Scaling on one node Hybrid engines with Multi-GPU on In-Memory database queries S23294 - Peter Strohm - Jedox AG GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 1
Jedox GPU Database What we are doing GPU Technology Conference Simplify Europe Planning, 2017 Analytics Peter and Strohm Reporting - @psjedox - #JedoxGTC17 2
GPU Database for Planning & Forecasting GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 33
Jedox 7 Architecture GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 4
Smart Modeling in Jedox Smart Data-driven Modeling and Smart Interfaces Start Insights Freeform Files File Exports (XLS, CSV) Custom Excel Workbook1 Smart data-driven Modeling Dimension types, attributes Structure & hierarchies Start Self-service Integration Scheduled Integration Job Model Pre-defined Jedox Model Custom Model Structured Template XLS, CSV, WSS Fill Data Custom hierarchies & data Smart Data Mapping Dimension load Data load GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 5
Jedox 7 GPU Advancement 1. Host RAM 2. GTX Support 3. Virtual Machine 4. GPU IN CLOUD 5. GPU Advisor GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 6
Jedox GPU Database Processing Large Calculation Trees GPU Technology Conference Simplify Europe Planning, 2017 Analytics Peter and Strohm Reporting - @psjedox - #JedoxGTC17 7
Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-Memory OLAP-Database on GPU IN-GPU-MEMORY OLAP All Data in Main/GPU Memory Hierarchical structure of consolidated elements Year Q1 Q2 Q3 Q4 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec All regions Europe France Italy UK North America USA Canada Mexico Count TweetID Longitude GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 8
Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-Memory OLAP-Database on GPU IN-GPU-MEMORY OLAP Calculations on the fly Store only non-zero values Year Q1 Q2 Q3 Q4 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec All regions Europe France Italy UK North America USA Canada Mexico Count TweetID Longitude GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 9
OLAP planner engine Query tree and subtrees Query root Calculation nodes Data nodes HANDLING LARGE TREES Group Nodes by Weighting Function Find Distinct Node Groups by Estimated Runtime Distribute Nodes to Engines Learning from Prior Queries GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 10
OLAP planner engine Query tree and subtrees Query root Calculation groups Data nodes HANDLING LARGE TREES Group Nodes by Weighting Function Find Distinct Node Groups by Estimated Runtime Distribute Nodes to Engines Learning from Prior Queries GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 11
Distribute subqueries to all engines HANDLING LARGE TREES Assign Calculation Engines Dynamically to Sub-Nodes Re-assign finished Engines to waiting Node Groups Repeat until end is reached Sub-Nodes Calculation engines.. GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 12
OLAP planner engine Multiple users User Nodes MULTI-USER SCENARIO Calculation node Data node Multiple Read Queries at the same Time How to assign Engines to Users What to do with free Engines How to estimate runtim How to sync different contexts GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 13
OLAP planner engine Multiple users User node Calculation node Data node Calculation engine.. GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 14
Jedox GPU Database Performance Overview GPU Technology Conference Simplify Europe Planning, 2017 Analytics Peter and Strohm Reporting - @psjedox - #JedoxGTC17 15
SpeedUp with up to 8 GPUs on average distributed data sets GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 16
SpeedUp with up to 8 GPUs on average distributed data sets SpeedUp factor when adding 1 GPU (compared to (n-1) GPUs) GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 17
SpeedUp with up to 8 GPUs on average distributed data sets PERFORMANCE BOOST Assuming similar behaviour with 16, 32 GPUs Artificial tests show that we can assume factor of ~1,5x when doubling number of GPUs GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 18
Jedox GPU Database Pl{ai}nning the Future of Planning GPU Technology Conference Simplify Europe Planning, 2017 Analytics Peter and Strohm Reporting - @psjedox - #JedoxGTC17 19
Predict Top-Level Automated Forecasting: AIssisted Planning On all Hierarchy Levels Legal Entities, Countries, Products Additional Information Lower/Upper bounds Accuracy Impact AI CALCULATED RECOMMENDATIONS IN MANAGEMENT REPORTS GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 20
Amount of Data in the Office of Finance FORECAST WITH NOTHING Short timelines (36 data points) Low Amount of Features Individual & Specific Data Sets Difficult Correlations What-if Analysis AI CALCULATED RECOMMENDATIONS ON LOW DATA GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 21
JEDOX GPU IN-MEMORY DATABASE 1 Jedox Social Analytics App download at: www.jedox-social-analytics.com DDE Legacy 2 Jedox Download (Trial for FREE) at: www.jedox.com 3 Contact: Steffen Wittmer - steffen.wittmer@jedox.com Peter Strohm peter.strohm@jedox.com GPU Technology Conference Europe 2017 Peter Strohm - @psjedox - #JedoxGTC17 22
Jedox GPU Database Thank you GPU Technology Conference Simplify Europe Planning, 2017 Analytics Peter and Strohm Reporting - @psjedox - #JedoxGTC17 23