service providers and AI innovators are expanding their services across data centers to harness scale-out benefits while minimizing wide-area network (WAN) communication overhead. Although Remote Direct Memory Access (RDMA) has become the de-facto standard for high-speed data center networks, extending its benefits to WAN faces fundamental challenges: high bandwidth-delay products (BDP), frequent packet loss, and the inherent tension between performance and resource efficiency in existing approaches. To address these fundamental limitations and empower cross-datacenter services, we propose OmniDMA, a novel PFC-free RDMA architecture designed for loss-prone WAN environments. Our design introduces two core innovations: 1) The Adamap data structure enables flexible loss recording through context compression and management, and 2) a three-tier control path architecture that decouples retransmission from primary transmission control path, achieving high performance at WAN scale with minimal RNIC SRAM consumption. This design further reduces context management and scheduling overhead to guarantee line-rate processing under packet loss. Preliminary results demonstrate that OmniDMA achieves efficient RDMA communication over lossy WANs with constant resource consumption.

The local deployment of large language models (LLMs) on mobile devices has garnered increasing attention due to its advantages in enhancing user privacy and enabling offline operation. However, given the limited computational resources of a single mobile device, only small language models (SLMs) with restricted capabilities can currently be supported. In this paper, we explore the potential of leveraging the collective computing power of multiple mobile devices to collaboratively support more efficient local LLM inference. We evaluate the feasibility and efficiency of existing parallelism techniques under the constraints of mobile devices and wireless network, identifying that chunked pipeline parallelism holds promise for realizing this vision. Building on this insight, we propose FlexSpark, a novel solution designed to achieve efficient and robust multi-device collaborative inference. FlexSpark incorporates priority scheduling, ordered communication, and elastic compression to maximize wireless bandwidth utilization, and thus accelerates distributed inference. Preliminary experimental results demonstrate that FlexSpark achieves up to a 2x speedup compared to state-of-the-art frameworks, significantly enhancing the practicality and scalability of LLM deployment on mobile devices.

Cross-silo privacy-preserving machine learning (PPML) adopts Partial Homomorphic Encryption (PHE) for secure data combination and high-quality model training across multiple organizations (e.g., medical and financial). However, PHE introduces significant computation and communication overheads due to data inflation. Batch optimization is an encouraging direction to mitigate the problem by compressing multiple data into a single ciphertext. While promising, it is impractical for a large number of cross-silo PPML applications due to the limited vector operations support and severe data corruption. In this paper, we present GeniBatch, a batch compiler that translates a PPML program with PHE into an efficient program with batch optimization. GeniBatch adopts a set of conversion rules to allow PHE programs involving all vector operations required in cross-silo PPML and ensures end-to-end result consistency before/after compiling. By proposing bit-reserving algorithms, GeniBatch avoids bit-overflow for the correctness of compiled programs and maximizes the compression ratio. We have integrated GeniBatch into FATE, a representative cross-silo PPML framework, and provided SIMD APIs to harness hardware acceleration. Experiments across six popular applications show that GeniBatch achieves up to 22.6x speedup and reduces network traffic by 5.4x-23.8x for generic cross-silo PPML applications.

Adaptive neural networks (NN) have been used to optimize OS kernel datapath functions because they can achieve superior performance under changing environments. However, how to deploy these NNs remains a challenge. One approach is to deploy these adaptive NNs in the userspace. However, such userspace deployments suffer from either high cross-space communication overhead or low responsiveness, significantly compromising the function performance. On the other hand, pure kernel-space deployments also incur a large performance degradation because the computation logic of model tuning algorithm is typically complex, interfering with the performance of normal datapath execution. This paper presents LiteFlow, a hybrid solution to build high-performance adaptive NNs for kernel datapath. At its core, LiteFlow decouples the control path of adaptive NNs into: (1) a kernel-space fast path for efficient model inference, and (2) a userspace slow path for effective model tuning. We have implemented LiteFlow with Linux kernel datapath and evaluated it with three popular datapath functions including congestion control, flow scheduling, and load balancing. Compared to prior works, LiteFlow achieves 44.4% better goodput for congestion control, and improves the completion time for long flows by 33.7% and 56.7% for flow scheduling and load balancing, respectively.

WiFi networks, widely utilized by wireless devices, have become increasingly complex and congested environments, leading to noticeable delays, jitter, and throughput degradation for end-to-end network flows in today's Internet. Through detailed experimental observations, we identified the TCP victim problem in WiFi, where TCP erroneously detects congestion and interacts with WiFi routers, resulting in a significant throughput decrease for certain hosts. In this paper, we introduce Cupid, a novel congestion control algorithm that relies on receiver-side WiFi physical layer measurements. Cupid accurately assesses congestion by measuring parameters such as airtime utilization, concurrency, and rates, allowing for precise rate adjustments. Our results demonstrate that whether employed alone or alongside other congestion controls, Cupid effectively mitigates the TCP victim problem. Furthermore, when used independently, Cupid can also reduce latency.

Network performance is critical to the user experience of many real-time interactive applications, such as video conferencing and live streaming. Empirical studies [46] show that transport latency over 300ms would become unacceptable, leading to significant user satisfaction declining. Unfortunately, due to the best-effort nature of Internet, such strict performance requirement can hardly be fully met. Despite continuous efforts have been made to improve the performance of Internet (e.g., overlay routing optimization and content delivery network), we are still far from delivering satisfying network performance for these applications. The stringent network requirements, the world-wide cross-continental network transfers, and the large-scale Internet-wide users, together make it a complex challenge to deliver ideal user experience for emerging real-time applications. In this paper, we present Panorama, a scalable system for delivering desired user experience to real-time applications over a globally distributed overlay network. Specifically, Panorama takes a centralized approach with end-to-end multi objective traffic engineering optimization to meet tight performance requirement for real-time applications, and in the meanwhile achieving good scalability with a batch of optimizations including intelligent measurement-based user/session grouping and parallelizable path calculation. We evaluate Panorama based on 81 million selected real-world traces in deployment environment with clients across 66 countries. The extensive evaluation demonstrates that Panorama can support a routing service for millions of users, while providing latency lower than 200ms for 96.34% of the communication sessions, and improving SLA satisfaction by up to 88.0%.

Data-driven solutions hold significant promise for improving network protocols and services in wide area networks. However, their practical adoption in production networks has been limited. This paper investigates the potential of leveraging network data itself to enhance the effectiveness of data-driven solutions. We evaluate a Quality of Service (QoS) forecasting model trained on directly collected network data to demonstrate the advantages of harnessing networking data for machine learning purposes. Our results reveal that training the model with network data effectively addresses the challenges of Data Drift. We also acknowledge the limitations of designing a generic framework to support all problem domains. To overcome this challenge, we propose a comprehensive set of potential solutions that leverage network data for machine learning (ML) applications.

TCP was designed to provide stream-oriented communication service for bulk data transfer applications (e.g., FTP and Email). With four-decade development, Internet applications have undergone significant changes, which now involve highly dynamic traffic pattern and message-oriented communication paradigm. However, the impact of this substantial evolution on congestion control (CC) has not been fully studied. Most of the network transports today still make the long-held assumption about application traffic, i.e., a byte stream with an unlimited data arrival rate. In this paper, we demonstrate, through both analyses and experiments, that the emerging traffic dynamics and message-level data structure have huge impacts on the correctness and effectiveness of CC, but none of the existing solutions treats these two characteristics appropriately. Therefore, we present Amphis, a new CC framework that re-architects the current pure network-oriented design into a dual-control architecture, which combines application-coordinated control and network-oriented control. Amphis contains two novel ideas, i.e., pattern-driven proactive probing for handling traffic dynamics and message-driven adaptive optimization for optimizing message transmission performance. Our preliminary results show that Amphis holds great promise in terms of accurate bandwidth estimation under dynamic traffic conditions and effective data transfer at message granularity.

Adaptive neural networks (NN) have been used to optimize OS kernel datapath functions because they can achieve superior performance under changing environments. However, how to deploy these NNs remains a challenge. One approach is to deploy these adaptive NNs in the userspace. However, such userspace deployments suffer from either high cross-space communication overhead or low responsiveness, significantly compromising the function performance. On the other hand, pure kernel-space deployments also incur a large performance degradation because the computation logic of model tuning algorithm is typically complex, interfering with the performance of normal datapath execution.

In recent years, enterprises are increasingly moving from private WANs to hybrid WANs, for the purpose of cost saving, better scalability and improved user experience. To utilize network resource of both private and public networks, existing traffic engineering (TE) solutions for hybrid WANs perform traffic classification at the application level, i.e., sending traffic of high-priority applications to private WANs while letting other traffic go to the Internet. Unfortunately, such a strategy falls short of achieving optimal performance and has inherent limitation in making desired tradeoff between cost and performance. In this paper, we present comboTE, a mixed-link based TE framework to optimize both performance and cost over hybrid WAN. Specifically, comboTE designs a fine-grained TE strategy that considers link(s) from both types of networks at every hop when deciding the routing path for each traffic demand. To address the scalability issues, we leverage Lagrangian relaxation, a decomposition technique for solving large scale integer linear programs, combined with a novel augmented-graph-based approach to derive the most cost-efficient link compositions within a segment. Our comboTE achieves the near-optimal solution with a theoretically proven gap that is better than linear programming relaxation. The experimental results on realistic topologies and traffic matrices show that comboTE scales well and achieves solutions of high quality. Compared to prior TE algorithms, comboTE achieves a solution with tighter gap in limited time, and takes 2.2× less execution time without a time budget.

As datacenter network bandwidth keeps growing, proactive transport becomes attractive, where bandwidth is proactively allocated as "credits" to senders who then can send "scheduled packets" at a right rate to ensure high link utilization, low latency, and zero packet loss. While promising, a fundamental challenge is that proactive transport requires at least one-RTT for credits to be computed and delivered. In this paper, we show such one-RTT "pre-credit" phase could carry a substantial amount of flows at high link-speeds, but none of existing proactive solutions treats it appropriately. We present Aeolus, a solution focusing on "pre-credit" packet transmission as a building block for proactive transports. Aeolus contains unconventional design principles such as scheduled-packet-first (SPF) that de-prioritizes the first-RTT packets, instead of prioritizing them as prior work. It further exploits the preserved, deterministic nature of proactive transport as a means to recover lost first-RTT packets efficiently. We have integrated Aeolus into ExpressPass[14], NDP[18] and Homa[29], and shown, through both implementation and simulations, that the Aeolus-enhanced solutions deliver signiicant performance or deployability advantages. For example, it improves the average FCT of ExpressPass by 56%, cuts the tail FCT of Homa by 20x, while achieving similar performance as NDP without switch modifications.

Today's cloud is shared among multiple tenants running different applications, and a desirable multi-tenant datacenter network infrastructure should provide bandwidth guarantees for throughput-intensive applications, low latency for latency-sensitive short messages, as well as work conservation to fully utilize the network bandwidth. Despite significant efforts in recent years, none of them can achieve these three properties simultaneously. In this paper, we identify the key deficiency of prior solutions and use this insight to motivate our design of Trinity-a simple, practical yet effective solution that achieves bandwidth guarantees, work conservation and low latency simultaneously in the cloud. We implement Trinity using existing commodity hardwares and demonstrate its superior performance over prior solutions using testbed experiments.

In recent years, federated learning (FL) has been widely applied for supporting decentralized collaborative learning scenarios. Among existing FL models, federated logistic regression (FLR) is a widely used statistic model and has been used in various industries. To ensure data security and user privacy, FLR leverages homomorphic encryption (HE) to protect the exchanged data among different collaborative parties. However, HE introduces significant computational overhead (i.e., the cost of data encryption/decryption and calculation over encrypted data), which eventually becomes the performance bottleneck of the whole system. In this paper, we propose HAFLO, a GPU-based solution to improve the performance of FLR. The core idea of HAFLO is to summarize a set of performance-critical homomorphic operators (HO) used by FLR and accelerate the execution of these operators through a joint optimization of storage, IO, and computation. The preliminary results show that our acceleration on FATE, a popular FL framework, achieves a 49.9$\times$ speedup for heterogeneous LR and 88.4$\times$ for homogeneous LR.

The link speed in production datacenters is growing fast, from 1Gbps to 40Gbps or even 100Gbps. However, the buffer size of commodity switches increases slowly, e.g., from 4MB at 1Gbps to 16MB at 100Gbps, thus significantly outpaced by the link speed. In such extremely shallow-buffered networks, today's TCP/ECN solutions, such as DCTCP, suffer from either excessive packet loss or substantial throughput degradation. To this end, we present BCC1, a simple yet effective solution that requires just one more ECN config (i.e., shared buffer ECN/RED) over prior solutions. BCC operates based on real-time global shared buffer utilization. When available buffer space suffices, BCC delivers both high throughput and low packet loss rate as prior work; Once it gets insufficient, BCC automatically triggers the shared buffer ECN to prevent packet loss at the cost of sacrificing little throughput. BCC is readily deployable with existing commodity switches. We validate BCC's hardware feasibility in a small 100G testbed and evaluate its performance using large-scale simulations. Our results show that BCC maintains low packet loss rate while slightly degrading throughput when the available buffer becomes insufficient. For example, compared to current practice, BCC achieves up to 94.4% lower 99th percentile flow completion time (FCT) for small flows while degrading average FCT for large flows by up to 3%.

近10年来,在盛行的网络应用(如搜索、在线零售和云计算等)的需求驱动下,数据中心在全球范围内以前所未有的速度和规模发展建立起来.特别地,数据中心网络引起了学术界和工业界的广泛关注.在这样的背景下,调研了数据中心网络的一个核心方面——传输层协议.虽然传输协议在因特网上已经有很长的历史,它却直到2010 年才在数据中心网络环境下被系统性地探索.数据中心网络有着和因特网不一样的特点(如单一控制域和同构网络架构),这给数据中心网络上的传输协议设计同时带来了机遇和挑战.在这驱使下,一系列的传输协议被设计提出.将早期(2010-2015 年)数据中心网络传输设计方面的工作分成3 类——基于端主机的拥塞控制、网络仲裁机制和交换机优先级调度,对这3 类工作的优缺点作深入讨论.最后,分析近年来数据中心网络传输设计的研究趋势——接收端驱动的主动拥塞控制和RDMA传输协议设计.

With the increasing awareness of privacy protection and data fragmentation problem, federated learning has been emerging as a new paradigm of machine learning. Federated learning tends to utilize various privacy preserving mechanisms to protect the transferred intermediate data, among which homomorphic encryption strikes a balance between security and ease of utilization. However, the complicated operations and large operands impose significant overhead on federated learning. Maintaining accuracy and security more efficiently has been a key problem of federated learning. In this work, we investigate a hardware solution, and design an FPGA-based homomorphic encryption framework, aiming to accelerate the training phase in federated learning. The root complexity lies in searching for a compact architecture for the core operation of homomorphic encryption, to suit the requirement of federated learning about high encryption throughput and flexibility of configuration. Our framework implements the representative Paillier homomorphic cryptosystem with high level synthesis for flexibility and portability, with careful optimization on the modular multiplication operation in terms of processing clock cycle, resource usage and clock frequency. Our accelerator achieves a near-optimal execution clock cycle, with a better DSP-efficiency than existing designs, and reduces the encryption time by up to 71% during training process of various federated learning models.

Parameter/gradient exchange plays an important role in large-scale distributed machine learning (DML). However, prior solutions such as parameter server (PS) or ring-allreduce (Ring) fall short since they are not resilient to issues or uncertainties like oversubscription, congestion or failures that may occur in datacenter networks (DCN). This paper proposes RAT, a new solution that determines the communication pattern for DML. At its heart, RAT establishes allreduce trees taking into account the physical topology and its oversubscription condition. The allreduce trees specify the aggregation pattern in which each aggregator is responsible for aggregating gradients from all workers within an oversubscribed region at the reduce phase, and broadcasting the updates back to workers at the broadcast phase. We show that such an approach can effectively reduce cross-region traffic and shorten dependency chain compared to prior solutions. We have evaluated RAT in both oversubscribed network and network with failures and found that RAT is resilient to these issues or uncertainties. For example, it delivers an average of 25X and 5.7X speedup compared to PS in oversubscribed network and Ring in network with failures, respectively.

Remote direct memory access over converged Ethernet deployments is vulnerable to deadlocks induced by priority flow control. Prior solutions for deadlock prevention either require significant changes to routing protocols or require excessive buffers in the switches. In this paper, we propose Tagger, a scheme for deadlock prevention. It does not require any changes to the routing protocol and needs only modest buffers. Tagger is based on the insight that given a set of expected lossless routes, a simple tagging scheme can be developed to ensure that no deadlock will occur under any failure conditions. Packets that do not travel on these lossless routes may be dropped under extreme conditions. We design such a scheme, prove that it prevents deadlock, and implement it efficiently on commodity hardware.

Today's cloud networks are shared among many tenants. Bandwidth guarantees and work conservation are two key properties to ensure predictable performance for tenant applications and high network utilization for providers. Despite significant efforts, very little prior work can really achieve both properties simultaneously even some of them claimed so. In this paper, we present QShare, a comprehensive in-network solution to achieve bandwidth guarantees and work conservation simultaneously. QShare leverages weighted fair queuing on commodity switches to slice network bandwidth for tenants, and solves the challenge of queue scarcity through balanced tenant placement and dynamic tenant-queue binding. We have implemented a QShare prototype and evaluated it extensively via both testbed experiments and simulations. Our results show that QShare ensures bandwidth guarantees while driving network utilization to over 91% even under unpredictable traffic demands.