An official website of the United States government
As data analytics applications become increasingly important in a wide range of domains, the ability to develop large-scale and sustainable platforms and software infrastructure to support these applications has significant potential to drive research and innovation in both science and business domains. This paper characterizes performance and power-related behavior trends and tradeoffs of the two predominant frameworks for Big Data analytics (i.e., Apache Hadoop and Spark) for a range of representative applications. It also evaluates system design knobs, such as storage and network technologies and power capping techniques. Experimental results from empirical executions provide meaningful data points for exploring the potential of software-defined infrastructure for Big Data processing systems through simulation. The results provide better understanding of the design space to build multi-criteria application-centric models as well as show significant advantages of software-defined infrastructure in terms of execution time, energy and cost. It motivates further research focused on in-memory processing formulations regarding systems with deeper memory hierarchies and software-defined infrastructure.
more »
« less
ConEx: Efficient Exploration of Big-Data System Configurations for Better Performance
Configuration space complexity makes the big-data software systems hard to configure well. Consider Hadoop, with over nine hundred parameters, developers often just use the default configurations provided with Hadoop distributions. The opportunity costs in lost performance are significant. Popular learning-based approaches to auto-tune software does not scale well for big-data systems because of the high cost of collecting training data. We present a new method based on a combination of Evolutionary Markov Chain Monte Carlo (EMCMC)} sampling and cost reduction techniques tofind better-performing configurations for big data systems. For cost reduction, we developed and experimentally tested and validated two approaches: using scaled-up big data jobs as proxies for the objective function for larger jobs and using a dynamic job similarity measure to infer that results obtained for one kind of big data problem will work well for similar problems. Our experimental results suggest that our approach promises to improve the performance of big data systems significantly and that it outperforms competing approaches based on random sampling, basic genetic algorithms (GA), and predictive model learning. Our experimental results support the conclusion that our approach strongly demonstrates the potential toimprove the performance of big data systems significantly and frugally.
more »
« less
- Award ID(s):
- 1909414
- PAR ID:
- 10196300
- Date Published:
- Journal Name:
- IEEE Transactions on Software Engineering
- ISSN:
- 0098-5589
- Page Range / eLocation ID:
- 1 to 1
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
DOI 10.1109/COMPSAC.2019.10284 Abstract—Hadoop is a popular data-analytics platform based on the MapReduce model. When analyzing extremely big data, hard disk drives are commonly used and Hadoop performance can be optimized by improving I/O performance. Hard disk drives have different performance depending on whether data are placed in the outer or inner disk zones. In this paper, we propose a method that uses knowledge of job characteristics to place data in hard disk drives so that Hadoop performance is improved. Files of a job that intensively and sequentially accesses the storage device are placed in outer disk tracks which have higher sequential access speed than inner tracks. Temporary and permanent files are placed in the outer and inner zones, respectively. This enables repeated usage of the faster zones by avoiding the use of the faster zones by permanent files. Our evaluation demonstrates that the proposed method improves the performance of Hadoop jobs by 15.0% over the normal case when file placement is not used. The proposed method also outperforms a previously proposed placement approach by 9.9%.more » « less
-
The ability to accurately estimate job runtime properties allows a scheduler to effectively schedule jobs. State-of-the-art online cluster job schedulers use history-based learning, which uses past job execution information to estimate the runtime properties of newly arrived jobs. However, with fast-paced development in cluster technology (in both hardware and software) and changing user inputs, job runtime properties can change over time, which lead to inaccurate predictions. In this paper, we explore the potential and limitation of real-time learning of job runtime properties, by proactively sampling and scheduling a small fraction of the tasks of each job. Such a task-sampling-based approach exploits the similarity among runtime properties of the tasks of the same job and is inherently immune to changing job behavior. Our analytical and experimental analysis of 3 production traces with different skew and job distribution shows that learning in space can be substantially more accurate. Our simulation and testbed evaluation on Azure of the two learning approaches anchored in a generic job scheduler using 3 production cluster job traces shows that despite its online overhead, learning in space reduces the average Job Completion Time (JCT) by 1.28×, 1.56×, and 1.32× compared to the prior-art history-based predictor. We further analyze the experimental results to give intuitive explanations to why learning in space outperforms learning in time in these experiments. Finally, we show how sampling-based learning can be extended to schedule DAG jobs and achieve similar speedups over the prior-art history-based predictor.more » « less
-
Aleti A., Panichella A (Ed.)Users of highly-configurable software systems often want to optimize a particular objective such as improving a functional outcome or increasing system performance. One approach is to use an evolutionary algorithm. However, many applications today are data-driven, meaning they depend on inputs or data which can be complex and varied. Hence, a search needs to be run (and re-run) for all inputs, making optimization a heavy-weight and potentially impractical process. In this paper, we explore this issue on a data-driven highly-configurable scientific application. We build an exhaustive database containing 3,000 configurations and 10,000 inputs, leading to almost 100 million records as our oracle, and then run a genetic algorithm individually on each of the 10,000 inputs. We ask if (1) a genetic algorithm can find configurations to improve functional objectives; (2) whether patterns of best configurations over all input data emerge; and (3) if we can we use sampling to approximate the results. We find that the original (default) configuration is best only 34% of the time, while clear patterns emerge of other best configurations. Out of 3,000 possible configurations, only 112 distinct configurations achieve the optimal result at least once across all 10,000 inputs, suggesting the potential for lighter weight optimization approaches. We show that sampling of the input data finds similar patterns at a lower cost.more » « less
-
The ability to accurately estimate job runtime properties allows a scheduler to effectively schedule jobs. State-of-the-art online cluster job schedulers use history-based learning, which uses past job execution information to estimate the runtime properties of newly arrived jobs. However, with fast-paced development in cluster technology (in both hardware and software) and changing user inputs, job runtime properties can change over time, which lead to inaccurate predictions. In this paper, we explore the potential and limitation of real-time learning of job runtime properties, by proactively sampling and scheduling a small fraction of the tasks of each job. Such a task-sampling-based approach exploits the similarity among runtime properties of the tasks of the same job and is inherently immune to changing job behavior. Our analytical and experimental analysis of 3 production traces with different skew and job distribution shows that learning in space can be substantially more accurate. Our simulation and testbed evaluation on Azure of the two learning approaches anchored in a generic job scheduler using 3 production cluster job traces shows that despite its online overhead, learning in space reduces the average Job Completion Time (JCT) by 1.28x, 1.56x, and 1.32x compared to the prior-art history-based predictor. Finally, we show how sampling-based learning can be extended to schedule DAG jobs and achieve similar speedups over the prior-art history-based predictor.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1109/TSE.2020.3007560
