1. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 1 Satisfying Data-Intensive Queries Using GPU Clusters Haicheng Wu, Jeff Young Sudhakar Yalamanchili Computer Architecture and Systems Laboratory Center for Experimental Research in Computer Systems School of Electrical and Computer Engineering Georgia Institute of Technology Sponsors: AIC, AMD, LogicBlox Inc., National Science Foundation, NEC, NVIDIA

2. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 2 Application: Data Warehousing On-line and off-line analysis Retail analysis Forecasting Pricing Etc… Combination of relational data queries and computational kernels Current applications process 1 to 50 TBs of data [1] Techniques can be applied to other “Big Data” problems like irregular graphs, sorting [1] Independent Oracle Users Group. A New Dimension to Data Warehousing: 2011 IOUG Data Warehousing Survey. …… LargeQty(p) <- Qty(q), q > 1000. ……

3. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 3 Proposed System Model Red Fox: Compilation and optimization of queries for GPUs Remove need for application developer to optimize applications to run on GPUs Oncilla: Global Address Space (GAS) layer Create an API to simplify data movement and scheduling

4. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 4 Red Fox Compilation Flow RA-to-PTX ( nvcc + RA-Lib) Primitives Library Runtime Manager LogicBlox Front-End Language Front-End Translation Layer Back-End Datalog Queries Query PlanPTX/Binary Kernel Kernel Weaver

5. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 5 Relational Algebra Primitives on GPUs Multi-stage algorithm Simple primitives are close to maximum performance More complex primitives could show better performance with newer implementations (in progress with NVIDIA Research) Raw Performance (NVIDIA C2050) Fastest known for GPUs! Practical MAX

6. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 6 Red Fox: TPC-H Q1 Results GPU computation scales well with problem size Improved primitives could lead to further 10x speedup

7. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 7 Oncilla: Fabrics for Accelerator Clouds Goal: Transparent, efficient host memory aggregation across node for accelerators Solution: Use Global Address Spaces (GAS) and commodity fabrics (HT, QPI, PCIe, 10GE, IB) Support in-core databases using software from Red Fox project Companies: LogicBlox, NVIDIA, AIC

8. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 8 Oncilla aims to combine support for multiple types of data transfer and CUDA- based optimizations under a simplified runtime. Ex: “oncilla_malloc(2 GB, node2, gpumem)” Enable application developers and schedulers to take advantage of high- performance GAS without needing to be experts in specialized hardware Oncilla: Efficient Data Movement

9. SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING | GEORGIA INSTITUTE OF TECHNOLOGY HPCDB 2012 - Satisfying Data-Intensive Queries Using GPU Clusters November 11 th, 2012 9 Questions? For more information: Red Fox: H. Wu, G. Diamos, H. Cadambi, and S. Yalamanchili, “KernelWeaver: Automatically Fusing Database Primitives for Efficient GPU Computation,” MICRO, December 2012 http://gpuocelot.gatech.edu/projects/compiler-projects/ Oncilla: http://gpuocelot.gatech.edu/projects/compiler-projects/oncilla-gas-infrastructure/ J. Young, S. Yalamanchili, Commodity Converged Fabrics for Global Address Spaces in Accelerator Clouds,” HPCC, June, 2012