Spark-GPU框架下海洋地理空间数据分布式并行处理任务调度
|
摘要
大规模长时间序列海洋地理空间数据处理属于计算密集型任务。本文重点介绍Spark框架下如何利用GPU并行计算机制实现海洋地理空间数据分布式并行处理的任务调度,以提高大规模长时间序列海洋地理空间数据处理效率,满足实时交互需求。Spark-GPU框架包括Spark-GPU调度器和Spark-GPU运行时两部分。任务计算量和GPU设备计算能力作为调度策略因子,采用一个多项式时间的2近似算法求解,是一个著名的无关并行机任务调度问题。本文以流场可视化线积分卷积算法作为测试用例,1 000~2 000场的任务调度测试结果表明与原生Spark调度算法相比,Spark-GPU框架执行时间减少了14%~18%,GPU占用比提高了10%~20%。
Long time and large scale Oceanographic Geospatial Data(OGD)processing is computation-intensive.This paper focuses on the method of task scheduling for ODG distributed parallel processing based on Spark with GPU,to imporve processing efficiency of long time and large scale OGD,and satisfy realtime interaction requirements.Spark original scheduling algorithms(FIFO,FAIR)shows severe problem,low efficiency and more execution time when running computation-intensive tasks,To solve the problem,this paper presents a Spark-GPU Framework(SGF).SGF includes Spark-GPU Scheduler(SGS)and Spark-GPU Runtime(SGR).SGS takes into consideration of GPU tasks with different computation and GPU devices with different computing capacity.The scheduling is on Unrelated Parallel Machines and deal with a polynomial 2-Approximation Algorithm.SGR uses JNI+CUDA as GPU task runtime.The method of JNI+CUDA use only one JNI call to achieve high efficiency,and is easy to programming and debug.The main contribution of this paper is as follow:(1)Improved Spark-GPU Framework can support more balance scheduling of large scale computation task running,(2)Describe a scheduling algorithm for large scale computation task on heterogeneous GPU devices by consider GPU tasks with different computation and GPU devices with different computing capacity.Flow Field Visualization is as the test application.On a cluster with 10 GPU nodes,1 000~2 000 field tasks evaluation show the SGF can reduce 14%~18%execution time,improve GPU time occupancy ratio 10%~20%.
Long time and large scale Oceanographic Geospatial Data(OGD)processing is computation-intensive.This paper focuses on the method of task scheduling for ODG distributed parallel processing based on Spark with GPU,to imporve processing efficiency of long time and large scale OGD,and satisfy realtime interaction requirements.Spark original scheduling algorithms(FIFO,FAIR)shows severe problem,low efficiency and more execution time when running computation-intensive tasks,To solve the problem,this paper presents a Spark-GPU Framework(SGF).SGF includes Spark-GPU Scheduler(SGS)and Spark-GPU Runtime(SGR).SGS takes into consideration of GPU tasks with different computation and GPU devices with different computing capacity.The scheduling is on Unrelated Parallel Machines and deal with a polynomial 2-Approximation Algorithm.SGR uses JNI+CUDA as GPU task runtime.The method of JNI+CUDA use only one JNI call to achieve high efficiency,and is easy to programming and debug.The main contribution of this paper is as follow:(1)Improved Spark-GPU Framework can support more balance scheduling of large scale computation task running,(2)Describe a scheduling algorithm for large scale computation task on heterogeneous GPU devices by consider GPU tasks with different computation and GPU devices with different computing capacity.Flow Field Visualization is as the test application.On a cluster with 10 GPU nodes,1 000~2 000 field tasks evaluation show the SGF can reduce 14%~18%execution time,improve GPU time occupancy ratio 10%~20%.
引文
[1] Grossman M,Breternitz M,Sarkar V.Hadoopcl:Mapreduce on distributed heterogeneous platforms through seamless integration of hadoop and opencl[C].Parallel and Distributed Processing Symposium Workshops&PhD Forum(IPDPSW)IEEE:2013IEEE27th International.2013:1918-1927.
[2] Gary F.Aparapi:An open source tool for extending the java promise of write once run anywhereto include the gpu[EB/OL].[2012-09-10]http://conferences.oreilly.com/oscon/oscon2012/public/schedule/detail/23434:O′Reilly,2012.
[3] Segal O,Colangelo P,Nasiri N,et al.SparkCL:A unified programming framework for accelerators on heterogeneous clusters[J].arXiv Preprint arXiv:1505.01120,2015.
[4] Elteir M,Lin H,Feng W,et al.StreamMR:an optimized MapReduce framework for AMD GPUs[C].Parallel and Distributed Systems(ICPADS),IEEE:2011IEEE 17th International Conference on.2011:364-371.
[5] Grossman M,Breternitz M,Sarkar V.Hadoopcl2:Motivating the design of a distributed,heterogeneous programming system with machine-learning applications[J].IEEE Transactions on Parallel and Distributed Systems,2016,27(3):762-775.
[6] Choi W,Hong S,Jeong W K.Vispark:GPU-accelerated distributed visual computing using spark[J].SIAM Journal on Scientific Computing,2016,38(5):S700-S719.
[7] Wang H,Xiao B,Wang L,et al.Accelerating large-scale image retrieval on heterogeneous architectures with Spark[C].ACM:Proceedings of the 23rd ACM International Conference on Multimedia.2015:1023-1026.
[8] Li P,Luo Y,Zhang N,et al.HeteroSpark:A heterogeneous CPU/GPU Spark platform for machine learning algorithms[C].Networking,Architecture and Storage(NAS),IEEE:2015IEEE International Conference on.IEEE,2015:347-348.
[9] Tsiomenko R,Rees B S.Accelerating fast fourier transforms using hadoop and CUDA[J].arXiv Preprint arXiv:1407.6915,2014.
[10] Liu J,Hu Y.Gpu support in spark and gpu/cpu mixed resource scheduling at production scale[EB/OL].[2016-03-25]https://spark-summit.org/2016/events/gpu-support-in-spark-and-gpu-cpumixed-resource-scheduling-at-production-scale/:Databricks,2016.
[11] Lenstra J K,Shmoys D B,Tardos E.Approximation algorithms for scheduling unrelated parallel machines[J].Mathematical programming,1990,46(1-3):259-271.
[12] Cabral B,Leedom L C.Imaging vector fields using line integral convolution[C].ACM:Proceedings of the 20th Annual Conference on Computer Graphics and Interactive techniques.1993:263-270.
[13] Yan Y,Grossman M,Sarkar V.JCUDA:A programmer-friendly interface for accelerating Java programs with CUDA[C].European Conference on Parallel Processing.Berlin:Springer,2009:887-899.
[14] Hunt J.Java native interface[M].Java for Practitioners.London:Springer,1999:417-425.
[2] Gary F.Aparapi:An open source tool for extending the java promise of write once run anywhereto include the gpu[EB/OL].[2012-09-10]http://conferences.oreilly.com/oscon/oscon2012/public/schedule/detail/23434:O′Reilly,2012.
[3] Segal O,Colangelo P,Nasiri N,et al.SparkCL:A unified programming framework for accelerators on heterogeneous clusters[J].arXiv Preprint arXiv:1505.01120,2015.
[4] Elteir M,Lin H,Feng W,et al.StreamMR:an optimized MapReduce framework for AMD GPUs[C].Parallel and Distributed Systems(ICPADS),IEEE:2011IEEE 17th International Conference on.2011:364-371.
[5] Grossman M,Breternitz M,Sarkar V.Hadoopcl2:Motivating the design of a distributed,heterogeneous programming system with machine-learning applications[J].IEEE Transactions on Parallel and Distributed Systems,2016,27(3):762-775.
[6] Choi W,Hong S,Jeong W K.Vispark:GPU-accelerated distributed visual computing using spark[J].SIAM Journal on Scientific Computing,2016,38(5):S700-S719.
[7] Wang H,Xiao B,Wang L,et al.Accelerating large-scale image retrieval on heterogeneous architectures with Spark[C].ACM:Proceedings of the 23rd ACM International Conference on Multimedia.2015:1023-1026.
[8] Li P,Luo Y,Zhang N,et al.HeteroSpark:A heterogeneous CPU/GPU Spark platform for machine learning algorithms[C].Networking,Architecture and Storage(NAS),IEEE:2015IEEE International Conference on.IEEE,2015:347-348.
[9] Tsiomenko R,Rees B S.Accelerating fast fourier transforms using hadoop and CUDA[J].arXiv Preprint arXiv:1407.6915,2014.
[10] Liu J,Hu Y.Gpu support in spark and gpu/cpu mixed resource scheduling at production scale[EB/OL].[2016-03-25]https://spark-summit.org/2016/events/gpu-support-in-spark-and-gpu-cpumixed-resource-scheduling-at-production-scale/:Databricks,2016.
[11] Lenstra J K,Shmoys D B,Tardos E.Approximation algorithms for scheduling unrelated parallel machines[J].Mathematical programming,1990,46(1-3):259-271.
[12] Cabral B,Leedom L C.Imaging vector fields using line integral convolution[C].ACM:Proceedings of the 20th Annual Conference on Computer Graphics and Interactive techniques.1993:263-270.
[13] Yan Y,Grossman M,Sarkar V.JCUDA:A programmer-friendly interface for accelerating Java programs with CUDA[C].European Conference on Parallel Processing.Berlin:Springer,2009:887-899.
[14] Hunt J.Java native interface[M].Java for Practitioners.London:Springer,1999:417-425.