With the expansion of applications and areas of blockchains, there is an increasing need for faster transaction validation platforms. The Practical Byzantine Fault Tolerance (PBFT) algorithm is one of the best validation processes to meet the need for velocity and security of successful consensus. Despite the implementations in several platforms such as Hyperledger Fabric, the performance does not reach what was hoped for their consensuses. The objective of this article is to present a performance evaluation model based on an Open Queuing Network to study the limits of PBFT and pipelined PBFT and to succeed in presenting a new parallelization approach in the form of parallel clusters. The advantage of this model is that it closely evaluates the different stages of PBFT execution according to the load submitted by the clients. This model is validated with the ns-3 simulator and has the advantage of quickly reaching evaluations that the simulator is limited in time to have them. The results show the importance of having consensus executions on parallel clusters to distribute the loads between nodes and clients. With cluster-based parallel PBFT, we can reach even twenty thousand transactions per second close to the validation values of Visa transactions. This opens the horizons towards the effective use of blockchains for new types of fast applications.
Improving network performance and user experience by tuning network configurations is crucial to cellular service providers (CSPs). However, predicting the performance impacts of configuration changes is non-trivial. The large scale, diversity and complexity of configuration parameters and base station deployments, and more importantly, the uncontrollable external factors ( e.g. , weather and called latents ), lead to confounding effects between configurations and performance metrics. In this article, we show that the effects of latents can be properly mitigated by considering intermediates, called Mobility, Access, and Traffic (MAT) metrics , which separate the configurations and latents from performance metrics. Then, we propose the C onfiguration I mpact P rediction A nalysis T oolkit (CIPAT), a novel two-stage toolkit, driven by a large real-world dataset from live LTE and 5G networks. Our extensive evaluation shows that CIPAT enables network operators to confidently predict the performance impact of candidate configuration settings with an accuracy of up to 86% and an efficacy of up to 85%.
We present a methodology for automated performance regression testing at the unit level, designed to integrate seamlessly into the development workflow of high-performance computing (HPC) applications. The proposed framework emphasizes the integration of performance tests with existing functional unit tests by leveraging their shared structure, thereby enabling the development of performance test suites with minimal developer effort and supporting the early detection of performance degradations. As a proof of concept, we implement and release the framework as a publicly available Julia package. It generates architecture-agnostic test suites that remain valid across heterogeneous hardware environments. To demonstrate its effectiveness, we apply it to representative HPC use cases, including memory throughput analysis in stencil-based computations and roofline modeling of a sparse linear system solver. Experiments conducted on a diverse set of CPU and GPU architectures confirm the framework’s ability to detect performance regressions consistently, validating its potential for use in continuous performance testing pipelines.
This paper examines how environmental regulation shapes pricing, capacity allocation, and infrastructure investment in the competitive Content Delivery Network (CDN) market. We develop a hierarchical game-theoretic framework modeling strategic interactions among multiple CDNs, a service provider, and end-users, capturing both economic and environmental objectives. Using backward induction, we analyze equilibrium behavior under two regulatory regimes–a baseline scenario without intervention and a regulated scenario, where CDNs incur taxes proportional to dirty capacity usage. Our analysis yields four principal findings. First, a structural tax threshold τ * , governed by the gap between renewable and fossil-fuel energy costs, determines the transition from carbon-intensive to clean infrastructure. Second, a moderate tax slightly above τ * achieves near-complete clean adoption at a total welfare loss of only 6.5 %. Third, government revenue is non-monotonic in the tax rate, peaking at low levels and collapsing as CDNs shift to untaxed clean capacity. Fourth, since the infrastructure-transition threshold is governed by the energy-cost gap of each provider, a uniform carbon tax imposes asymmetric burdens once these costs are heterogeneous across CDNs, motivating tiered tax designs. These results provide insights for policymakers seeking to align sustainability objectives with competitive dynamics in the content delivery market.
Communication overhead is a main bottleneck in federated learning (FL) especially in the wireless environment due to the limited data rate and unstable radio channels. The communication challenge necessitates holistic selection of participating clients that accounts for both the computation needs and communication cost, as well as judicious allocation of the limited transmission resource. Meanwhile, the random unpredictable nature of both the training data samples and the communication channels requires an online optimization approach that adapts to the changing system state over time. In this work, we consider a general framework of online joint client sampling and power allocation for wireless FL under time-varying communication channels. We formulate it as a stochastic network optimization problem that admits a Lyapunov-typed solution approach. This leads to per-training-round subproblems with a special bi-convex structure, which we leverage to propose globally optimal solutions, culminating in a meta algorithm that provides strong performance guarantees. We further study three specific FL problems covering multiple scenarios, namely, with IID or non-IID data, whether robustness against data drift is required, and with unbiased or biased client sampling. We derive detailed algorithms for each of these problems. Simulation with standard classification tasks demonstrate that the proposed communication-aware algorithms outperform their counterparts under a wide range of learning and communication scenarios.
Graph Neural Networks (GNNs) have become increasingly popular for their ability to learn the complex features of graph-structured data effectively. However, many real-world graphs are dynamic and change over time in terms of graph structures and features. A large dynamic graph is commonly stored in distributed graph stores and learned through distributed GNN training. Classical graph partition algorithms focus on partition balance and cross-partition edge reduction, which do not serve the need of distributed dynamic GNN learning well. We propose DistDy, a novel online graph partition framework tailored for distributed dynamic GNN learning, aiming to minimize dynamic graph storage overhead and inter-server communication. We design distributed additive storage to store changes in the large dynamic graph, and decide graph partition (aka change storage) on the go by formulating it into a communication utility maximization problem. An efficient online graph partition algorithm is proposed, which computes near-optimal partition strategies according to refined resource prices and additive storage rewards, achieving a proven competitive ratio. Experiments on various real-world and synthetic dynamic graph datasets show that DistDy can achieve 92.2% storage saving and up to 1.39× speed-up in distributed GNN training as compared to using representative graph partition algorithms.
Federated learning is a new learning paradigm which utilizes crowdsourced data stored at dispersed user devices (aka clients) to learn a global model. Studies have shown that even though data are kept on local devices, an adversary is still able to infer client information during the training process or from the learned model. Differential privacy has recently been introduced to deep learning model training, to protect data privacy of clients. Nonetheless, it exacerbates unfairness with the learned model among participating clients due to its uniform clipping and noise addition, even when the training loss function explicitly considers unfairness. To validate the impact of the differential privacy mechanism in federated learning, we carefully approximate the correlation between fairness performance across clients and the fundamental operations within the differential privacy mechanism and quantify the influence of differential privacy mechanisms on model performance across various clients. Subsequently, leveraging our theoretical findings regarding the effect of the differential privacy mechanism, we formulate the unfairness mitigation problem and propose an algorithm based on the modified method of differential multipliers. Extensive evaluation shows that our method outperforms state-of-the-art differentially private federated learning algorithm by about 30% for non-i.i.d. data distribution in terms of the variance of model performance across clients.
Distributed ledgers are common in the industry. Some of them can use blockchains as their underlying infrastructure. A blockchain requires participants to agree on its contents. This can be achieved via a consensus protocol. How do these protocols differ in performance, and how are the differences affected by the communication network? Moreover, such a protocol would need a timer to ensure progress, but how should the timer be set? This article presents an analytical model to address these and related issues when there are crash faults. Specifically, it focuses on two consensus protocols (Istanbul BFT and HotStuff) and two network topologies (Folded-Clos and Dragonfly). The model provides formulas that express the consensus time in terms of protocol and topology parameters. No other model in the literature provides such a global view of the parameter space. Analysis of the closed-form expressions yield new insights into how the timers should be set, how faults affect the consensus time, when one protocol is faster than the other, and how the two topologies differ in their impact. The formulas and analyses are validated with simulations. The conclusion also offers some tips for the analytical modeling of similar protocols.
Traditional workload analysis uses discrete times measured by data accesses. An example is the classic independent reference model (IRM). Effective solutions have been developed to model workloads with stochastic access patterns, but they incur a high cost for Zipfian workloads, which may contain millions of items each accessed with a different frequency. This article first presents a continuous-time model of locality for workloads with stochastic access patterns. It shows that two previous techniques by Dan and Towsley and by Denning and Schwartz can be interpreted as a single model using different discrete times. Using continuous time, it derives a closed-form solution for an item and a general solution, that is, a differentiable function. In addition, the article presents an approximation technique by grouping items into partitions. When evaluated using Zipfian workloads, it shows that a workload with millions of items can be approximated using a small number of partitions, and the continuous-time model has greater accuracy and is faster to compute numerically. For the largest data size verifiable using trace generation and simulation, the new techniques reduce the time of locality analysis by six orders of magnitude.
We consider the performance analysis of an M/M/1/N queueing system where the server consumes energy from a battery during the service of customers. The energy supply of the battery depends on a randomly changing environment. When the battery gets empty, the server goes on vacation for a random amount of time. We model the energy level of the battery as the fluid in an infinite fluid buffer and the environment as a continuous-time Markov chain (CTMC). The analytical framework resembles the one used in fluid vacation models with exhaustive discipline, but the properties of the considered queueing model require the extension of the available methodology, because the model evolution is different during the service and the vacation period due to the inactivity of the server during vacation. Essentially, new results are derived to cope with the general properties of the considered model. Consequently, the results in this article extend the analysis of fluid vacation models to a more general class than the fluid vacation models with exhaustive discipline. The steady-state vector density of the fluid level, its Laplace transform, and the mean fluid level are derived together with the probability mass function of the queue size. A special case of the model with phase-type (PH)-distributed vacation time is also analyzed.
Growing concerns about centralized mining of personal data threatens to stifle further proliferation of machine learning (ML) applications. Consequently, a recent trend in ML training advocates for a paradigm shift – moving the computation of ML models from a centralized server to a federation of edge devices owned by the users whose data is to be mined. Though such decentralization aims to alleviate concerns related to raw data sharing, it introduces a set of challenges due to the hardware heterogeneity among the devices possessing the data. The heterogeneity may, in the most extreme cases, impede the participation of low-end devices in the training or even prevent the deployment of the ML model to such devices. Recent research in distributed collaborative machine learning (DCML) promises to address the issue of ML model training over heterogeneous devices. However, the actual extent to which the issue is solved remains unclear, especially as an independent investigation of the proposed methods’ performance in realistic settings is missing. In this paper, we present a detailed survey and an evaluation of algorithms that aim to enable collaborative model training across diverse devices. We explore approaches that harness three major strategies for DCML, namely Knowledge Distillation, Split Learning, and Partial Training, and we conduct a thorough experimental evaluation of these approaches on a real-world testbed of 14 heterogeneous devices. Our analysis compares algorithms based on the resulting model accuracy, memory consumption, CPU utilization, network activity, and other relevant metrics, and provides guidelines for practitioners as well as pointers for future research in DCML.
The Federated Learning paradigm is a distributed machine learning strategy, developed for settings where training data is owned by distributed devices and cannot be shared with others. Federated Learning circumvents this constraint by carrying out model training in distribution, so that each participant, or client, trains a local model only on its own data. The parameters of these local models are shared intermittently among participants and aggregated to enhance model accuracy. This strategy has shown impressive success, and has been rapidly adopted by the industry in efforts to overcome confidentiality and resource constraints in model training. However, the application of FL to real-world settings brings additional challenges, many associated with heterogeneity between participants. Research into mitigating these difficulties in Federated Learning has largely focused on only two particular types of heterogeneity: the unbalanced distribution of training data, and differences in client resources. Yet many more types of heterogeneity exist, and some are becoming increasingly relevant as the capability of FL expands to cover more and more complex real-world problems, from the tuning of large language models to enabling machine learning on edge devices. In this work, we discuss a novel type of heterogeneity that is likely to become increasingly relevant in future applications: this is preference heterogeneity, emerging when clients learn under multiple objectives, with different importance assigned to each objective on different clients. In this work, we discuss the implications of this type of heterogeneity and propose a FedPref, a first algorithm designed to facilitate personalised federated learning in this setting. We demonstrate the effectiveness of the algorithm across several different problems, preference distributions and model architectures. In addition, we introduce a new analytical point of view, based on multi-objective metrics, for evaluating the performance of federated algorithms in this setting beyond the traditional client-focused metrics. We perform a second experimental analysis based in this view, and show that FedPref outperforms compared algorithms.
Federated Learning (FL) is a promising technique for decentralized privacy-preserving Machine Learning (ML) with a diverse pool of participating devices with varying device capabilities. However, existing approaches to handle such heterogeneous environments do not consider “fairness” in model aggregation, resulting in significant performance variation among devices. Meanwhile, prior works on FL fairness remain hardware-oblivious and cannot be applied directly without severe performance penalties. To address this issue, we propose a novel hardware-sensitive FL method called \(\mathsf {FairHetero}\) that promotes fairness among heterogeneous federated clients. Our approach offers tunable fairness within a group of devices with the same ML architecture as well as across different groups with heterogeneous models. Our evaluation under MNIST, FEMNIST, CIFAR10, and SHAKESPEARE datasets demonstrates that \(\mathsf {FairHetero}\) can reduce variance among participating clients’ test loss compared to the existing state-of-the-art (SOTA) techniques, resulting in increased overall performance.
Federated Learning (FL) enables collaborative model training across decentralized edge devices while preserving data privacy. However, existing FL methods often assume clean annotated datasets, impractical for resource-constrained edge devices. In reality, noisy labels are prevalent, posing significant challenges to FL performance. Prior approaches attempt label correction and robust training techniques but exhibit limited efficacy, particularly under high noise levels. This paper introduces ClipFL (Federated Learning Client Pruning), a novel framework addressing noisy labels from a fresh perspective. ClipFL identifies and excludes noisy clients based on their performance on a clean validation dataset, tracked using a Noise Candidacy Score (NCS). The framework comprises three phases: pre-client pruning to identify potential noisy clients and calculate their NCS, client pruning to exclude a percentage of clients with the highest NCS, and post-client pruning for fine-tuning the global model with standard FL on clean clients. Empirical evaluation demonstrates ClipFL’s efficacy across diverse datasets and noise levels, achieving accurate noisy client identification, superior performance, faster convergence, and reduced communication costs compared to state-of-the-art FL methods. Our code is available at https://github.com/MMorafah/ClipFL.
Federated Learning (FL) has gained attention across various industries for its capability to train machine learning models without centralizing sensitive data. While this approach offers significant benefits such as privacy preservation and decreased communication overhead, it presents several challenges, including deployment complexity and interoperability issues, particularly in heterogeneous scenarios or resource-constrained environments. Over-the-air (OTA) FL was introduced to tackle these challenges by disseminating model updates without necessitating direct device-to-device connections or centralized servers. However, OTA-FL brought forth limitations associated with heightened energy consumption and network latency. In this paper, we propose a multi-attribute client selection framework employing the grey wolf optimizer (GWO) to strategically control the number of participants in each round and optimize the OTA-FL process while considering accuracy, energy, delay, reliability, and fairness constraints of participating devices. We evaluate the performance of our multi-attribute client selection approach in terms of model loss minimization, convergence time reduction, and energy efficiency. In our experimental evaluation, we assessed and compared the performance of our approach against the existing state-of-the-art methods. Our results demonstrate that the proposed GWO-based client selection outperforms these baselines across various metrics. Specifically, our approach achieves a notable reduction in model loss, accelerates convergence time, and enhances energy efficiency while maintaining high fairness and reliability indicators.
We propose and experimentally evaluate a novel secure aggregation algorithm targeted at cross-organizational federated learning applications with a fixed set of participating learners. Our solution organizes learners in a chain and encrypts all traffic to reduce the controller of the aggregation to a mere message broker. We show that our algorithm scales better and is less resource demanding than existing solutions, while being easy to implement on constrained platforms. With 36 nodes our method outperforms state-of-the-art secure aggregation by 70x, and 56x with and without failover, respectively.
In the dynamic realm of machine learning (ML), federated learning (FL) emerges as a pivotal method for training models on decentralized devices without the need for central data aggregation. This technique confronts the challenges of data heterogeneity, which disrupts the independent and identically distributed (IID) assumptions, adversely affecting the accuracy of the overall model. To tackle this issue, we introduce the federated non-performing-node-resilient neural selector (FNNS), an advanced client selection algorithm grounded in a combinatorial contextual neural bandit framework. This algorithm enhances the extraction of contextual data by assessing each local client using a universally standardized dataset, thereby providing a deeper, context-specific insight suitable for federated environments. In addition, we introduce selection robustness score (SRS), a novel metric designed to quantify the efficacy of client selection in the presence of non-performing-nodes (NPN) conditions. Using this metric, we demonstrate FANS’s effectiveness in enhancing the FL process. Empirical evaluations across diverse settings reveal our method’s superiority over current state-of-theart solutions, with significant improvements in both SRS and global model accuracy.
Since inception, DNS has used a TTL-based replication scheme that allows the source (i.e., an authoritative domain server) to control the frequency of record eviction from client caches. Existing studies of DNS predominantly focus on reducing query latency and source bandwidth, both of which are optimized by increasing the cache hit rate. However, this causes less-frequent contacts with the source and results in higher staleness of retrieved records. Given high data-churn rates at certain providers (e.g., dynamic DNS, CDNs) and importance of consistency to their clients, we propose that cache models include the probability of freshness as an integral performance measure. We derive this metric under general update/download processes and present a novel framework for measuring its value using remote observation (i.e., without access to the source or the cache). Besides freshness, our methods can estimate the inter-update distribution of DNS records, cache hit rate, distribution of TTL, and query arrival rate from other clients. Furthermore, these algorithms do not require any changes to the existing infrastructure/protocols.
