Decentralized Online Learning (DOL) extends online learning to the domain of distributed networks. However, limitations of local data in decentralized settings lead to a decrease in the accuracy of decisions or models compared to centralized methods. Considering the increasing requirement to achieve a high-precision model or decision with distributed data resources in a network, applying ensemble methods is attempted to achieve a superior model or decision with only transferring gradients or models. A new boosting method, namely Boosting for Distributed Online Convex Optimization (BD-OCO), is designed to realize the application of boosting in distributed scenarios. BD-OCO achieves the regret upper bound
- Article type
- Year
- Co-author
Apache Spark provides a well-known MapReduce computing framework, aiming to fast-process big data analytics in data-parallel manners. With this platform, large input data are divided into data partitions. Each data partition is processed by multiple computation tasks concurrently. Outputs of these computation tasks are transferred among multiple computers via the network. However, such a distributed computing framework suffers from system overheads, inevitably caused by communication and disk I/O operations. System overheads take up a large proportion of the Job Completion Time (JCT). We observed that excessive computational resources incurs considerable system overheads, prolonging the JCT. The over-allocation of individual jobs not only prolongs their own JCTs, but also likely makes other jobs suffer from under-allocation. Thus, the average JCT is suboptimal, too. To address this problem, we propose a prediction model to estimate the changing JCT of a single Spark job. With the support of the prediction method, we designed a heuristic algorithm to balance the resource allocation of multiple Spark jobs, aiming to minimize the average JCT in multiple-job cases. We implemented the prediction model and resource allocation method in ReB, a Resource-Balancer based on Apache Spark. Experimental results showed that ReB significantly outperformed the traditional max-min fairness and shortest-job-optimal methods. The average JCT was decreased by around 10%-30% compared to the existing solutions.
Visible-Light Communication (VLC) has the potential to provide dense and fast connectivity at low cost. In this paper we propose SFNet, a novel VLC-enabled hybrid data center network that extends the design of wireless Data Center Networks (DCNs) into three further dimensions: (1) fully wireless at the inter-rack level; (2) no need for a centralized control mechanism on wireless links; and (3) no need for any infrastructure-level alterations to data centers. Previous proposals typically cannot realize these three rationales simultaneously. To achieve this vision, the proposed SFNet augments fat-tree by organizing all racks into a wireless small-world network via VLC links. The use of VLC links eliminates hierarchical switches and cables in the wireless network, and thus reduces hardware investment and maintenance costs. To fully exploit the benefits of the topology of SFNet, we further propose its topology design and optimization method, routing scheme, and online flow scheduling algorithm. Comprehensive experiments indicate that SFNet exhibits good topological properties and network performance.
To satisfy the ever-increasing bandwidth demand of modern data centers, researchers have proposed hybrid Data Center Networks (DCNs), which employ high-bandwidth Optical Circuit Switches (OCSs) to compensate for Electrical Packet Switches (EPS). Existing designs, such as Helios and c-Through, mainly focus on reconfiguring optical devices to meet the estimated traffic requirements. However, these designs face two major challenges in their OCS-based networks, namely, the complex control mechanism and cabling problems. To solve these challenges, we propose TIO, a hybrid DCN that employs Visible Light Communication (VLC) instead of wired OCS design to connect racks. TIO integrates the wireless VLC-based Jellyfish and wired EPS-based Fat Tree seamlessly and combines the opposite and complementary characteristics, including wireless VLC direct connection and wired electrical packet switching, random graph, and Clos topology properties. To further exploit the merits of TIO, we design a hybrid routing scheme and congestion-aware flow scheduling method. Comprehensive evaluations indicate that TIO outperforms the Jellyfish and Fat Tree in both topology properties and network performance, and the flow scheduling method also evidently improves performance.
Set reconciliation between two nodes is widely used in network applications. The basic idea is that each member of a node pair has an object set and seeks to deliver its unique objects to the other member. The Standard Bloom Filter (SBF) and its variants, such as the Invertible Bloom Filter (IBF), are effective approaches to solving the set reconciliation problem. The SBF-based method requires each node to represent its objects using an SBF, which is exchanged with the other node. A receiving node queries the received SBF against its local objects to identify the unique objects. Finally, each node exchanges its unique objects with the other node in the node pair. For the IBF-based method, each node represents its objects using an IBF, which is then exchanged. A receiving node subtracts the received IBF from its local IBF so as to decode the different objects between the two sets. Intuitively, it would seem that the IBF-based method, with only one round of communication, entails less communication overhead than the SBF-based method, which incurs two rounds of communication. Our research results, however, indicate that neither of these two methods has an absolute advantages over the others. In this paper, we aim to provide an in-depth understanding of the two methods, by evaluating and comparing their communication overhead. We find that the best method depends on parameter settings. We demonstrate that the SBF-based method outperforms the IBF-based method in most cases. But when the number of different objects in the two sets is below a certain threshold, the IBF-based method outperforms the SBF-based method.