Cluster-based Servers Research Articles

Yuz: Improving Performance of Cluster-Based Services by Near-L4 Session-Persistent Load Balancing

Yuz: Improving Performance of Cluster-Based Services by Near-L4 Session-Persistent Load Balancing

SC-FGCL: Self-Adaptive Cluster-Based Federal Graph Contrastive Learning

As a self-supervised learning method, the graph contrastive learning achieve admirable performance in graph pre-training tasks, and can be fine-tuned for multiple downstream tasks such as protein structure prediction, social recommendation, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">etc.</i> One prerequisite for graph contrastive learning is the support of huge graphs in the training procedure. However, the graph data nowadays are distributed in various devices and hold by different owners, like those smart devices in Internet of Things. Considering the non-negligible consumptions on computing, storage, communication, data privacy and other issues, these devices often prefer to keep data locally, which significantly reduces the graph contrastive learning performance. In this paper, we propose a novel federal graph contrastive learning framework. First, it is able to update node embeddings during training by means of a federation method, allowing the local GCL to acquire anchors with richer information. Second, we design a Self-adaptive Cluster-based server strategy to select the optimal embedding update scheme, which maximizes the richness of the embedding information while avoiding the interference of noise. Generally, our method can build anchors with richer information through a federated learning approach, thus alleviating the performance degradation of graph contrastive learning due to distributed storage. Extensive analysis and experimental results demonstrate the superiority of our framework.

Self-adaptive load-balancing strategy based on a time series pattern for concurrent user access on Web map service

Load-balancing strategies address challenges stemming from intensive access and heavy communication traffic on a Web map service platform (WMSP) by collecting workload-related information from cluster-based servers, and distributing tasks to minimize the consumption of computational and caching resources. However, intensive user access has time series patterns that create temporal periodic variations that can be exploited to improve the performance of the WMSP. In this paper, we propose a variable feedback strategy based on time series variations in the intensity of user access to increase the efficiency and reliability of workload feedback with little resource consumption. A task distribution strategy, based on the expected values of the arrival rate that match real-time conditions in workload feedback periods and the real-time processing capability of each cluster-based server, is devised simply and accurately by an association strategy for workload and service rate that supports services insensitive to massive numbers of concurrent access requests. The results of experiments show that the proposed strategy provides quick responses and high throughput for large-scale user access. It implements efficient load balancing for service resource utilization, and can thus improve the stability and capacity of the WMSP server using heterogeneous back-end cloud cluster-based servers.

Energy management for the homogeneous server clusters offering web services

The population of the Internet users has exceeded 2.4 billion. Data centers provide Internet services to fulfill the demand from these users. Many data centers adopt the cluster-based server systems to host the required Internet services. These server systems consume significant amount of energy, but much of the power is used to maintain service capacity during idle or low workload periods. This paper surveys some recent approaches addressing this issue. As learned from the traditional telephone call center planning processes, a queueing model is adopted to model the server clusters with homogeneous architecture. By analyzing the model, a set of the factors affecting the energy cost is identified. Based on the identified factors, an on-line energy management technique is then deigned. The proposed approach is simulated with a real-world workload trace. The simulation result shows that approximately 70 to 75 % of the originally consumed energy can be saved.

HCBLS: A Hierarchical Cluster-Based Location Service in Urban Environment

Vehicle location information is central to many location-based services and applications in VANETs. Tracking vehicles positions and maintaining an accurate up-to-date view of the entire network are not easy due to the high mobility of vehicles and consequently rapid topology changes. The design of a scalable, accurate, and efficient location service is still a very challenging issue. In this paper, we propose a lightweight hierarchical cluster-based location service in city environments (HCBLS). HCBLS integrates a logical clustering based on the city digital map and consequently does not involve extra signaling overhead. An advanced location update aggregation at different levels of the assumed hierarchy is adopted to maintain up-to-date and accurate location information. Simulation results show that HCBLS achieves much better performances than the Efficient Map-Based Location Service (EMBLS) and any regular (non-cluster-based) updating scheme. HCBLS increases the success rate by around 10%, improves the overview of the network by more than 30%, lowers the location update and query costs by more than 7 times, lowers the message delivery latency by around 3 times, and presents around 4 times better localization accuracy.

A Load-balancing method for network GISs in a heterogeneous cluster-based system using access density

The uneven distribution of data access and imbalances in the processing capability of heterogeneous servers are two important factors that affect load balancing for network geographic information systems. This article presents a load-balancing method that considers both localized access control and balanced load allocation. First, the method considers access patterns for terrain data (tiles) that follow the Zipf law as well as the different processing performance of servers in a heterogeneous cluster-based environment. Adapting to intense user access by distributing heterogeneous cluster-based caching, the proposed method balances the access load for hotspot data to yield a higher hit rate. Then, queue theory is applied to solve the minimum processing cost for data requests in view of the overall heterogeneous cluster-based server performance, balancing the load for each server according to its processing capability and as a result, obtaining the optimal response time. Finally, using the cache distribution strategy mentioned above, data requests are distributed according to their content to prevent over-concentration of loads caused by hotspot data access. This approach takes into account large-scale traffic and highly aggregated user access preferences, adapting to the intensity of data access requests and hence handles more access traffic per unit time. Experimental results reveal that the proposed method obtains a good response performance and higher system throughput and, consequently, improves the utilization efficiency of large-scale network geographic information systems.

클러스터 VOD 서버의 부분적 장애에서 QoS 보장

For large scale VOD service, cluster servers are spotlighted to their high performance and low cost. A cluster server usually consists of a front-end node and multiple back-end nodes. Though increasing the number of back-end nodes can result in the more QoS streams for clients, the possibility of failures in back-end nodes is proportionally increased. The failure causes not only the stop of all streaming service but also the loss of the current playing positions. In this paper, when a back-end node becomes a failed state, the recovery mechanisms are studied to support the unceasing streaming service. For the actual VOD service environment, we implement a cluster-based VOD servers composed of general PCs and adopt the parallel processing for MPEG movies. From the implemented VOD server, a video block recovery mechanism is designed on parity algorithms. However, without considering the architecture of cluster-based VOD server, the application of the basic technique causes the performance bottleneck of the internal network for recovery and also results in the inefficiency CPU usage of back-end nodes. To address these problems, we propose a new failure recovery mechanism based on the pipeline computing concept.

Parallel failure recovery techniques in cluster-based media servers

For large-scale video-on-demand (VOD) service, cluster servers are highlighted due to their high performance and low cost. A cluster server consists of a front-end node and multiple backend nodes. Though the increase in backend nodes provides more quality of service (QoS) streams, the possibility of backend node failure is proportionally increased. The failure causes not only the cessation of streaming services but also the loss of current playing positions. In this paper, when a backend node fails, recovery mechanisms are studied to support the streaming service continuously. Without considering the characteristics of cluster-based servers and MPEG media, the basic redundant array of independent disks (RAID) techniques cause a network bottleneck in the internal network path and demonstrate inefficient CPU usage in backend nodes. To address these problems, a new failure recovery mechanism is proposed based on the pipeline computing concept. The proposed method not only distributes the internal network traffic generated from the recovery operations but also utilizes the CPU time available in the backend nodes. In the experiments, even if a backend node fails, the proposed method provides continuous streaming media services within a short MTTR value as well as more QoS streams than the existing method.

Recovery Strategies for Streaming Media Service in a Cluster-Based VOD Server with a Fault Node

Due to the economic cost and good scalability, cluster-based server architecture is used for VOD services. This server consists of a front-end node and multiple backend nodes. In this server architecture, backend nodes are added simply to support large-scale on-demand clients. However, as the number of backend nodes increases, the possibility of backend node failure also increases. A backend node fault not only degrades the quality of serviced streaming media but also decreases the number of streams supported in the VOD server. For successful VOD service, even if a backend node enters a fault state, the streaming service in progress should be re-continued after a short recovery time. As the recovery strategy, when legacy RAID methods are applied to cluster-based VOD servers, the excessive internal network traffic between the backend nodes causes performance degradation. In addition, the backend nodes demonstrate inefficient CPU utilization for the recovery process. In this paper, to address these problems, a new fault recovery strategy is proposed based on the pipeline computing concept. The proposed method not only distributes the network traffic generated from the recovery operations but also makes efficient use of the CPU time available in the backend nodes. Based on these advantages, even if the cluster-based server has a backend node that fails, the proposed method provides more QoS streams compared to the existing recovery method. In addition, since the proposed method needs a very short recovery time, the streaming services in progress are sustained without degradation of media quality.

Resource allocation optimization for quantitative service differentiation on server clusters

The goal of service differentiation is to provide different service quality levels to meet changing system configuration and resource availability and to satisfy different requirements and expectations of applications and users. In this paper, we investigate the problem of quantitative service differentiation on cluster-based delay-sensitive servers. The goal is to support a system-wide service quality optimization with respect to resource allocation on a computer system while provisioning proportionality fairness to clients. We first propose and promote a square-root proportional differentiation model. Interestingly, both popular delay factors, queueing delay and slowdown, are reciprocally proportional to the allocated resource usage. We formulate the problem of quantitative service differentiation as a generalized resource allocation optimization towards the minimization of system delay, defined as the sum of weighted delay of client requests. We prove that the optimization-based resource allocation scheme essentially provides square-root proportional service differentiation to clients. We then study the problem of service differentiation provisioning from an important relative performance metric, slowdown. We give a closed-form expression of the expected slowdown of a popular heavy-tailed workload model with respect to resource allocation on a server cluster. We design a two-tier resource management framework, which integrates a dispatcher-based node partitioning scheme and a server-based adaptive process allocation scheme. We evaluate the resource allocation framework with different models via extensive simulations. Results show that the square-root proportional model provides service differentiation at a minimum cost of system delay. The two-tier resource allocation framework can provide fine-grained and predictable service differentiation on cluster-based servers.

Cluster-based service discovery for heterogeneous wireless sensor networks

We propose an energy-efficient service discovery protocol for heterogeneous wireless sensor networks that interact with the users in open and dynamic environments. Our solution exploits a cluster overlay, where the clusterhead nodes form a distributed service registry. A service lookup results in visiting only the clusterhead nodes. We aim for minimising the communication costs during discovery of services and maintenance of a functional distributed service registry. To achieve these objectives we propose a clustering algorithm which makes decisions based on 1-hop neighbourhood information, avoids chain reactions and constructs a set of sparsely distributed clusterheads. We analyse how the properties of the clustering structure influence the performance of the service discovery protocol, by comparing theoretically and through simulations our proposed clustering algorithm with distributed mobility-adaptive clustering (DMAC). To validate our simulations we implement the proposed solution on resource-constraint sensor nodes and we measure the performance of the protocol running on different testbeds.

Effective load balancing for cluster-based servers employing job preemption

A cluster-based server consists of a front-end dispatcher and multiple back-end servers. The dispatcher receives incoming jobs, and then decides how to assign them to back-end servers, which in turn serve the jobs according to some discipline. Cluster-based servers have been widely deployed, as they combine good performance with low costs. Several assignment policies have been proposed for cluster-based servers, most of which aim to balance the load among back-end servers. There are two main strategies for load balancing: The first aims to balance the amount of workload at back-end servers, while the second aims to balance the number of jobs assigned to back-end servers. Examples of policies using these strategies are Dynamic and LC (Least Connected), respectively. In this paper we propose a policy, called LC*, which combines the two aforementioned strategies. The paper shows experimentally that when preemption is admitted (i.e., when jobs execute concurrently on back-end servers), LC* substantially outperforms both Dynamic and LC in terms of response-time metrics. This improved performance is achieved by using only information readily available to the dispatcher, rendering LC* a practical policy to implement. Finally, we study a refinement, called ALC* (Adaptive LC*), which further improves on the response-time performance of LC* by adapting its actions to incoming traffic rates.

Self-determined mechanisms in complex networks

Self-organized networks are pervasive in communication systems such as the Internet, overlay networks, peer-to-peer networks, and cluster-based services. These networks evolve into complex topologies, under specific driving forces, i.e. user demands, technological innovations, design objectives and so on. Our study focuses on the driving forces behind individual evolutions of network components, and their stimulation and domination to the self-organized networks which are defined as self-determined mechanisms in this paper. Understanding forces underlying the evolution of networks should enable informed design decisions and help to avoid unwanted surprises, such as congestion collapse. A case study on the macroscopic evolution of the Internet topology of autonomous systems under a specific driving force is then presented. Using computer simulations, it is found that the power-law degree distribution can originate from a connection preference to larger numbers of users, and that the small-world property can be caused by rapid growth in the number of users. Our results provide a new feasible perspective to understand intrinsic fundamentals in the topological evolution of complex networks.

A Network Processor-Based, Content-Aware Switch

A content-aware switch can optimize cluster-based server architectures by examining requests and distributing them to servers on the basis of application-level information. A network-processor-based switch can reduce HTTP processing latency and improve packet throughput.

Deferred Assignment Scheduling in Cluster-Based Servers

This paper proposes a new scheduling policy for cluster-based servers called DAS (Deferred Assignment Scheduling). The main idea in DAS is to defer scheduling as much as possible in order to make better use of the accumulated information on job sizes. In broad outline, DAS operates as follows: (1) incoming jobs are held by the dispatcher in a buffer; (2) the dispatcher monitors the number of jobs being processed by each server; (3) when the number of jobs at a server queue drops below a prescribed threshold, the dispatcher sends to it the shortest job in its buffer. To gauge the efficacy of DAS, the paper presents simulation studies, using various data traces. The studies collected response times and slowdowns for two cluster configurations under multi-threaded and multi-process back-end server architectures. The experimental results show that in both architectures, DAS outperforms the Round-Robin policy in all traffic regimes, and the JSQ (Join Shortest Queue) policy in medium and heavy traffic regimes.

Weblins: A scalable WWW cluster-based server

With the ever-growing web traffic, cluster-based web servers have become very important to the Internet infrastructure. Thus, making the best use of all available resources in the cluster to achieve high performance is a significant research issue. In this paper, we present Weblins, a cluster-based web server that can achieve good throughput. Weblins has Gobelins operating system as platform. Gobelins is an efficient single system image operating system that transparently makes use of the resources available in the cluster. The architecture of Weblins is fully distributed. Weblins implements a content-aware request distribution policy via a new interface on top of Gobelins. Popular web files are dynamically replicated on all nodes via a cooperative caching mechanism. For the non-popular files, the requests are handed-off to the corresponding nodes via the TCP Handoff protocol. Simulation results show that the strategy used by Weblins is more suitable for cluster-based Web severs in comparison with pure content-aware strategy and pure cooperative caching strategy.

Quantifying the performability of cluster-based services

In this paper, we propose a two-phase methodology for systematically evaluating the performability (performance and availability) of cluster-based Internet services. In the first phase, evaluators use a fault-injection infrastructure to characterize the service's behavior in the presence of faults. In the second phase, evaluators use an analytical model to combine an expected fault load with measurements from the first phase to assess the service's performability. Using this model, evaluators can study the service's sensitivity to different design decisions, fault rates, and other environmental factors. To demonstrate our methodology, we study the performability of a multitier Internet service. In particular, we evaluate the performance and availability of three soft state maintenance strategies for an online bookstore service in the presence of seven classes of faults. Among other interesting results, we clearly isolate the effect of different faults, showing that the tier of Web servers is responsible for an often dominant fraction of the service unavailability. Our results also demonstrate that storing the soft state in a database achieves better performability than storing it in main memory (even when the state is efficiently replicated) when we weight performance and availability equally. Based on our results, we conclude that service designers may want an unbalanced system in which they heavily load highly available components and leave more spare capacity for components that are likely to fail more often.

Integrated resource management for cluster-based Internet services

Client request rates for Internet services tend to be bursty and thus it is important to maintain efficient resource utilization under a wide range of load conditions. Network service clients typically seek services interactively and maintaining reasonable response time is often imperative for such services. In addition, providing differentiated service qualities and resource allocation to multiple service classes can also be desirable at times. This paper presents an integrated resource management framework (part of Neptune system ) that provides flexible service quality specification, efficient resource utilization, and service differentiation for cluster-based services. This framework introduces the metric of quality-aware service yield to combine the overall system efficiency and individual service response time in one flexible model. Resources are managed through a two-level request distribution and scheduling scheme. At the cluster level, a fully decentralized request distribution architecture is employed to achieve high scalability and availability. Inside each service node, an adaptive scheduling policy maintains efficient resource utilization under a wide range of load conditions. Our trace-driven evaluations demonstrate the performance, scalability, and service differentiation achieved by the proposed techniques.

LSMAC vs. LSNAT: Scalable cluster-based Web servers

Server scalability is more important than ever in today's client/server dominated network environments. Recently, researchers have begun to consider cluster-based computers using commodity hardware as an alternative to expensive specialized hardware for building scalable Web servers. In this paper, we present performance results comparing two cluster-based Web servers based on different server architectures: OSI layer two dispatching (LSMAC) and OSI layer three dispatching (LSNAT). Both cluster-based server systems were implemented as application-space programs running on commodity hardware in contrast to other, similar, solutions which require specialized hardware/software. We point out the advantages and disadvantages of both systems. We also identify when servers should be clustered and when clustering will not improve performance.

Analytical and experimental evaluation of cluster-based network servers

In this paper we use analytic modeling and simulation to evaluate network servers implemented on clusters of workstations. More specifically, we model the potential benefits of locality-conscious request distribution within the cluster and evaluate the performance of a cluster-based server (called L2S) we designed in light of our experience with the model. Our most important modeling results show that locality-conscious distribution on a 16-node cluster can increase server throughput with respect to a locality-oblivious server by up to 5-fold, depending on the average size of the files requested and on the size of the server's working set. Our simulation results demonstrate that L2S achieves throughput that is within 28% of the full potential of locality-conscious distribution on 16 nodes, outperforming and significantly outscaling the best-known locality-conscious server. Based on our results and on the fact that the files serviced by network servers are becoming larger and more numerous, we conclude that our locality-conscious network server should prove very useful for its performance, scalability, and availability properties.