SQREAM DB on IBM Power9
•Download as PPTX, PDF•
2 likes•564 views
GPU data warehouse Sqream DB provides a massively parallel processing engine powered by GPUs that is faster and more efficient than CPU-based systems. It can ingest terabytes of data per hour onto a single GPU and handle petabytes of data stored in a compact 2U server. With familiar SQL queries and connectors, Sqream DB accelerates analytics by 100x over traditional warehouses through its GPU-accelerated processing and columnar storage.
Report
Share
More Related Content
What's hot
What's hot (20)
Similar to SQREAM DB on IBM Power9
Similar to SQREAM DB on IBM Power9 (20)
More from Ganesan Narayanasamy
More from Ganesan Narayanasamy (20)
Recently uploaded
Recently uploaded (20)
SQREAM DB on IBM Power9
- 1. GPU DATA WAREHOUSE OF MASSIVE DATA ANALYTICS TACKLING THE CHALLENGES David Garber, Sales Manager, West
- 2. HQ in 7 WTC New York | R&D in Tel Aviv CORPORATE PROFILE FOUNDED IN 2010 with Alibaba Cloud Strategic Partnership Patents 10 Employees 70+
- 3. 2008 <1-4TB 2010 <10TB 2016 TB-PB DATA STORES ARE GROWING EXPONENTIALLY Technology CPU Technology GPU
- 4. BUT DATA WAREHOUSES WERE NOT BUILT TO HANDLE THIS LEVEL OF DATA NoSQL & Hadoop GPU Database Relational DB 1970s-1990s 1990-2010 MPP 2005-2010 In-Memory 2010… Massive Data Hive Kinetica Aerospike Mongo DB SQREAM DBMapD MemSQL VoltDB DB2 BLU IBM Netezza IBM Oracle DB2 Teradata Vertica Redshift Exadata Oracle Server SQL Classic Relational
- 5. X86 CPU SYSTEMS ARE NOT ADVANCING PROCESS TAKES A REALLY LONG TIME 3-5 hours30 minutes Data lake Legacy MPP DB 1000 of CPUs 1-2 hours BI customersData sources ETL + Cubes + aggregation + index
- 6. SQL QUERIES AND BI ANALYTICS ARE TAKING WAY TOO LONG
- 7. VALUABLE INSIGHTS GO UNDISCOVERED BI Lost 90%Data Analyzed <10%
- 8. INTRODUCING SQREAM DB GPU-ACCELERATED DATA WAREHOUSE 100xfaster Queries 10%of resources Cost 20xmore data Analyze
- 9. SQREAM DB • Massively parallel engine • Faster and smaller than CPUs POWERED BY GPUs • Terabytes to petabytes • Not limited by RAM • Ingests 3 TB/hr/GPU • Powerful columnar storage • Always-on compression • Familiar ANSI SQL • Standard connectors • 100 TB in a 2U server • Highly cost-efficient • Python, AI, Jupyter, etc. • Built for data science COMPLEMENTS EXISTING INFRASTRUCTURE MASSIVELY SCALABLE SQL DATABASE EXTENSIBLE FOR ML/AI MINIMAL FOOTPRINT LIGHTNING FAST
- 10. SCALE-UP SOLUTION • SQream DB can scale up by expanding the attached storage, or out by adding additional compute nodes HP SN6000B 16Gb FC Switch 47434642454144403935383437333632312730262925282423192218211720161511141013912873625140 BI fabric Storage fabric
- 11. HIGH THROUGHPUT CONVERGED • SQream DB designed for high-throughput • IBM Power Systems is the only NVLink CPU-to-GPU enabled architecture • IBM AC922, with POWER9 and NVLINK can transfer data at up to 300GB/s, almost 9.5x faster than PCIe 3.0 found in x86- based architectures, reducing classic I/O bottlenecks 2x NVIDIA Tesla V100 2x NVIDIA Tesla V100 IBM Power 9 IBM Power 9
- 12. GPU-ACCELERATED DATA WAREHOUSE SQREAM DB BOOSTS QUERY PERFORMANCE BY UP TO 150% FOR IBM POWER9 USERS “GPU-accelerated analytics are an increasingly important part of our industry. The announcement of SQream on the IBM POWER9 platform takes this concept to another level of performance, as the POWER9 CPU with embedded NVIDIA NVLink interface to NVIDIA’s GPUs allows SQream to enable even faster processing of data on POWER9 servers.” Sumit Gupta, VP of HPC and AI for IBM Cognitive Systems
- 13. HIGH THROUGHPUT ARCHITECTURE IT’S NOT JUST THE CORES RAM Power9 CPU Tesla V100 GPU VRAM Tesla V100 GPU VRAM 170GB/s per CPU NVLink – 300GB/s BiDi 900GB/s RAM Power9 CPU Tesla V100 GPU VRAM Tesla V100 GPU VRAM IBM SMP bus
- 14. UP TO 2x FASTER LOADING SQREAM DB ON POWER9 • SQream DB relies on CPU as well as GPUs for loading • IBM’s Power9 multi-core architecture makes loading much faster than comparable x86 based systems • IBM Power9 system loaded data nearly twice as fast as the x86 based machine IBM Power9 AC922: 2x POWER9 16C @ 3.8GHz | 256 GB DDR4 2666 MHz | SSD storage | 4x NVIDIA Tesla V100 (SXM2 NVLINK - 16GB) Dell PowerEdge R740: 2x Intel Xeon Silver 4112 CPU @ 2.60GHz | 256GB DDR4 2666MHz | SSD storage | 4x NVIDIA Tesla V100 (PCIe - 16GB) 1,929 1,094 - 500 1,000 1,500 2,000 2,500 Load Time (seconds) LoadTime(seconds) Lowerisbetter Load time for 6 billion TPC-H records Dell Poweredge R740 IBM Power9 AC922
- 15. UP TO 3.7x FASTER QUERIES SQREAM DB ON POWER9 • SQream DB on Power9 is between 150% to 370% faster than comparable x86 architectures • The CPU-GPU NVLink bandwidth is key to performance in complex queries IBM Power9 AC922: 2x POWER9 16C @ 3.8GHz | 256 GB DDR4 2666 MHz | SSD storage | 4x NVIDIA Tesla V100 (SXM2 NVLINK - 16GB) Dell PowerEdge R740: 2x Intel Xeon Silver 4112 CPU @ 2.60GHz | 256GB DDR4 2666MHz | SSD storage | 4x NVIDIA Tesla V100 (PCIe - 16GB) 52.83 10.35 84.5 78.57 14.06 2.8 30.29 29.01 0 10 20 30 40 50 60 70 80 90 TPC-H Query 8 TPC-H Query 6 TPC-H Query 19 TPC-H Query 17 Querytime(seconds) Lowerisbetter Query SQream DB performance IBM Power9 vs Intel Xeon (Skylake) Dell PowerEdge R740 IBM Power9 AC922
- 16. DATA EXPLORATION MADE EASY Query raw data directly Immediate ad-hoc querying Ideal for data science and discovery Multiple JOINs on any field Time Series Regular Expressions ANSI-92 Compatible Window Analysis ODBC, JDBC Python Connectivity
- 17. HOW IT WORKS Chunking Data Data Data Automatic adaptive compression Data Data Data GPU Parallel chunk processing Data Skipping Data Data Data Columnar process + Metadata tagging Data DataDataData Raw data Data Data Data Data Data Data Data Data Data
- 18. 18 CONCEPT 1 • Columnar databases are very common, efficient for analytics • Good for big data analysis - aggregations over days, per accounts • Columnar databases compress data better because of the higher data locality COLUMNAR
- 19. 19 CONCEPT 2 SQream DB tables enable scalability by partitioning data in multiple dimensions. We call this chunking. Chunking is automatically and transparently performed during ingest. CHUNKING Table Chunks Columns
- 20. 20 CONCEPT 3 • Always on, calculated for every chunk • Example: SELECT * FROM t WHERE YEAR>2017 (all chunks with YEAR<=2017 can be skipped) ZONE MAPS day month year val1 val2 val3 10 2017 11 2017 12 2017 01 2018 02 2018 03 2018 Only this will be read This is skipped Automatic, transparent index replacement
- 21. FINANCE Fraud analysis Risk consolidation Customized services RETAIL Monitor Competitors Customer Experience Operational Decisions TELECOM Customer 360 Competitive Analysis Network Optimization HEALTHCARE Care Management IOT Devices Genomic Research
- 22. UNDERSTAND 40 MILLION CUSTOMERS TELECOM NVIDIA Tesla GPUs 96 GB RAM + 6 TB storage X86 HPDL380g9 $200K 40 NODES 5 full racks 7600 CPU cores $10,000,000 18M 10M 360M 120M Ingest time Reporting time Ownership Cost
- 23. INCREASE REVENUES AD-TECH Tesla GPUs Acquisition Sources 85 TB/day in ad impressions for constructing bidding histograms Data 2x NVIDIA Queries take 5 hours Extract Data Ingest Queries take 5 minutes
- 24. Tesla GPUs Acquisition Sources Data 8x NVIDIA Extract Not feasible X Queries take 5 minutes INCREASE REVENUES AD-TECH 360 TB/day ingested to enhance bid histogram accuracy Data Ingest
- 25. OF PERFORMANCE MEDIA CUT THE COST 4x NVIDIA Tesla GPUs 512 GB RAM + iSCSI JBOD (20TB) X86 Dell C4130 8 full 42U racks, 56 S-Blades 7 TB RAM Compression ratio Netezza Ownership Cost 33.70 Average query time (seconds) Processing Units (S-Blade / GPUs) 4.0 56 $12,000,000 31.70 4.7 4 $500,000 ACV calculation on 24 TB of data, 300B rows, 8 tables with complex, nested joins
- 26. FEEL FREE TO ADDRESS Headquarters, 7 WTC 250 Greenwich Street New York, New York David Garber, Sales Manager, West davidg@sqream.com | sqream.com WE ARE SOCIAL CONTACT