Computation of Large Sparse Aggregated Areas for Analytic Database Queries

Transcription

1 Computation of Large Sparse Aggregated Areas for Analytic Database Queries Steffen Wittmer Tobias Lauer Jedox AG Collaborators: Zurab Khadikov Alexander Haberstroh Peter Strohm

2 Business Intelligence and Corporate Planning

3 Jedox BI Workflow Database Data Integration Extract Transform Load Jedox ETL OLAP Calculation Aggregation Enterprise rules Write-back Jedox OLAP User Frontend Reporting Analysis Planning Simulation Jedox for Excel ERP Text Files Extract Transform Load Aggregation Enterprise rules Write-back Jedox Web Jedox Mobile

4 Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year Online Analytical Processing (OLAP) Data modeled as multidimensional cube Operations: Analysis Reporting Planning Simulation All regions Europe France Italy UK North America USA Canada Mexico Deviation Actual Budget Dimensions are structured hierarchically: Consolidated elements Base elements Year Q1 Q2 Q3 Q4 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

5 Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-memory OLAP storage model All data stored in main memory All regions Europe France Italy UK North America USA Canada Mexico Deviation Actual Budget

6 Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-memory OLAP storage model All data stored in main memory Only store base cells All regions Europe France Italy UK North America USA Canada Mexico Deviation Actual Budget

7 Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-memory OLAP storage model All data stored in main memory Only store base cells Do not store zero-value cells Memory saving, data consistency All regions Europe France Italy UK North America USA Canada Mexico Deviation Actual Budget Represent cells as (key, value) pairs, e.g. ( (2, 1, 0), 4.0 ) Note: Values are double precision!

8 Jan Feb Mar Q1 Apr May Jun Q2 Jul Aug Sep Q3 Oct Nov Dec Q4 Year In-memory OLAP storage model All data stored in main GPU memory Only store base cells Do not store zero-value cells Memory saving, data consistency Compute other cells on the fly when needed All regions Europe France Italy UK North America USA Canada Mexico Deviation Actual Budget Use GPU to accelerate

9 GPU aggregation solutions Target driven aggregation Parallel reductions Very fast for small/dense target areas Source driven aggregation Exploits OLAP characteristic of sparsity Well suited for large aggregated areas

10 Target driven aggregation Fast parallel aggregation step Utilizes shared, global and constant memory Coalesced memory access Almost no thread divergence Multi-GPU solution Performance optimized bulk aggregations prefiltering

11 Large sparse aggregated areas Large: up to millions of target cells Top 10 products for each customer Sparse: most target values are zero = zero = non-zero slice all years customer year product...

12 Handling large sparse areas Requirements Performance and memory consumption that correlate with number of non-zero target cells Solution Source driven approach - Serialized aggregation with atomics - Utilize hash tables on GPU source cells target cells

13 Source driven aggregation Jan, 2011 sold units source cells Q1, all years sold units target cell area Q1 Year Q2 Jan... Apr Year parent map h(x) Year, all years sold units target cell hash table atomic add atomic add

14 Atomics: Contention thread serialization h(x) atomic add Great improvement in Fermi and Kepler over CC 1.X

15 Reducing contention ballot: merge with first? warp warp warp preaggregation warp-wise different hash functions h 1 (x) h 2 (x)

16 Speedup vs. CPU Speedup vs. CPU Speedup: Small target areas (1-11 cells) GPU Target Driven GPU Source Driven 80,0 70,0 76,7 80,0 70,0 60,0 50,0 40,0 55,9 54,9 60,0 50,0 40,0 30,0 26,4 27,8 20,0 13,7 10,0 0,0 0,9 2,0 3,5 0, ,0001 0,001 0,01 0,1 1 Selectivity 30,0 20,0 18,3 13,2 10,0 9,5 12,9 1,6 3,0 0,1 0,5 0,7 0,0 0, ,0001 0,001 0,01 0,1 1 Selectivity Database: 1B records (filled cube cells) GPU: 3x Tesla C2070 (18 GB RAM)

17 Time in seconds Larger areas Calculation times Speedup , target cells 9, ,6 CPU GPU source driven ,2 3, target cells target cells 1547 target cells 4,9 0,0 2,0 4,0 6,0 8,0 10,0 Speedup over CPU Database: 40M records (filled cube cells) GPU: 2x Tesla K20 (10 GB RAM)

18 Selectivity Comparison: aggregation algorithms CPU algorithm good for low aggregation Target driven algorithm: good for small and/or dense target areas GPU target driven GPU source driven Source driven algorithm: optimized for large and sparse areas Sparsity CPU algorithm (multi-core) Target size

19 Decision mechanism Future challenges When GPU, when CPU? Which GPU algorithm in which situation? Find suitable thresholds What about large and dense target areas? GPU memory problem

20 Come visit us! Exhibit Hall: Booth 718 Online: