GPU Support In Spark And GPU/CPU Mixed Resource Scheduling At Production Scale
7 likes3,887 views
1. GPU support in Spark can accelerate analytics workloads through automatically generating CUDA code from Spark Java code or integrating Spark with GPU-enabled libraries and applications. 2. Production deployments face challenges in identifying GPU vs CPU execution, data preparation for GPU, and low resource utilization. Scheduling must handle mixed GPU and CPU workloads across non-identical hosts to avoid overload and improve utilization. 3. IBM Conductor with Spark provides solutions through fine-grained scheduling that recognizes GPU tasks, prioritizes and allocates resources independently, and allows adaptive scheduling between CPU and GPU. This improves time to results through better resource utilization.
Ad
More Related Content
What's hot (19)
Viewers also liked (20)
Ad
Similar to GPU Support In Spark And GPU/CPU Mixed Resource Scheduling At Production Scale (20)
Ad
More from Spark Summit (20)
Recently uploaded (20)
GPU Support In Spark And GPU/CPU Mixed Resource Scheduling At Production Scale
- 1. GPU Support in Spark and GPU/ CPU Mixed Resource Scheduling at Production Scale Yonggang Hu, IBM, DE Junfeng Liu, IBM, Architect
- 2. About us • Yonggang Hu Distinguished Engineer, IBM Chief Architect at Platform Computing, IBM. Vice President and Application Architect at JPMorgan Chase Working on distributed computing, grid, cloud and big data for the past 20 years. • Junfeng Liu IBM Platform Computing Architect, focusing on Big data platform design and implementation. Successfully delivering solutions to several key customers.
- 3. Agenda • GPU and Spark integration motivation • The challenges in production deployments • Solutions in IBM Conductor with Spark • Demo
- 4. Spark & GPU Spark apps are CPU intensive Need to handle more data and bigger models Machine Learning Predicative analytics, Logistic regression, ALS Kmeans, etc. Graph Analytics Security, Fraud Detection Social Network Analytics GraphX Video/Speech Analytics Object Recognition Dialog Financial Risk Analytics Market simulation Credit risk. home-grown, apps from Murex, Misys Spark-enable existing GPU appsGPU-enable Spark apps
- 5. Various ways to enable Spark & GPU • Use GPUs for accelerating Spark Libraries and operations without changing interfaces and underlying programming model. • Automatically generate CUDA code from the source Spark Java code • Integrate Spark with GPU-enabled application & system (e.g., Spark integrated with Caffe, TensorFlow and customer applications)
- 6. Production Challenges • However – Identification of GPU execution vs. CPU execution in DAG – Data preparation for GPU execution – Low resource utilization for CPU or GPU or both • Cannot assume all compute hosts are identical and have GPU resource available • GPU is a lot more expensive !!! – Overload and contention when running mixed GPU & CPU workload – Long tail & GPU & CPU tasks failover – Task ratio control on different resources Stage 2 Stage 1 reduceByKey collect Stage GPU GPU Group CPU Group
- 7. A typical example – Personalized Medicine – Adverse Drug Reaction Workload - 30X faster at learning speed and 4.3 X speed up at end-2-end - Need to fully utilize both GPU and CPU resources to get economic benefits
- 8. Scheduling Granularity • Scheduling at application level – Mesos and Yarn tag the GPU machine with label – Schedule the application on GPU hosts based resource requirement of application – Corse grained scheduling leads to low utilization of CPU/GPU. • Scheduling at DAG level – Need fine grained sharing for GPU resources rather than reserving entire GPU machines – Identify GPU operation – Optimize the DAG tree by decupling GPU operations from CPU operations and by inserting new GPU stages – Reduce GPU wait time, enable sharing GPU among different jobs and therefore improve the overall GPU utilization
- 9. GPU tasks recognition • GPU and CPU tasks mixed together • Separate the workload is necessary for scheduling control GPUFunction() Python-C Wrapper to Invoke Native Function Function implemented by CUDA/OpenCL GPU library Python-C/C++ Wrapper
- 10. GPU tasks recognition • Mark the GPU workload by DAG operation – Go through the DAG tree to identify the stages with GPU requirement – Optimize the distribution by inserting GPU stage
- 11. Policies • RM needs capability to identify the GPU hosts and manage along with CPU resources • Prioritization policy - share GPU resource among applications • Allocation policy – control GPU and CPU allocation independently – multi-dimensional scheduling • Fine grained policy to schedule tasks according to GPU optimized DAG plan
- 12. Adaptive Scheduling • CPU & GPU tasks are convertible in many applications • Scheduling needs adaptive capability – If GPU is available, use a portion of GPU – Otherwise run rest of tasks on CPU dstInBlocks.join(merged).mapV alues { …. if (useGPU) { loadGPULib() callGPU () } else { //CPU version } }
- 13. Adaptive Scheduling CPU CPU CPU GPU GPU GPU Driver Executors Executors Executors Tasks -------------- Node 0 Node 1 Node n
- 14. Efficiency Considerations • Do we need to wait GPU resource if there is CPU available? • Do we need rerun the CPU tasks on GPU if tasks on CPU are long-tail? • Do we need to have failover cross resource type?
- 15. Defer Scheduling • Traditional defer Scheduling – Wait for data locality – Cache, Host, Rack • Resource based defer scheduling – Necessary if the GPU can greatly speed up task execution – Wait time is acceptable
- 16. Future works • Global optimization – Consider the cost of additional shuffle stage – Consider data locality of CPU and GPU stage – Add time dimension to MDS – Optimize global DAG tree execution – Use historical data to optimize future execution, e.g, future iteration
- 17. Fine grain, dynamic allocation of resources maximizes efficiency of Spark instances sharing a common resource pool. Multi-tenant, multi-framework support. Eliminates cluster sprawl. 2 Run Spark natively on a shared infrastructure without the dependency of Hadoop. Reduce application wait time, improving time to results. 1 Building Spark Centric Shared Service with IBM Conductor End-to-End Enterprise Class Solution Improve Time to Results Proven architecture at extreme scale, with enterprise class workload management, multi-version support for Spark, monitoring, reporting, and security capabilities. 3 Reduce Administration Costs Increase Resource Utilization • IBM STC Spark Distribution • IBM Platform Resource Orchestrator / Session Scheduler, application service manager. • IBM Spectrum Scale FPO 4
- 18. IBM Conductor with Spark IBM Bluemix Spark Cloud Service in production – thousands of users and tenants. Third party audited benchmark indicated significant performance/ throughput/SLA advantages https://stacresearch.com/news/2016/03/29/IBM160229
- 19. IBM Systems 19© 2016 IBM Corpora*on IBM Conductor with Spark Monitor and Reporting with Elastic (ELK) ! Integrated Elastic Search, Logstash, Kibana for customizable monitoring ! Built-in monitoring Metrics ! Cross Spark Instance Groups ! Cross Spark Applications within Spark Instance Group ! Within Spark Application ! Built-in monitoring inside Zeppelin Notebook
- 20. Demo
- 21. THANK YOU. Contact information or call to action goes here.
- 22. Acceleration Opportunities for GPUs & Spark Analytics Model Computational Patterns suitable for GPU Acceleration Regression Analysis Cholesky Factorization, Matrix Inversion, Transpose Clustering Cost-based iterative convergence Nearest-neighbor Search Distance calculations, Singular Value Decomposition, Hashing Neural Networks Matrix Multiplications, Convolutions, FFTs, Pair-wise dot-products Support Vector Machines Linear Solvers, Dot-product Association Rule Mining Set Operations: Intersection, union Recommender Systems Matrix Factorizations, Dot-product Time-series Processing FFT, Distance and Smoothing functions Text Analytics Matrix multiplication, factorization, Set operations, String computations, Distance functions Monte Carlo Methods Random number generators, Probability distribution generators Mathematical Programming Linear solvers, Dynamic Programming OLAP/BI Aggregation, Sorting, Hash-based grouping, User-defined functions Graph Analytics Matrix multiplications, Path traversals