Scheduling concurrent applications on a cluster of CPU–GPU nodes

Heterogeneous architectures comprising a multi-core CPU and many-core GPU(s) are increasingly being used within cluster and cloud environments. In this paper, we study the problem of optimizing the overall throughput of a set of applications deployed on a cluster of such heterogeneous nodes. We consider two different scheduling formulations. In the first formulation, we consider jobs that can be executed on either the GPU or the CPU of a single node. In the second formulation, we consider jobs that can be executed on the CPU, GPU, or both, of any number of nodes in the system. We have developed scheduling schemes addressing both of the problems. In our evaluation, we first show that the schemes proposed for first formulation outperform a blind round-robin scheduler and approximate the performances of an ideal scheduler that involves an impractical exhaustive exploration of all possible schedules. Next, we show that the scheme proposed for the second formulation outperforms the best of existing schemes for heterogeneous clusters, TORQUE and MCT, by up to 42%. Additionally, we evaluate the robustness of our proposed scheduling policies under inaccurate inputs to account for real execution scenarios. We show that, with up to 20% of inaccuracy in the input, the degradation in performance is marginal (less than 7%) on the average.     

Distributed multi-join query processing in data grids

Query processing in data grids is a difficult issue due to the heterogeneous, unpredictable and volatile behaviors of the grid resources. Applying join operations on remote relations in data grids is a unique and interesting problem. However, to the best of our knowledge, little is done to date on multi-join query processing in data grids. An approach for processing multi-join queries is proposed in this paper. Firstly, a relation-reduction algorithm for reducing the sizes of operand relations is presented in order to minimize data transmission cost among grid nodes. Then, a method for scheduling computer nodes in data grids is devised to parallel process multi-join queries. Thirdly, an innovative method is developed to efficiently execute join operations in a pipeline fashion. Finally, a complete algorithm for processing multi-join queries is given. Analytical and experimental results show the effectiveness and efficiency of the proposed approach.     

Collection trees for event-monitoring queries

In this paper we present algorithms for building and maintaining efficient collection trees that provide the conduit to disseminate data required for processing monitoring queries in a wireless sensor network. While prior techniques base their operation on the assumption that the sensor nodes that collect data relevant to a specified query need to include their measurements in the query result at every query epoch, in many event monitoring applications such an assumption is not valid. We introduce and formalize the notion of event monitoring queries and demonstrate that they can capture a large class of monitoring applications. We then show techniques which, using a small set of intuitive statistics, can compute collection trees that minimize important resources such as the number of messages exchanged among the nodes or the overall energy consumption. Our experiments demonstrate that our techniques can organize the data collection process while utilizing significantly lower resources than prior approaches.     

基于改进秃鹰搜索算法的Kubernetes资源调度应用

Kubernetes是一个管理容器化应用的开源平台,其默认的调度算法在优选阶段仅把CPU和内存两种资源来作为计算节点的评分指标,同时还忽略了不同类型的Pod对节点资源的占用比例是不同的,容易导致某一资源达到性能瓶颈,从而造成节点对资源使用失衡.针对上述问题,本文在Kubernetes原有的资源指标基础上增加了带宽和磁盘容量,考虑到CPU、内存、带宽和磁盘容量这4类资源在节点上的占用比例对节点的性能的影响,可能造成Pod中应用的非正常运行,甚至杀死Pod,从而影响集群整体的高可靠性.本文将等待创建的Pod区分为可压缩消耗型、不可压缩消耗型以及均衡型,并为每种类型的Pod设置相应的权重,最后通过改进的秃鹰搜索算法(TBESK)来寻找出最优节点进行调度.实验结果表明,随着集群中Pod的数量在不断增加,在集群负载较大的情况下, TBESK算法的综合负载标准差和默认的调度算法相比提升了24%.     

分时EDF算法及其在多媒体操作系统中的应用

提出了一种新的CPU调度算法－－分时EDF（Earliest Deadine First)算法,该算法能保证硬实时任务不丢失死线,并易于在分时系统中实现。以分时EDF算法为基础,提出一种新的CPU层次调度算法－－HRFSFQ,该算法用于多媒体操作系统时能保证各类任务的QoS。最后通过大量实验证明了上述算法的有效性和正确性。     

Kubernetes异构资源细粒度调度策略的设计与实现

在异构资源环境中高效利用计算资源是提升任务效率和集群利用率的关键。Kuberentes作为容器编排领域的首选方案,在异构资源调度场景下调度器缺少GPU细粒度信息无法满足用户自定义需求,并且CPU/GPU节点混合部署下调度器无法感知异构资源从而导致资源竞争。综合考虑异构资源在节点上的分布及其硬件状态,提出一种基于Kubernetes的CPU/GPU异构资源细粒度调度策略。利用设备插件机制收集每个节点上GPU的详细信息,并将GPU资源指标提交给调度算法。在原有CPU和内存过滤算法的基础上,增加自定义GPU信息的过滤,从而筛选出符合用户细粒度需求的节点。针对CPU/GPU节点混合部署的情况,改进调度器的打分算法,动态感知应用类型,对CPU和GPU应用分别采用负载均衡算法和最小最合适算法,保证异构资源调度策略对不同类型应用的正确调度,并且在CPU资源不足的情况下充分利用GPU节点的碎片资源。通过对GPU细粒度调度和CPU/GPU节点混合部署情况下的调度效果进行实验验证,结果表明该策略能够有效进行GPU调度并且避免资源竞争。     

PostgreSQL数据库在不同调节器中的能效比较

近年来,能效数据库系统成为数据库领域的一个研究议题.CPU动态电压频率调节(DVFS)是一种有效的动态功率节能技术.探寻PostgreSQL数据库在ACPI不同调节器下查询操作的性能、能耗、功率之间潜在联系,发现动态功耗管理与数据库系统的能效关系,通过运行TPC-H测试基准生成的数据库与相应22个查询,总结出调节器对数据库查询处理各种操作的影响.实验结果表明,DVFS可以对DBMS进行动态功耗管理是有效的,查询处理的不同操作具有各自特性,利用这些特性来设计效率更高的调节器是颇有前途的.     

Multiple similarity queries: a basic DBMS operation for mining inmetric databases

Metric databases are databases where a metric distance function is defined for pairs of database objects. In such databases, similarity queries in the form of range queries or k-nearest-neighbor queries are the most important query types. In traditional query processing, single queries are issued independently by different users. In many data mining applications, however, the database is typically explored by iteratively asking similarity queries for answers of previous similarity queries. We introduce a generic scheme for such data mining algorithms and we investigate two orthogonal approaches, reducing I/O cost as well as CPU cost, to speed-up the processing of multiple similarity queries. The proposed techniques apply to any type of similarity query and to an implementation based on an index or using a sequential scan. Parallelization yields an additional impressive speed-up. An extensive performance evaluation confirms the efficiency of our approach     

Large matrix–vector products on distributed bus networks with communication delays using the divisible load paradigm: performance analysis and simulation

We present a performance analysis and experimental simulation results on the problem of scheduling a divisible load on a bus network. In general, the computing requirement of a divisible load is CPU intensive and demands multiple processing nodes for efficient processing. We consider the problem of scheduling a very large matrix–vector product computation on a bus network consisting of a homogeneous set of processors. The experiment was conducted on a PC-based networking environment consisting of Pentium II machines arranged in a bus topology. We present a theoretical analysis and verify these findings on the experimental test-bed. We also developed a software support system with flexibility in terms of scalability of the network and the load size. We present a detailed discussion on the experimental results providing directions for possible future extensions of this work.     

Parallel-machine scheduling to minimize tardiness penalty and power cost

Traditional research on machine scheduling focuses on job allocation and sequencing to optimize certain objective functions that are defined in terms of job completion times. With regard to environmental concerns, energy consumption becomes another critical issue in high-performance systems. This paper addresses a scheduling problem in a multiple-machine system where the computing speeds of the machines are allowed to be adjusted during the course of execution. The CPU adjustment capability enables the flexibility for minimizing electricity cost from the energy saving aspect by sacrificing job completion times. The decision of the studied problem is to dispatch the jobs to the machines as well as to determine the job sequence and processing speed of each machine with the objective function comprising of the total weighted job tardiness and the power cost. We give a formal formulation, propose two heuristic algorithms, and develop a particle swarm optimization (PSO) algorithm to effectively tackle the problem. Since the existing solution representations do not befittingly encode the decisions involved in the studied problem into the PSO algorithm, we design a tailored encoding scheme which can embed all decisional information in a particle. A computational study is conducted to investigate the performances of the proposed heuristics and the PSO algorithm.     

移动云计算环境下多工作流任务调度的联合优化方法

传统移动云计算环境下的任务调度通过random算法来决定任务执行位置,通过动态电压调节技术来调节工作频率,通过任务间的差异性判别进行任务的整合,这往往带来了很多不合理的任务迁移,并导致CPU负载严重,造成了系统损害和大量能耗。针对多工作流任务提出了CCS算法,它包括consolidation算法与多任务并发算法,通过增加任务之间传输与执行的并发性,增加任务集整合的概率,提高任务的处理速率,减少任务的响应时间,增加CPU使用率的同时将主机和内核CPU使用率控制在阈值上限以下,避免CPU过载并根据多任务并发来优化local算法,调整任务执行位置,提高迁移效率的同时也避免了随机算法的局限性,实验结果表明该算法可以有效地提高系统性能,避免CPU过载问题,并且优化了能耗和工作流的完成时间。     

UpStream: storage-centric load management for streaming applications with update semantics

This paper addresses the problem of minimizing the staleness of query results for streaming applications with update semantics under overload conditions. Staleness is a measure of how out-of-date the results are compared with the latest data arriving on the input. Real-time streaming applications are subject to overload due to unpredictably increasing data rates, while in many of them, we observe that data streams and queries in fact exhibit “update semantics” (i.e., the latest input data are all that really matters when producing a query result). Under such semantics, overload will cause staleness to build up. The key to avoid this is to exploit the update semantics of applications as early as possible in the processing pipeline. In this paper, we propose UpStream, a storage-centric framework for load management over streaming applications with update semantics. We first describe how we model streams and queries that possess the update semantics, providing definitions for correctness and staleness for the query results. Then, we show how staleness can be minimized based on intelligent update key scheduling techniques applied at the queue level, while preserving the correctness of the results, even for complex queries that involve sliding windows. UpStream is based on the simple idea of applying the updates in place, yet with great returns in terms of lowering staleness and memory consumption, as we also experimentally verify on the Borealis system.     

Energy consumption analysis of data stream processing: a benchmarking approach

Energy efficiency of data analysis systems has become a very important issue in recent times because of the increasing costs of data center operations. Although distributed streaming workloads have increasingly been present in modern data centers, energy‐efficient scheduling of such applications remains as a significant challenge. In this paper, we conduct an energy consumption analysis of data stream processing systems in order to identify their energy consumption patterns. We follow stream system benchmarking approach to solve this issue. Specifically, we implement Linear Road benchmark on six stream processing environments (S4, Storm, ActiveMQ, Esper, Kafka, and Spark Streaming) and characterize these systems' performance on a real‐world data center. We study the energy consumption characteristics of each system with varying number of roads as well as with different types of component layouts. We also use a microbenchmark to capture raw energy consumption characteristics. We observed that S4, Esper, and Spark Streaming environments had highest average energy consumption efficiencies compared with the other systems. Using a neural networkbased technique with the power/performance information gathered from our experiments, we developed a model for the power consumption behavior of a streaming environment. We observed that energy‐efficient execution of streaming application cannot be specifically attributed to the system CPU usage. We observed that communication between compute nodes with moderate tuple sizes and scheduling plans with balanced system overhead produces better power consumption behaviors in the context of data stream processing systems. Copyright © 2016 John Wiley & Sons, Ltd.     

Energy management in solar cells powered wireless sensor networks for quality of service optimization

Sensor nodes equipped with solar cells and rechargeable batteries are useful in many outdoor, long-lasting applications. In these sensor nodes, the cycles of energy harvesting and battery recharge need to be managed appropriately in order to avoid sensor node unavailability due to energy shortages. To this purpose, we suggest sensor nodes to be programmed with alternative scheduling plans, each corresponding to a given energy requirement and meeting a given quality level. Thus, sensor nodes can select the scheduling plan that best suits to the expected energy production and the residual battery charge, in order to avoid sensor nodes' unavailability. We then propose an algorithm for the selection of the scheduling plan that aims at keeping the overall energy consumption neutral and that can sustain the sensor nodes’ activities uninterruptedly.     

Minimizing write operation for multi-dimensional DSP applications via a two-level partition technique with complete memory latency hiding

Most scientific and digital signal processing (DSP) applications are recursive or iterative. The execution of these applications on a chip multiprocessor (CMP) encounters two challenges. First, as most of the digital signal processing applications are both computation intensive and data intensive, an inefficient scheduling scheme may generate huge amount of write operation, cost a lot of time, and consume significant amount of energy. Second, because CPU speed has been increased dramatically compared with memory speed, the slowness of memory hinders the overall system performance. In this paper, we develop a Two-Level Partition (TLP) algorithm that can minimize write operation while achieving full parallelism for multi-dimensional DSP applications running on CMPs which employ scratchpad memory (SPM) as on-chip memory (e.g., the IBM Cell processor). Experiments on DSP benchmarks demonstrate the effectiveness and efficiency of the TLP algorithm, namely, the TLP algorithm can completely hide memory latencies to achieve full parallelism and generate the least amount of write operation to main memory compared with previous approaches. Experimental results show that our proposed algorithm is superior to all known methods, including the list scheduling, rotation scheduling, Partition Scheduling with Prefetching (PSP), and Iterational Retiming with Partitioning (IRP) algorithms. Furthermore, the TLP scheduling algorithm can reduce write operation to main memory by 45.35% and reduce the schedule length by 23.7% on average compared with the IRP scheduling algorithm, the best known algorithm.     

Multiple query scheduling for distributed semantic caches

In distributed query processing systems, load balancing plays an important role in maximizing system throughput. When queries can leverage cached intermediate results, improving the cache hit ratio becomes as important as load balancing in query scheduling, especially when dealing with computationally expensive queries. The scheduling policies must be designed to take into consideration the dynamic contents of the distributed caching infrastructure. In this paper, we propose and discuss several distributed query scheduling policies that directly consider the available cache contents by employing distributed multidimensional indexing structures and an exponential moving average approach to predicting cache contents. These approaches are shown to produce better query plans and faster query response times than traditional scheduling policies that do not predict dynamic contents in distributed caches. We experimentally demonstrate the utility of the scheduling policies using MQO, which is a distributed, Grid-enabled, multiple query processing middleware system we developed to optimize query processing for data analysis and visualization applications.     

Adaptive processing of historical spatial range queries in peer-to-peer sensor networks

We investigate the problem of processing historical queries on a sensor network. Since data is considered to have been already collected at the sensor nodes, the main issue is exploring the spatial component of the query in order to minimize its cost represented by the energy consumption. We assume queries can be issued at any network node, i.e., there is no central base station and all nodes have only local knowledge of the network. On the one hand, a globally optimum query processing plan is desirable but its construction is not possible due to the lack of global knowledge of the network. On the other hand, while a simple network flooding is feasible, it is not a practical choice from a cost perspective. To address this problem we propose a two-phase query processing strategy, where in the first phase a path from the query originator to the query region is found and in the second phase the query is processed within the query region itself. This strategy is supported by analytical models that are used to dynamically select the best processing strategy depending on the query specifics. Our extensive analytical and experimental results show that our analytical models are accurate and that the two-phase strategy is better suited for small to medium sized queries, being up to 10 times more cost effective than a typical network flooding. In addition, the dynamic selection of a query processing technique proved itself capable of always delivering at least as good performance as the most energy efficient strategy for all query sizes. Research supported in part by NSERC Canada.     

面向边缘计算的Storm边缘节点调度优化方法

边缘计算有高实时性和大数据交互处理的需求,边缘异构节点间的调度时耗长、通信时延高以及负载不均衡是影响边缘计算性能的核心问题,传统的云计算平台难以满足新的要求。文中研究了在边缘计算环境下Storm边缘节点的调度优化方法,建立了面向边缘计算的Storm任务卸载调度模型。针对拓扑任务在边缘异构节点间的实时动态分配问题,提出了一种启发式动态规划算法(Inspire Dynamic Programming,IDP),通过改变Storm的Task实例的排序分配方式以及Task实例和Slot任务槽的映射关系实现全局的优化调度;同时,针对拓扑任务的并发度受限于JVM栈深度的缺陷,提出了一种基于蝙蝠算法的调度策略。实验结果表明,与Storm调度算法相比,所提算法在边缘节点CPU利用率指标上平均提升了约60%,在集群的吞吐量指标上平均提升了约8.2%,因此能够满足边缘节点之间的高实时性处理要求。     

Range nearest-neighbor query

A range nearest-neighbor (RNN) query retrieves the nearest neighbor (NN) for every point in a range. It is a natural generalization of point and continuous nearest-neighbor queries and has many applications. In this paper, we consider the ranges as (hyper)rectangles and propose efficient in-memory processing and secondary memory pruning techniques for RNN queries in both 2D and high-dimensional spaces. These techniques are generalized for kRNN queries, which return the k nearest neighbors for every point in the range. In addition, we devise an auxiliary solution-based index EXO-tree to speed up any type of NN query. EXO-tree is orthogonal to any existing NN processing algorithm and, thus, can be transparently integrated. An extensive empirical study was conducted to evaluate the CPU and I/O performance of these techniques, and the study showed that they are efficient and robust under various data sets, query ranges, numbers of nearest neighbors, dimensions, and cache sizes.