4226 - IBM watsonx.data 中采用 NVIDIA CuDF 实现原生 GPU 加速.pdf

1、Orlando,FLOctober 69IBM TechXchange 20254226Luis Garcs-Erice IBM ResearchTodd Mostak NVIDIARohan Pednekar IBM Watsonx.dataNative GPU Acceleration in IBM watsonx.data with NVIDIA CuDFAgenda010203040506Velox&GPUsWhat a single GPU can doMultiple GPUs:Presto ExchangeFaster I/O:StorageNext gen I/OWatsonx

2、.dataIBM TechXchange|2025 IBM CorporationVeloxIBM TechXchange|2025 IBM Corporation3A composable and fully extensible C+execution engine library for data management systems.Velox cuDF integrationIBM TechXchange|2025 IBM Corporation4Source:VeloxCon 2025,“Accelerating Velox with cuDF”-https:/ cuDF inte

3、gration example-OrderByIBM TechXchange|2025 IBM Corporation5Source:VeloxCon 2025,“Accelerating Velox with cuDF”-https:/ a single GPU can doMultiple GPUs:Presto ExchangeFaster I/O:StorageNext gen I/OWatsonx.dataIBM TechXchange|2025 IBM CorporationVelox cuDF integration TPC-H SF100 speedup(single GPU)

4、IBM TechXchange|2025 IBM Corporation7Source:VeloxCon 2025,“Accelerating Velox with cuDF”-https:/ cuDF integration TPC-H SF100 cuDF breakdown(single GPU)TPC-H benchmark is 85%TableScan for Velox-cuDF,due to high throughput joins,aggregations,expressions and filteringWhat is in the work breakdown?IO:c

5、opying parquet data from OS page cache to user space,then HtoD copy to GPUDecompress:GPU kernel to decompress the compressed parquet pagesDecode:processing page headers,dictionary,RLE decode,column factoriesFilter-Project:compute expressions and apply boolean masksJoins,Aggregations,Cast,Concatenate

6、 batches8IBM TechXchange|2025 IBM CorporationSource:VeloxCon 2025,“Accelerating Velox with cuDF”-https:/ a single GPU can doMultiple GPUs:Presto ExchangeFaster I/O:StorageNext gen I/OWatsonx.dataIBM TechXchange|2025 IBM CorporationPresto ArchitectureIBM TechXchange|2025 IBM Corporation10How can weex