2015年BPOE研讨会：大数据基准、性能优化与新兴硬件论文集

需积分: 9 165 浏览量更新于2024-07-20 收藏 10.26MB PDF 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

《大规模数据基准、性能优化与新兴硬件：第六届研讨会论文集》（BPOE 2015）是一本收录了2015年8月31日至9月4日在夏威夷科哈拉海岸举行的VLDB 2015国际大会卫星活动——第六届大数据基准、性能优化与新兴硬件研讨会（BPOE 2015）的精选论文。该书共包含8篇经过精心评审的论文，旨在促进数据管理、架构和系统研究社区之间的交流，探讨这些领域交叉中的研究问题。书中的内容涵盖四个方面： 1. **Benchmarking**（基准测试）：第1章讨论了大数据基准测试的原则和方法学，重新审视并提出适应大型数据环境的标准化实践。第2章介绍了BigDataBench-MT工具，这是一个用于生成现实混合数据中心工作负载的基准工具，有助于评估系统的性能和可扩展性。 2. **Benchmarking and Workload Characterization**（基准测试与工作负载分析）：第3章探讨了为在线社交网络时代设计的带有真实工作负载和实时数据的大数据基准套件，以及第4章分析了现实生活中知识图谱的统计特性。第5章则关注多核操作系统Mbench的基准测试，通过混合工作负载来衡量系统的性能。 3. **Performance Optimization and Evaluation**（性能优化与评估）：在这一部分，第6章分析了Shark到Spark SQL的演进，并进行初步评估；第7章探讨了数据量如何影响规模扩展服务器上的Spark数据分析；第8章提出了一种基于采样数据倾斜的最优reduce放置算法；第9章AAA方法介绍了一个针对大规模系统监控的大数据采集策略。 4. **Emerging Hardware**（新兴硬件）：最后一部分聚焦于硬件创新，第10章介绍了一种基于插件的策略，以利用RDMA（远程直接内存访问）技术提升Apache和企业级HDFS的性能；第11章则提出了一个流式无损数据压缩硬件方案，通过自适应频率表管理提高效率。这本书不仅提供了实用的工具和方法，还反映了当时大数据领域内关于性能优化、工作负载理解和硬件发展的前沿研究，对于理解和支持大规模数据处理系统的开发和优化具有重要的参考价值。

资源详情

资源推荐

Revisiting Benchmarking Principles and Methodologies 7

3.1 TPC-C Methodology

TPC-C is an on-line transaction processing (OLTP) benchmark with multiple

transaction types, complex database and overall execution structure. The con-

structing of TPC-C introduces the concept of abstractions – functions of abstrac-

tion and functional workload model [6].

Functions of Abstraction. Through investigating multiple transactions in

OLTP application domain, TPC-C concentrates on ﬁnding the compute units

that frequently appeared with repetitions or similarities, such as duplicate opera-

tions and processing ﬂow on diﬀerent tables. After that, TPC-C ﬁnally abstracts

ﬁve units of computation within an order-entry environment, with diﬀerent pro-

portion of read and write.

Functional Workload Model. Functional workload model includes functions

of abstraction, data set and load pattern. Among which, data set involves data

structure, inter-dependence and content. Load pattern involves execution fre-

quency, distribution and arrival rate of each function of abstraction. That is to

say, combining ﬁve primitive transactions, with specifying scheduling mode and

deﬁning data structure, a functional workload model can be formed.

3.2 Methodology of BigDataBench

Following the above ﬁve principles, BigDataBench [13] takes an incremental and

iterative approach as illustrated in Fig. 1.

First, it investigates important or emerging application domains according to

widely acceptable metrics. Second, it understands selected application domains

from the perspectives of data sets and workloads [14], including data models

of diﬀerent types, i.e. structured, un-/semi-structured data sets and semantics,

i.e. text, table, graph, image and workload dwarfs of units of computation that

frequently appear in big data analytics [7].

Fig. 1. Methodology of BigDataBench.

8L.Zhaoetal.

Third, it deﬁnes benchmark speciﬁcations [10] for each application domain to

guide its implementations from aspects of data model description and impor-

tant workloads extraction. Fourth, three aspects are considered to implement

BigDataBench: representative and diverse real-world data sets, big data gener-

ator suite which preserve characteristics of real-world data [11], and important

workloads implementation with competitive techniques.

Finally, it provides two versions for speciﬁc requirements and purposes: multi-

tenancy version with diﬀerent percentages of service and analytics workloads,

and BigDataBench subset with a small number of representative workloads

according to workload characteristics from a speciﬁc perspective.

4 Conclusion

Big data benchmarking is important and challenging. Revisiting successful bench-

marks in other domains, we ﬁnd that most of them specify benchmark principles

according to domain-oriented benchmarking requirements, and deﬁne methodolo-

gies accord with principles. In terms of big data domain, the benchmarking prin-

ciples should take its complexity and diversity into consideration, moreover, the

methodologies should be capable to keep pace with the rapid evolution of big data

domains.

Acknowledgements. This work is supported by the National High Technology

Research and Development Program of China (Grant No. 2015AA015308), the Major

Program of National Natural Science Foundation of China (Grant No. 61432006), and

the Key Technology Research and Development Programs of Guangdong Province,

China (Grant No. 2015B010108006).

References

1. Bdbc. http://clds.sdsc.edu/bdbc

2. Spec. https://www.spec.org/

3. Tpc. http://www.tpc.org

4. Angles, R.: Benchmark principles and methods. In: Linked Data Benchmark Coun-

cil (LDBC). Project No 317548, European Community’s Seventh Framework Pro-

gramme FP7 (2012–2014)

5. Bienia, C., Li, K.: Benchmarking Modern Multiprocessors. Princeton University,

New York (2011)

6. Chen, Y., Raab, F., Katz, R.: From TPC-C to big data benchmarks: a functional

workload model. In: Rabl, T., Poess, M., Baru, C., Jacobsen, H.-A. (eds.) WBDB

2012. LNCS, vol. 8163, pp. 28–43. Springer, Heidelberg (2014)

7. Gao, W., Luo, C., Zhan, J., Ye, H., He, X., Wang, L., Zhu, Y., Tian, X.: Identifying

dwarfs workloads in big data analytics (2015). arXiv preprint arXiv:1505.06872

8. Gray, J.: Benchmark Handbook: For Database and Transaction Processing Sys-

tems. Morgan Kaufmann Publishers Inc., San Francisco (1992)

9. Levine, C.: TPC benchmarks. In: SIGMOD International Conference on Manage-

mentof Data - Industrial Session (1997)

BigDataBench-MT: A Benchmark Tool

for Generating Realistic Mixed Data Center

Workloads

Rui Han

)

, Shulin Zhan

, Chenrong Shao

,JunweiWang

Lizy K. John

, Jiangtao Xu

,GangLu

1,2

,andLeiWang

Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China

{hanrui,xujiangtao,lugang,wl}@ict.ac.cn

University of Chinese Academy of Sciences, Beijing, China

ICarsclub, Beijing, China

zhanshulin@ppzuche.com

Xi’an Jiaotong University, Xi’an, China

scr1994@stu.xjtu.edu.cn

Kingsoft Cloud, Beijing, China

wangjunwei@kingsoft.com

Department of Electrical and Computer Engineering,

TheUniversityofTexas,Austin,TX,USA

ljohn@ece.utexas.edu

Abstract. Long-running service workloads (e.g. web search engine) and

short-term data analysis workloads (e.g. Hadoop MapReduce jobs) co-

locate in today’s data centers. Developing realistic benchmarks to reﬂect

such practical scenario of mixed workload is a key problem to pro-

duce trustworthy results when evaluating and comparing data center

systems. This requires using actual workloads as well as guaranteeing

their submissions to follow patterns hidden in real-world traces. How-

ever, existing benchmarks either generate actual workloads based on

probability models, or replay real-world workload traces using basic I/O

operations. To ﬁll this gap, we propose a benchmark tool that is a ﬁrst

step towards generating a mix of actual service and data analysis work-

loads on the basis of real workload traces. Our tool includes a combiner

that enables the replaying of actual workloads according to the workload

traces, and a multi-tenant generator that ﬂexibly scales the workloads

up and down according to users’ requirements. Based on this, our demo

illustrates the workload customization and generation process using a

visual interface. The proposed tool, called BigDataBench-MT, is a multi-

tenant version of our comprehensive benchmark suite BigDataBench

and it is publicly available from http://prof.ict.ac.cn/BigDataBench/

multi-tenancyversion/.

Keywords: Data center

· Benchmark · Workload trace · Mixed work-

loads

 Springer International Publishing Switzerland 2016

J. Zhan et al. (Eds.): BPOE 2015, LNCS 9495, pp. 10–21, 2016.

DOI: 10.1007/978-3-319-29006-5

BigDataBench-MT: A Benchmark Tool for Generating Realistic Mixed Data 11

1 Introduction

In modern cloud data centers, a large number of tenants are consolidated to share a

common computing infrastructure and execute a diverse mix of workloads. Bench-

marking and understanding these workloads is a key problem for system designers,

programmers and researchers to optimize the performance and energy eﬃciency

of data center systems and to promote the development of data center technology.

This work focuses on two classes of popular data center workloads [46]:

– Long-running services. These workloads oﬀer online services such as web

search engines and e-commerce sites to end users and the services usually

keep running for months and years. The tenants of such workloads are service

end users.

– Short-term data analysis jobs. These workloads process input data of many

scales using relatively short periods (e.g. in Google and Facebook data centers,

a majority (over 90 %) of analytic jobs complete within a few minutes [30,46]).

The tenants of such workloads are job submitters.

As data analysis systems such as Hadoop and Spark mature, both types

of workloads widely co-locate in today’s data centers, hence the pressure to

benchmark and understand these mixed workloads rises. Within this context,

we believe that it will be of interest to the data management community and a

large user base to generate realistic workloads such that trustworthy benchmark-

ing reﬂecting the practical data center scenarios can be conducted. Considering

the heterogeneity and dynamic nature of data center workloads and their aggre-

gated resource demands and arrival patterns, this requires overcoming two major

challenges.

Benchmarking Using Actual Workloads Based on Real-World Work-

load Rraces. Data analysis jobs usually have various computation semantics

(i.e. implementation logics or source codes) and input data sizes (e.g. ranging

from KB to PB), and their behaviors also heavily rely on the underlying software

stacks (such as Hadoop or MPI). Hence it is diﬃcult to emulate the behaviors of

such highly diverse workloads just using synthetic workloads such as I/O oper-

ations. On the other hand, generating workloads whose arrival patterns follow

real-world traces is an equally important aspect of realistic workloads. This is

because these traces are the most realistic data sources including both explicit

and implicit arrival patterns (e.g. sequences of time stamped requests or jobs).

Benchmarking Using Scalable Workloads with Realistic Mixes. A good

benchmark needs to ﬂexibly adjust the scale of workloads to meet the require-

ments of diﬀerent benchmarking scenarios. Based on our experience, we noticed

that in many cases, obtaining real workload traces is diﬃcult due to conﬁdential

issues. The limited trace data also restrict the scalability of benchmark. It is

therefore challenging to produce workloads at diﬀerent scales while still guaran-

teeing their realistic mix corresponding to real-world scenarios.

In this paper, we propose a benchmark tool that is a ﬁrst step towards gener-

ating realistic mixed data center workloads. This tool, called BigDataBench-MT,

剩余150页未读，继续阅读

ramissue

粉丝: 354
资源: 1487

2015年BPOE研讨会：大数据基准、性能优化与新兴硬件论文集

benchmarks_test_pso.rar_benchmark_benchmarks

Web Microanalysis of Big Image Data

Graduate thesis in financial management

verisilicon video decoder performance

解释mv user_example_test/ ../../../out-of-box-linux/benchmarks/

sysbench下载

ModuleNotFoundError: No module named 'avalanche.benchmarks'什么意思

parser = argparse.ArgumentParser(description='k-FP benchmarks')

android benchmark

python实现将OpenKE知识图谱基于transH嵌入成向量

从哪里下载边缘检测的数据集

pso_standard_benchmarks_test.m

应该如何搭建mlperf inference测试，测试的项目为ResNet50测试ImageNet数据集

benchmarks.py是什么意思

pytorch yolov7

那里可以找到moead算法c++源码

jetson_benchmarks

财务管理第3章静态分析指标(

No module named 'benchmarks'

最新资源