Dynamic load balancing on heterogeneous multi-GPU systems

Actual HPC systems are composed by multicore processors and powerful graphics processing units. Adapting existing code and libraries to these new systems is a fundamental problem due to the important increment on programming difficulties. The heterogeneity, both at architectural and programming levels at the same time, raises the programmability wall. The performance of the code is affected by the large interdependence between the code and the parallel architecture. We have developed a dynamic load balancing library that allows parallel code to be adapted to a wide variety of heterogeneous systems. The overhead introduced by our system is minimal and the cost to the programmer negligible. This system has been successfully applied to solve load imbalance problems appearing in homogeneous and heterogeneous multiGPU platforms. We consider the Dynamic Programming technique as case of study to validate our proposals using different heterogeneous scenarios in multiGPU systems.     

A novel dynamic load balancing scheme for parallel systems

Adaptive mesh refinement (AMR) is a type of multiscale algorithm that achieves high resolution in localized regions of dynamic, multidimensional numerical simulations. One of the key issues related to AMR is dynamic load balancing (DLB), which allows large-scale adaptive applications to run efficiently on parallel systems. In this paper, we present an efficient DLB scheme for structured AMR (SAMR) applications. This scheme interleaves a grid-splitting technique with direct grid movements (e.g., direct movement from an overloaded processor to an underloaded processor), for which the objective is to efficiently redistribute workload among all the processors so as to reduce the parallel execution time. The potential benefits of our DLB scheme are examined by incorporating our techniques into a SAMR cosmology application, the ENZO code. Experiments show that by using our scheme, the parallel execution time can be reduced by up to 57% and the quality of load balancing can be improved by a factor of six, as compared to the original DLB scheme used in ENZO.     

Distributed algorithms for QoS load balancing

We consider a dynamic load balancing scenario in which users allocate resources in a non-cooperative and selfish fashion. The perceived performance of a resource for a user decreases with the number of users that allocate the resource. In our dynamic, concurrent model, users may reallocate resources in a round-based fashion. As opposed to various settings analyzed in the literature, we assume that users have quality of service demands. A user has zero utility when falling short of a certain minimum performance threshold and having positive utility otherwise. Whereas various load-balancing protocols have been proposed for the setting without quality of service requirements, we consider protocols that satisfy an additional locality constraint: The behavior of a user depends merely on the state of the resource it currently allocates. This property is particularly useful in scenarios where the state of other resources is not readily accessible. For instance, if resources represent channels in a mobile network, then accessing channel information may require time-intensive measurements. We consider several variants of the model, where the quality of service demands may depend on the user, the resource, or both. For all cases we present protocols for which the dynamics converge to a state in which all users are satisfied. More importantly, the time to reach such a state scales nicely. It is only logarithmic in the number of users, which makes our protocols applicable in large-scale systems.     

HeDPM: load balancing of linear pipeline applications on heterogeneous systems

This work presents a new algorithm, called Heterogeneous Dynamic Pipeline Mapping, that allows for dynamically improving the performance of pipeline applications running on heterogeneous systems. It is aimed at balancing the application load by determining the best replication (of slow stages) and gathering (of fast stages) combination taking into account processors computation and communication capacities. In addition, the algorithm has been designed with the requirement of keeping complexity low to allow its usage in a dynamic tuning tool. For this reason, it uses an analytical performance model of pipeline applications that addresses hardware heterogeneity and which depends on parameters that can be known in advance or measured at run-time. A wide experimentation is presented, including the comparison with the optimal brute force algorithm, a general comparison with the Binary Search Closest algorithm, and an application example with the Ferret pipeline included in the PARSEC benchmark suite. Results, matching those of the best existing algorithms, show significant performance improvements with lower complexity (\(O(N^3\)), where N is the number of pipeline stages).     

A load index and load balancing algorithm for heterogeneous clusters

This paper presents a load balancing algorithm specifically designed for heterogeneous clusters, composed of nodes with different computational capabilities. The method is based on a new index, which takes into consideration two levels of processors heterogeneity: the number of cores per node and the computational power of each core. The experimental results show that this index allows achieving balanced workload distributions even on those clusters where heterogeneity can not be neglected.     

Disk load balancing for video-on-demand systems

For a video-on-demand computer system, we propose a scheme which balances the load on the disks, thereby helping to solve a performance problem crucial to achieving maximal video throughput. Our load-balancing scheme consists of two components. The static component determines good assignments of videos to groups of striped disks. The dynamic component uses these assignments, and features a “DASD dancing” algorithm which performs real-time disk scheduling in an effective manner. Our scheme works synergistically with disk striping. We examine the performance of the proposed algorithm via simulation experiments.     

Distributed strategic interleaving with load balancing

In a previous paper, we developed an algebraic theory of threads, interleaving of threads, and interaction of threads with services. In the current paper, we assume that the threads and services are distributed over the nodes of a network. We extend the theory developed so far to the distributed case by introducing distributed interleaving strategies that support explicit thread migration and see to load balancing or capability searching by implicit thread migration. The extension to the distributed case provides insight into details of multi-threading that come up in a networked environment.     

Accelerated diffusion algorithms for dynamic load balancing

In this paper we consider the application of accelerated techniques in order to increase the rate of convergence of the diffusive iterative load balancing algorithms. In particular, we compare the application of Semi-Iterative, Second Degree and Variable Extrapolation techniques on the basic diffusion method for various types of network graphs.     

Cooperative distributed dynamic load balancing

Summary This paper explores and applies the concept of cooperation to the load balancing problem in a computer network. We discuss an analytical model and propose a scheme which can be classified as distributed, dynamic, and stochastic. In the case of a homogeneous network, we guarantee that the load is balanced and no communication cost or information exchange is necessary.     

Dynamic load balancing on Web-server systems

Popular Web sites cannot rely on a single powerful server nor on independent mirrored-servers to support the ever-increasing request load. Distributed Web server architectures that transparently schedule client requests offer a way to meet dynamic scalability and availability requirements. The authors review the state of the art in load balancing techniques on distributed Web-server systems, and analyze the efficiencies and limitations of the various approaches     

Strategies for dynamic load balancing on highly parallel computers

Dynamic load balancing strategies for minimizing the execution time of single applications running in parallel on multicomputer systems are discussed. Dynamic load balancing (DLB) is essential for the efficient use of highly parallel systems when solving non-uniform problems with unpredictable load estimates. With the evolution of more highly parallel systems, centralized DLB approaches which make use of a high degree of knowledge become less feasible due to the load balancing communication overhead. Five DLB strategies are presented which illustrate the tradeoff between 1) knowledge - the accuracy of each balancing decision, and 2) overhead - the amount of added processing and communication incurred by the balancing process. All five strategies have been implemented on an Inter iPSC/2 hypercube     

LTE系统中的分布式负载迁移协商均衡算法

为了提高长期演进计划(long term evolution,LTE)系统中业务处理资源的利用率,降低由于系统处理资源分配不均而导致的呼损率,给出了一种适用于LTE系统的分布式负载迁移协商均衡算法.分析了LTE系统通信的特点,指出可利用LTE系统eNodeB (evolved NodeB)间使用X2接口进行数据通信的特点,在组网的eNodeB问进行负载迁移.在此基础上建立了负载迁移协商均衡模型,指出了算法的适用场景,说明了算法均衡粒度的选取标准和性能评价指标.通过与传统的轮询算法进行比较,使用分布式事件驱动方法对该算法进行了性能分析,验证结果表明了该算法的有效性.     

Semi-distributed load balancing for massively parallelmulticomputer systems

A semidistributed approach is given for load balancing in large parallel and distributed systems which is different from the conventional centralized and fully distributed approaches. The proposed strategy uses a two-level hierarchical control by partitioning the interconnection structure of a distributed or multiprocessor system into independent symmetric regions (spheres) centered at some control points. The central points, called schedulers, optimally schedule tasks within their spheres and maintain state information with low overhead. The authors consider interconnection structures belonging to a number of families of distance transitive graphs for evaluation, and, using their algebraic characteristics, show that identification of spheres and their scheduling points is in general an NP-complete problem. An efficient solution for this problem is presented by making exclusive use of a combinatorial structure known as the Hadamard matrix. The performance of the proposed strategy has been evaluated and compared with an efficient fully distributed strategy through an extensive simulation study. The proposed strategy yielded much better results     

Game-theoretic static load balancing for distributed systems

In this paper, we present a game theoretic approach to solve the static load balancing problem for single-class and multi-class (multi-user) jobs in a distributed system where the computers are connected by a communication network. The objective of our approach is to provide fairness to all the jobs (in a single-class system) and the users of the jobs (in a multi-user system). To provide fairness to all the jobs in the system, we use a cooperative game to model the load balancing problem. Our solution is based on the Nash Bargaining Solution (NBS) which provides a Pareto optimal solution for the distributed system and is also a fair solution. An algorithm for computing the NBS is derived for the proposed cooperative load balancing game. To provide fairness to all the users in the system, the load balancing problem is formulated as a non-cooperative game among the users who try to minimize the expected response time of their own jobs. We use the concept of Nash equilibrium as the solution of our non-cooperative game and derive a distributed algorithm for computing it. Our schemes are compared with other existing schemes using simulations with various system loads and configurations. We show that our schemes perform near the system optimal schemes and are superior to the other schemes in terms of fairness.     

Distributed load balancing for parallel main memory hash join

Parallel joins have been widely studied during the past decade and a number of efficient algorithms were presented. While it is known that the performance of these algorithms may suffer greatly in the presence of skewed input data, the work on load balancing schemes for parallel join has been limited. The main contribution of this paper is the development and analysis of a new distributed data structure and an effective load balancing scheme for parallel main memory hash join on NUMA architecture. Multiprocessors based on this architecture are scalable in both size of main memory and number of processors, and provide very high memory bandwidth. The load balancing scheme is based on random probing to avoid the hot spot problems caused by probing sequentially. We have modeled this load balancing scheme both analytically and experimentally. The experiments were run on a BBN TC2000 multiprocessor system     

A repartitioning hypergraph model for dynamic load balancing

In parallel adaptive applications, the computational structure of the applications changes over time, leading to load imbalances even though the initial load distributions were balanced. To restore balance and to keep communication volume low in further iterations of the applications, dynamic load balancing (repartitioning) of the changed computational structure is required. Repartitioning differs from static load balancing (partitioning) due to the additional requirement of minimizing migration cost to move data from an existing partition to a new partition. In this paper, we present a novel repartitioning hypergraph model for dynamic load balancing that accounts for both communication volume in the application and migration cost to move data, in order to minimize the overall cost. The use of a hypergraph-based model allows us to accurately model communication costs rather than approximate them with graph-based models. We show that the new model can be realized using hypergraph partitioning with fixed vertices and describe our parallel multilevel implementation within the Zoltan load balancing toolkit. To the best of our knowledge, this is the first implementation for dynamic load balancing based on hypergraph partitioning. To demonstrate the effectiveness of our approach, we conducted experiments on a Linux cluster with 1024 processors. The results show that, in terms of reducing total cost, our new model compares favorably to the graph-based dynamic load balancing approaches, and multilevel approaches improve the repartitioning quality significantly.     

An optimal migration algorithm for dynamic load balancing

The problem of redistributing the work load on parallel computers is considered. An optimal redistribution algorithm, which minimises the Euclidean norm of the migrating load, is derived. The relationship between this algorithm and some existing algorithms is discussed and the convergence of the new algorithm is studied. Finally, numerical results on randomly generated graphs as well as on graphs related to real meshes are given to demonstrate the effectiveness of the new algorithm. © 1998 John Wiley & Sons, Ltd.     

一种动态的客户端负载均衡机制

互联网应用的并发用户数量不仅多变并且这种变化常常不可预测.将系统容量配置为固定值的惯用做法在面临多变的请求时常常会导致两种结果,一种是因为配置过低而引起用户的不满,另一种则因为配置过高而造成计算资源的浪费.而通过运用云架构按需提供、按使用收费的能力,系统具有了实时地按需配置计算资源的能力.然而静态的客户端负载均衡方法作为一种主要的负载均衡技术很难适应云架构条件下更加易变的集群结构.本文提出了一种动态的客户端负载均衡机制.通过引入分布式的集群视图更新、控制流等技术,该机制在保持客户端负载均衡机制分布式、可伸缩性强的基础上,又为集群节点的动态加入与退出和负载均衡策略的动态调整提供了有效的支持.同时,本文分析了该机制在一个开源JEE应用服务器PKUAS中的关键实现问题,并通过实验从多方面验证了该机制的有效性.     

一种改进的基于动态反馈的负载均衡算法

负载均衡是集群系统研究的一个重要问题,负载均衡算法是集群任务分配的核心,介绍了LVS中的负载均衡算法,讨论了常用算法的不足,在分析这些算法各自优缺点的基础上,提出了一种改进的基于反馈的负载均衡算法,算法引入一个负载容余参数以更准确地描述集群节点的负载状况,在考虑服务节点真实负载,处理能力的基础上,尽量简化负载均衡器的任务分配算法.测试结果显示该算法优于静态算法.     

A dynamic load balancing scheme with incentive mechanism in heterogeneous structured P2P networks

Dynamic load imbalance is a basic and inherent problem in structured P2P networks. Most existing research suffers from the problems of inefficiency in globally managing the nodes’ load information and consumption of network bandwidth. This paper describes the mechanisms for collecting and globally managing the dynamic load of each node, and based on which to present a load balancing strategy which transfers the load from overloaded to under loaded nodes so as to improve load balancing efficiency. In order to encourage the rational and selfish nodes to actively participate in the load balancing process, we also propose an incentive mechanism in dynamic load balancing, by which the differentiated services could be provided for the nodes according to their load balancing abilities. The simulation results indicate that our approach could tackle the load imbalance problem in structured P2P networks effectively and efficiently in terms of the load distribution and the transferred load volume.