面向CPU-GPU异构系统的数据分析负载均衡策略
|
摘要
应用于高性能计算领域的通用GPU拥有强大的并行计算能力,以通用GPU作为主处理器的数据分析系统相较于传统数据库能够提供更好的性能。在大数据场景下,如何根据CPU和GPU的资源在处理器之间合理分配工作负载是亟待解决的问题。提出了一种CPU-GPU异构数据分析系统上的负载均衡处理策略。该策略采用流水线模型将工作负载分解,基于流水线设计了负载均衡模型,将工作负载合理分配至异构处理器,减少系统总执行时间开销,实现了性能提升。实验结果表明,提出的基于流水线的负载均衡模型能适应不同查询请求下的不同数据量场景,具有良好的性能。
With strong parallel computation power, GPU-based data analytic systems can achieve better performance than traditional CPU-based data analytic systems. However, how to leverage the resources of CPU and GPU to dispatch workloads appropriately in massive data scenarios remains to be solved. We propose a load balancing strategy on heterogeneous CPU-GPU data analytic systems. It adopts a pipeline model to split workloads, and a load balancing model based on pipeline is proposed to dispatch workloads to heterogeneous processors reasonably and appropriately, which reduces the total execution time of the system and enhances the performance. Experimental results show that the load balancing model based on pipeline is suitable for various queries with different dataset volumes and has great performance.
With strong parallel computation power, GPU-based data analytic systems can achieve better performance than traditional CPU-based data analytic systems. However, how to leverage the resources of CPU and GPU to dispatch workloads appropriately in massive data scenarios remains to be solved. We propose a load balancing strategy on heterogeneous CPU-GPU data analytic systems. It adopts a pipeline model to split workloads, and a load balancing model based on pipeline is proposed to dispatch workloads to heterogeneous processors reasonably and appropriately, which reduces the total execution time of the system and enhances the performance. Experimental results show that the load balancing model based on pipeline is suitable for various queries with different dataset volumes and has great performance.
引文
[1] Big data SQL database [DB/OL]. [2018-05-10].https://sqream.com/.
[2] MapD:GPU database for fast, interactive and visual analytics[EB/OL].[2018-05-10].https://www.mapd.com.
[3] GPU database designed from the ground up to deliver truly real-time insights[EB/OL].[2018-05-10].https://www.kinetica.com/product/.
[4] He B,Lu M,Yang K,et al.Relational query coprocessing on graphics processors[J].ACM Transactions on Database Systems (TODS),2009,34(4):1-39.
[5] He B,Yu J X.High-throughput transaction executions on graphics processors[J].Proceedings of the VLDB Endowment,2011,4(5):314-325.
[6] Breβ S,Saake G.Why it is time for a HyPE:A hybrid query processing engine for efficient GPU coprocessing in DBMS[J].Proceedings of the VLDB Endowment,2013,6(12):1398-1403.
[7] Li J,Tseng H W,Lin C,et al.HippogriffDB:Balancing I/O and GPU bandwidth in big data analytics[J].Proceedings of the VLDB Endowment,2016,9(14):1647-1658.
[8] Yuan Y,Lee R,Zhang X.The Yin and Yang of processing data warehousing queries on GPU devices[J].Proceedings of the VLDB Endowment,2013,6(10):817-828.
[9] Mittal S,Vetter J S.A survey of CPU-GPU heterogeneous computing techniques[J].ACM Computing Surveys (CSUR),2015,47(4):1-35.
[10] Chen L,Villa O,Krishnamoorthy S,et al.Dynamic load balancing on single-and multi-GPU systems[C]//Proc of 2010 IEEE International Symposium on Parallel & Distributed Processing (IPDPS),2010:1-12.
[11] Momcilovic S,Ilic A,Roma N,et al.Dynamic load balancing for real-time video encoding on heterogeneous CPU+GPU systems[J].IEEE Transactions on Multimedia,2014,16(1):108-121.
[12] Chen Y,Qiao Z,Davis S,et al.Pipelined multi-GPU MapReduce for big-data processing[M]//Computer and Information Science.Berlin:Springer Heidelberg,2013:231-246.
[13] Shahvarani A,Jacobsen H A.A hybrid b+-tree as solution for in-memory indexing on CPU-GPU heterogeneous computing platforms[C]//Proc of the 2016 International Conference on Management of Data,2016:1523-1538.
[14] Kaldewey T,Lohman G,Mueller R,et al.GPU join processing revisited[C]//Proc of the 8th International Workshop on Data Management on New Hardware,2012:55-62.
[15] Wang Z,Yang J,Melhem R,et al.Simultaneous multikernel GPU:Multi-tasking throughput processors via fine-grained sharing[C]//Proc of 2016 IEEE International Symposium on High Performance Computer Architecture (HPCA),2016:358-369.
[16] Breβ S,K?cher B,Heimel M,et al.Ocelot/HyPE:Optimized data processing on heterogeneous hardware[J].Proceedings of the VLDB Endowment,2014,7(13):1609-1612.
[17] Breβ S,Siegmund N,Bellatreche L,et al.An operator-stream-based scheduling engine for effective GPU coprocessing[M]//Advances in Databases and Information Systems.Berlin:Springer Berlin Heidelberg,2013:288-301.
[18] Riha L,Shea C,Malik M,et al.Task scheduling for GPU accelerated OLAP systems[C]//Proc of the 2011 Conference of the Center for Advanced Studies on Collaborative Research,2011:107-119.
[19] Fermi architecture [EB/OL].[2018-05-10].http://www.nvdia.cn/object/fermi-architecture-cn.html.
[20] Kepler architecture [EB/OL]. [2018-05-16]. https://www.nvidia.cn/object/nvidia-kepler-cn.html.
[2] MapD:GPU database for fast, interactive and visual analytics[EB/OL].[2018-05-10].https://www.mapd.com.
[3] GPU database designed from the ground up to deliver truly real-time insights[EB/OL].[2018-05-10].https://www.kinetica.com/product/.
[4] He B,Lu M,Yang K,et al.Relational query coprocessing on graphics processors[J].ACM Transactions on Database Systems (TODS),2009,34(4):1-39.
[5] He B,Yu J X.High-throughput transaction executions on graphics processors[J].Proceedings of the VLDB Endowment,2011,4(5):314-325.
[6] Breβ S,Saake G.Why it is time for a HyPE:A hybrid query processing engine for efficient GPU coprocessing in DBMS[J].Proceedings of the VLDB Endowment,2013,6(12):1398-1403.
[7] Li J,Tseng H W,Lin C,et al.HippogriffDB:Balancing I/O and GPU bandwidth in big data analytics[J].Proceedings of the VLDB Endowment,2016,9(14):1647-1658.
[8] Yuan Y,Lee R,Zhang X.The Yin and Yang of processing data warehousing queries on GPU devices[J].Proceedings of the VLDB Endowment,2013,6(10):817-828.
[9] Mittal S,Vetter J S.A survey of CPU-GPU heterogeneous computing techniques[J].ACM Computing Surveys (CSUR),2015,47(4):1-35.
[10] Chen L,Villa O,Krishnamoorthy S,et al.Dynamic load balancing on single-and multi-GPU systems[C]//Proc of 2010 IEEE International Symposium on Parallel & Distributed Processing (IPDPS),2010:1-12.
[11] Momcilovic S,Ilic A,Roma N,et al.Dynamic load balancing for real-time video encoding on heterogeneous CPU+GPU systems[J].IEEE Transactions on Multimedia,2014,16(1):108-121.
[12] Chen Y,Qiao Z,Davis S,et al.Pipelined multi-GPU MapReduce for big-data processing[M]//Computer and Information Science.Berlin:Springer Heidelberg,2013:231-246.
[13] Shahvarani A,Jacobsen H A.A hybrid b+-tree as solution for in-memory indexing on CPU-GPU heterogeneous computing platforms[C]//Proc of the 2016 International Conference on Management of Data,2016:1523-1538.
[14] Kaldewey T,Lohman G,Mueller R,et al.GPU join processing revisited[C]//Proc of the 8th International Workshop on Data Management on New Hardware,2012:55-62.
[15] Wang Z,Yang J,Melhem R,et al.Simultaneous multikernel GPU:Multi-tasking throughput processors via fine-grained sharing[C]//Proc of 2016 IEEE International Symposium on High Performance Computer Architecture (HPCA),2016:358-369.
[16] Breβ S,K?cher B,Heimel M,et al.Ocelot/HyPE:Optimized data processing on heterogeneous hardware[J].Proceedings of the VLDB Endowment,2014,7(13):1609-1612.
[17] Breβ S,Siegmund N,Bellatreche L,et al.An operator-stream-based scheduling engine for effective GPU coprocessing[M]//Advances in Databases and Information Systems.Berlin:Springer Berlin Heidelberg,2013:288-301.
[18] Riha L,Shea C,Malik M,et al.Task scheduling for GPU accelerated OLAP systems[C]//Proc of the 2011 Conference of the Center for Advanced Studies on Collaborative Research,2011:107-119.
[19] Fermi architecture [EB/OL].[2018-05-10].http://www.nvdia.cn/object/fermi-architecture-cn.html.
[20] Kepler architecture [EB/OL]. [2018-05-16]. https://www.nvidia.cn/object/nvidia-kepler-cn.html.