PDW Architecture Gets Real: Customer Implementations Brian Walker | Microsoft Corporation PDW Center of Excellence Murshed Zaman | Microsoft Corporation SQL Customer Advisory Team April 10-12, Chicago, IL Please silence cell phones April 10-12, Chicago, IL Agenda 3 Introducing Parallel Data Warehouse Pre-Built Hardware + Software Appliance • Co-engineered with HP and Dell • Pre-built Hardware • Pre-installed Software • Appliance installed in 1-2 days • Support - Microsoft provides first call support • Hardware partner provides onsite break/fix support Plug and Play Built-in Best Practices Save Time 5 The Power of PDW Massively Parallel Processing (MPP) Symmetric Multi-Processing (SMP) 6 The Basic Full Rack Infiniband & Ethernet SQL Server PDW 2012 • Reduce hardware footprint by virtualizing the entire control server rack down to a few nodes • 1.5x lower price/TB providing the one of the lowest price/TB in the industry • Save up to 70% of storage with up to ~15x compression via the xVelocity columstore • Resilient, scalable, and high performance storage features in Windows Server 2012 replace SAN with high density, low cost SAS JBODS • • • • 128 cores on 8 compute nodes 2TB of RAM on compute Up to 168 TB of temp DB Up to 1PB of user data • 70% more disk I/O bandwidth over SQL Server PDW 2008 R2 7 Data Layout PDW Compute Nodes Dimensional Model Date Dim Item Dim Date Dim ID Calendar Year Calendar Qtr Calendar Mo Calendar Day Prod Dim ID Prod Category Prod Sub Cat Prod Desc Sls Fact D s D S D Date Dim ID Store Dim ID Prod Dim ID Mktg Camp Id Qty Sold Dollars Sold S D Store Dim Store Dim ID Store Name Store Mgr Store Size Promo Dim Mktg Camp ID Camp Name Camp Mgr Camp Start Camp End S D S F 1 F 2 F 3 S 4 F 5 I P I P I P I P I P 8 Seamlessly Add Capacity Start Small Linearly Scale OUT Smallest (53TB) To Largest (6PB) Add Capacity • Start small with a few Terabyte warehouse • Add capacity up to 6 Petabytes Add Capacity 53 TB 6 PB Start Small And Grow Largest Warehouse PB 9 Any Size : Next-Gen Performance Country Supplier Sales Products Customer xVelocity - Fast Data Query Processing Columnstore Provides Dramatic Performance • Updateable and clustered xVelocity columnstore • Stores data in columnar format • Memory-optimized for next-generation performance • Updateable to support bulk and/or trickle loading Up to 50X Faster Up to 15x compression Save Time and Costs Batch Processing 10 The Power of Updatable ColumnStore Indexing on PDW 2012 Any Data: Hadoop Integration Polybase Details • External Tables and full SQL query access to data stored in HDFS • HDFS bridge for direct & fully parallelized access of data in HDFS • Joining ‘on-the-fly’ PDW data with data from HDFS • Parallel import of data from HDFS in PDW tables for persistent storage Regular T-SQL Results Enhanced PDW Query Engine PDW 2012 Structured data External Table HDFS Bridge • Parallel export of PDW data into HDFS including ‘round-tripping’ of data HDFS Data Nodes Unstructured data 12 Existing Excel Skillset With Big Data Familiar Tools To Analyze Structured/Unstructured Data Familiar Tools Analyse Big Data Hadoop Data Structured Data • Native Microsoft BI Integration to PDW • Structured and unstructured data in same spreadsheet • Widely adopted and familiar user tools High Adoption Of Excel No IT Intervention Analyze All Data Types 13 Simultaneous Reporting from Structured and Unstructured Data 14 15 Upgrading to PDW Gains 100x Improvement “…basic queries that previously took 20 minutes only took seconds using the SQL Server 2008 R2 Parallel Data Warehouse.” -Tom Settle, Assistant VP, Data Warehousing, Hy-Vee Benefits 16 16 Business Objectives Critical Provide Broader Range of Critical Customer Purchasing Data - Current system only supported 2 years of data – Business required 7 years Load Speed Improve Performance of Complex Transformations - Faster delivery of data within specified SLAs Save Time Enable Self-Service Reporting - SSAS/SSRS/SharePoint/Excel Query Enable User Ad hoc Reporting - Leveraging Excel/SharePoint Scale Provide solution that Scales to Meet Future Data Needs - Expansion of history, point of sale detail, and expansion into social media Save Costs Reduced IT Costs - Creating self-sufficient end users – Frees IT to focus on delivering new data 17 Shift from ETL to ELT Using the Power of MPP • Move their complex transformations and calculations to SQL Server Parallel Data Warehouse from ETL server • PDW has allowed Hy-Vee to create an enterprise data warehouse centralizing data from many sources • Complex Transformations Archiving point of sale source files for later data extraction 18 Upgrade to PDW 2012 Future Option • Improves their opportunity to further analyze social media data • Query data without having to move it into a relational database • Provides an alternative archive solution for point of sale data 19 Data Archive Challenge – Financial Customer Current Solution • Reporting Services Business only actively analyzes a rolling 12 months of data Archive Servers • Regulations require data is on-line and accessible for extended period • Data > 12 months is pushed to a farm of SQL servers to meet regulatory requirements Centralized EDW Data Archive Challenge – Financial Customer Future Solution • Replace archive farm with Hadoop cluster • PDW provides single point of access • Allows analyst to leverage existing SQL skills • Much lower maintenance and administration • Meets regulatory requirements Reporting Services Archive Servers Centralized EDW HDFS bridge HDFS Data Nodes Unstructured data AMD Boosts Performance with PDW “We used to worry about backlogs, but no more,” - Rajarao Chitturi, Database and Applications Manager at AMD AMD is also processing more reporting queries than it previously could—between 10,000 and 13,000 a day—with an average runtime of a few seconds and virtually no performance issues. Benefits AMD runs an average of 1,500 loads per day, and data loads to a given table range from fourminute to four-hour intervals. AMD averages about 500,000 file loads a day. Because of the user complaints about the previous system, the data warehouse team had one employee devoted full time to addressing performance-related support tickets. With Parallel Data Warehouse, AMD has reduced support work to just a few hours a week. 22 22 AMD Business Challenges Obstacles With SMP Oracle Linux Based Reporting Load Demand • Only supported 6 month data retention • Loading data always lagged behind by days • Issues loading concurrently with high query volume • Analyst couldn’t access recent data • Continuous data loads throughout the day while users were querying the system • Custom reporting tools hosted on Linux uses JDBC and ODBC drivers 23 Project Overview Critical Wafer Quality Assurance Data - 42 TB on PDW Save Space Space Saving PDW Index Lite Approach - Oracle required excessive non-clustered indexes to get any performance Load Speed Improved Loading Speed - 660 GB/hr. throughput Query 10,000 – 13,000 Analytic Queries per Day - Most are scan intensive Save Time Faster Backups – Complete in 1~2 hours per Database - Compared to a week on Oracle Save Costs Reduced Support Costs by 90% - No more chopping up queries to fit the data warehouse 24 Parallel Data Warehouse 2012 25 Other PDW Sessions Online Advertising: Hybrid Approach to Large-Scale Data Analysis (DAV-303-M) Data Analytics and Visualization Breakout Session (60 minutes) Fri April 12, 2013, 2:45 PM - 3:45 PM in Sheraton 3 Anna Skobodzinski Christian Bonilla Dmitri Tchikatilov Trevor Attridge 26 Win a Microsoft Surface Pro! Complete an online SESSION EVALUATION to be entered into the draw. Draw closes April 12, 11:59pm CT Winners will be announced on the PASS BA Conference website and on Twitter. Go to passbaconference.com/evals or follow the QR code link displayed on session signage throughout the conference venue. Your feedback is important and valuable. All feedback will be used to improve and select sessions for future events. 27 Thank you! Diamond Sponsor Platinum Sponsor April 10-12, Chicago, IL