In large-scale industrial recommendation systems, model checkpoints are instrumental in maintaining training goodput and numerical correctness during system failures and job preemptions. The increasing prevalence of multi-terabyte models has rendered frequent regular model checkpoints impractical, resulting in substantial lost progress when recovering from failures. As model sizes continue to grow, researchers and practitioners are compelled to investigate more efficient and scalable solutions. This paper presents DECK, a novel approach to delta model checkpointing designed for real-world industrial systems. Specifically, DECK focuses on extracting delta states with near-zero overhead, staging and streaming delta checkpoints without interrupting the training process, and merging delta checkpoints in an optimal and decoupled manner. Experimental results demonstrate that DECK achieves a 12-fold increase in checkpoint frequency while maintaining negligible impact on training throughput, thereby attaining state-of-the-art (SOTA) production performance.

Signal decay is the fundamental problem of wireless communications, especially in an indoor environment where line-of-sight (LOS) paths for signal propagation are often blocked and various indoor objects exacerbate signal fading. There are three reasons for signal decay: long transmission distance, signal penetration, and reflection. In this paper, we propose OptRe which optimally places metallic reflectors - providing a highly reflective surface that can reflect impinging signals almost 100% - in indoor environments to reduce the reflection loss and enhance wireless transmissions. It enhances both WiFi signal and low-power IoT devices without changing their configurations or network protocols. To enable OptRe, we first develop an empirical signal propagation model that can accurately estimate the signal strength and adapt itself to the reflectors' location. Using micro-benchmarks, our empirical signal propagation model is shown to be more accurate than the other existing path loss models. We also optimally place reflectors to maximize the worst-case signal coverage within the target indoor areas. Our extensive experimental evaluation results have shown OptRe to enhance signal strength for different types of wireless signals by almost 2x.

To meet the diverse traffic demands of different applications, emerging WLAN standards have been incorporating a variety of channel widths ranging from 5 to 160 MHz. The coexistence of variable-width channels imposes a new challenge to the 802.11 protocols, since the 802.11 MAC is agnostic of, and thus incapable of, adapting to the PHY-layer spectrum heterogeneity. To address this challenge, we uncover the cause and effect of variable-width channel coexistence, and develop a MAC-layer scheme, called Fine-grained Spectrum Sharing (FSS), that solves the general problem of fair and efficient spectrum sharing among users with heterogeneous channel-widths. Instead of deeming its spectrum band as an atomic block, an FSS user divides the spectrum into chunks, and adapts its chunk usage on a per-packet basis. FSS's spectrum adaptation is driven by a decentralized optimization framework. It preserves the 802.11 CSMA/CA primitives while allowing users to contend for each spectrum chunk, and can opportunistically split a wide-band channel or bond multiple (discontiguous) chunks to ensure fair and efficient access to available spectrum. In making such adaptation decisions, it balances the benefit from discontiguous chunks and the cost of guardband - a unique tradeoff in WLANs with heterogeneous channel-widths. Our in-depth evaluation demonstrates that FSS can improve throughput by multiple folds, while maintaining fairness of spectrum sharing for heterogeneous WLANs.

Detecting how a vehicle is steered and then alarming drivers in real time is of utmost importance to the vehicle and the driver's safety, since fatal accidents are often caused by dan- gerous steering. Existing solutions for detecting dangerous maneuvers are implemented either in only high-end vehicles or on smartphones as mobile applications. However, most of them rely on the use of cameras, the performance of which is seriously constrained by their high visibility requirement. Moreover, such an over/sole-reliance on the use of cameras can be a distraction to the driver.

Multiple-In-Multiple-Out (MIMO) offers great potential for increasing network capacity by exploiting spatial diversity with multiple antennas. Multiuser MIMO (MU-MIMO) further enables Access Points (APs) with multiple antennas to transmit multiple data streams concurrently to several clients. In MU-MIMO, clients need to estimate Channel State Information (CSI) and report it to APs in order to eliminate interference between them. We explore the vulnerability in clients' plaintext feedback of estimated CSI to the APs and propose two advanced attacks that malicious clients can mount by reporting forged CSI: (1) sniffing attack that enables concurrently transmitting malicious clients to eavesdrop other ongoing transmissions; (2) power attack that enables malicious clients to enhance their own capacity at the expense of others?. We have implemented and evaluated these two attacks in a WARP testbed. Based on our experimental results, we suggest a revision of the current CSI feedback scheme and propose a novel CSI feedback system, called the CSIsec, to prevent CSI forging without requiring any modification at the client side, thus facilitating its deployment.

In Cognitive Radio Networks (CRNs), Primary Users (PUs) can share their idle spectra with Secondary Users (SUs) under certain mechanisms. In this paper, we propose a distributed-centralized and incentive-aware spectrum sharing scheme for the multiple-PU scenario, which introduces a Random Leader who is elected randomly from SUs or PUs. The distributed aspect of our scheme lies in that it requires no central control entities, which can be independently implemented within a distributed spectrum market. The centralized aspect is that the leader draws up and assigns the socially optimal contracts for all PUs and SUs in a centralized manner, which maximizes the throughput of the whole network and attains the economic robustness (including Incentive Compatibility and Individual Rationality). Analysis shows that the proposed scheme takes in the advantages of both centralized and distributed schemes but overcomes their weaknesses. We use the proposed scheme to study two sharing scenarios: the short-term and the long-term spectrum sharing. The Short-Term Sharing (STS) focuses on distributing PUs' idle spectra within one time slot while the Long-Term Sharing (LTS) considers multiple slots, where the spectrum mobility must be investigated. As an integrated design of both STS and LTS, our scheme not only fulfils SUs' heterogeneous spectrum requirements but also obtains the socially optimal throughput while accounting for both SUs' and PUs' incentives.