1. SQL Server Parallel Data Warehouse: Supporting Large Scale Analytics José Blakeley, Software Architect Database Systems Group, Microsoft Corporation

2. SQL Server PDW Overview 3/18/2011JHU DIR March 20112

3. Workload Types 3/18/2011JHU DIR March 2011  Online Transaction Processing (OLTP)  Balanced read-update ratio (60%-40%)  Fine-grained inserts and updates  High transaction throughput e.g., 10s K/s  Usually very short transactions e.g., 1-3 tables  Sometimes multi-step e.g., financial  Relatively small data sizes e.g., few TBs  Data Warehousing and Business Analysis (DW)  Read-mostly (90%-10%)  Few updates in place, high-volume bulk inserts  Concurrent query throughput e.g., 10s K / hr  Per query response time < 2 s  Snowflake, star schemas are common e.g., 5-10 tables  Complex queries (filter, join, group-by, aggregation)  Very large data sizes e.g., 10s TB - PB Day-to-day business Analysis over historical data 3

4. SQL Server Parallel Data Warehouse  Shared-nothing, distributed, parallel DBMS  Built-in data and query partitioning  Provides single system view over a cluster of SQL Servers  Appliance concept  Software + hardware solution  Choice of hardware vendors (e.g., HP, Dell, NEC)  Optimized for DW workloads  Bulk loads (1.2 – 2.0 TB/hr)  Sequential scans (700 TB in 3hr)  Scale from 10 Terabytes to Petabytes  1 rack manages ~40 TB  1 PB will need ~25 racks 3/18/2011JHU DIR March 20114

5. Hardware Architecture 3/18/2011JHU DIR March 2011 Compute Nodes Dual Infiniband Control Nodes Active / Passive Spare Compute Node Dual Fiber Channel Client Drivers (ODBC, OLE- DB, ADO.NET) ETL Load Interface Corporate Backup Solution Data Center Monitoring 2 Rack Appliance 5

6. Software Architecture 3/18/2011JHU DIR March 2011 Compute Nodes Compute Node Query Tool MS BI (AS, RS) MS BI (AS, RS) Control Node 3 rd Party Tools DWSQL Landing Zone Node Internet Explorer SQL Server DW Authentication DW Configuration DW Schema TempDB SQL Server User Data Data Movement Service PDW Engine IIS Admin Console Data Access (OLEDB, ODBC, ADO.NET, JDBC) 6

7. Key Software Functionality 3/18/2011JHU DIR March 2011  PDW Engine  Provides single system image  SQL compilation  Global metadata and appliance configuration  Global query optimization and plan generation  Global query execution coordination  Global transaction coordination  Authentication and authorization  Supportability (HW and SW status info via DMVs)  Data Movement Service  Data movement across the appliance  Distributed query execution operators  Parallel Loader  Runs from the Landing Zone  SSIS or command line tool  Parallel Database Copy  High performance data export  Enables Hub-Spoke scenarios  Parallel Backup/Restore  Backup files stored on Backup Nodes  Backup files may be archived into external device/system 7

8. Query Processing  SQL statement compilation  Parsing, validation, optimization  Builds an MPP execution plan  A sequence of discrete parallel QE “steps”  Steps involve SQL queries to be executed by SQL Server at each compute node  As well as data movement steps  Executes the plan  Coordinates workflow among steps  Assembles the result set  Returns result set to client 3/18/2011JHU DIR March 20118

9. 3/18/2011JHU DIR March 2011 8/20/2015 18,000,048,306 rows 4,500,000,000 rows 450,000,000 rows 600,000,000 rows Example DW Schema 30,000,000 rows 25 rows 5 rows 2,400,000,000 rows SELECT TOP 10 L_ORDERKEY, SUM(L_EXTENDEDPRICE*(1-L_DISCOUNT)) AS REVENUE, O_ORDERDATE, O_SHIPPRIORITY FROMCUSTOMER, ORDERS, LINEITEM WHERE C_MKTSEGMENT = 'BUILDING' AND C_CUSTKEY = O_CUSTKEY AND L_ORDERKEY = O_ORDERKEY AND O_ORDERDATE < ‘2010-03-05' AND L_SHIPDATE > ‘2010-03-05' GROUP BY L_ORDERKEY, O_ORDERDATE, O_SHIPPRIORITY ORDER BY REVENUE DESC, O_ORDERDATE 9

10. Example – Schema TPCH 3/18/2011JHU DIR March 2011 ---------------------------------------------------------------------- -- Customer Table -- distributed on c_custkey ---------------------------------------------------------------------- CREATE TABLE customer ( c_custkey bigint, c_name varchar(25), c_address varchar(40), c_nationkey integer, c_phone char(15), c_acctbal decimal(15,2), c_mktsegment char(10), c_comment varchar(117)) WITH (distribution=hash(c_custkey)) ; ---------------------------------------------------------------------- -- Orders Table ---------------------------------------------------------------------- CREATE TABLE orders ( o_orderkey bigint, o_custkey bigint, o_orderstatus char(1), o_totalprice decimal(15,2), o_orderdate date, o_orderpriority char(15), o_clerk char(15), o_shippriority integer, o_comment varchar(79)) WITH (distribution=hash(o_orderkey)) ; ---------------------------------------------------------------------- -- Customer Table -- distributed on c_custkey ---------------------------------------------------------------------- CREATE TABLE customer ( c_custkey bigint, c_name varchar(25), c_address varchar(40), c_nationkey integer, c_phone char(15), c_acctbal decimal(15,2), c_mktsegment char(10), c_comment varchar(117)) WITH (distribution=hash(c_custkey)) ; ---------------------------------------------------------------------- -- Orders Table ---------------------------------------------------------------------- CREATE TABLE orders ( o_orderkey bigint, o_custkey bigint, o_orderstatus char(1), o_totalprice decimal(15,2), o_orderdate date, o_orderpriority char(15), o_clerk char(15), o_shippriority integer, o_comment varchar(79)) WITH (distribution=hash(o_orderkey)) ; ---------------------------------------------------------------------- -- LineItem Table -- distributed on l_orderkey ---------------------------------------------------------------------- CREATE TABLE lineitem ( l_orderkey bigint, l_partkey bigint, l_suppkey bigint, l_linenumber bigint, l_quantity decimal(15,2), l_extendedprice decimal(15,2), l_discount decimal(15,2), l_tax decimal(15,2), l_returnflag char(1), l_linestatus char(1), l_shipdate date, l_commitdate date, l_receiptdate date, l_shipinstruct char(25), l_shipmode char(10), l_comment varchar(44)) WITH (distribution=hash(l_orderkey)) ; ---------------------------------------------------------------------- -- LineItem Table -- distributed on l_orderkey ---------------------------------------------------------------------- CREATE TABLE lineitem ( l_orderkey bigint, l_partkey bigint, l_suppkey bigint, l_linenumber bigint, l_quantity decimal(15,2), l_extendedprice decimal(15,2), l_discount decimal(15,2), l_tax decimal(15,2), l_returnflag char(1), l_linestatus char(1), l_shipdate date, l_commitdate date, l_receiptdate date, l_shipinstruct char(25), l_shipmode char(10), l_comment varchar(44)) WITH (distribution=hash(l_orderkey)) ; 10

11. Example - Query 3/18/2011JHU DIR March 2011 SELECT TOP 10 L_ORDERKEY, SUM(L_EXTENDEDPRICE*(1-L_DISCOUNT))AS REVENUE, O_ORDERDATE, O_SHIPPRIORITY FROM CUSTOMER, ORDERS, LINEITEM WHERE C_MKTSEGMENT = 'BUILDING' AND C_CUSTKEY = O_CUSTKEY AND L_ORDERKEY = O_ORDERKEY AND O_ORDERDATE < ‘2010-03-05' AND L_SHIPDATE > ‘2010-03-05' GROUP BY L_ORDERKEY, O_ORDERDATE, O_SHIPPRIORITY ORDER BY REVENUE DESC, O_ORDERDATE Ten largest “building” orders shipped since March 5, 2010 11

12. Example – Execution Plan 3/18/2011JHU DIR March 2011 ------------------------------ -- Step 1: create temp table at control node ------------------------------ CREATE TABLE [tempdb].[dbo].[Q_[TEMP_ID_664]] ( [l_orderkey] BIGINT, [REVENUE] DECIMAL(38, 4), [o_orderdate] DATE, [o_shippriority] INTEGER ); ------------------------------ -- Step 2: create temp tables at all compute nodes ------------------------------ CREATE TABLE [tempdb].[dbo].[Q_[TEMP_ID_665]_[PARTITION_ID]] ( [l_orderkey] BIGINT, [l_extendedprice] DECIMAL(15, 2), [l_discount] DECIMAL(15, 2), [o_orderdate] DATE, [o_shippriority] INTEGER, [o_custkey] BIGINT, [o_orderkey] BIGINT ) WITH ( DISTRIBUTION = HASH([o_custkey]) ); ------------------------------- -- Step 3: SHUFFLE_MOVE -------------------------------- SELECT [l_orderkey], [l_extendedprice], [l_discount], [o_orderdate], [o_shippriority], [o_custkey], [o_orderkey] FROM [dwsys].[dbo].[orders] JOIN [dwsys].[dbo].[lineitem] ON ([l_orderkey] = [o_orderkey]) WHERE ([o_orderdate] < ‘2010-03-05' AND [o_orderdate] >= ‘2010-09-15 00:00:00.000') INTO Q_[TEMP_ID_665]_[PARTITION_ID] SHUFFLE ON (o_custkey); ------------------------------ -- Step 1: create temp table at control node ------------------------------ CREATE TABLE [tempdb].[dbo].[Q_[TEMP_ID_664]] ( [l_orderkey] BIGINT, [REVENUE] DECIMAL(38, 4), [o_orderdate] DATE, [o_shippriority] INTEGER ); ------------------------------ -- Step 2: create temp tables at all compute nodes ------------------------------ CREATE TABLE [tempdb].[dbo].[Q_[TEMP_ID_665]_[PARTITION_ID]] ( [l_orderkey] BIGINT, [l_extendedprice] DECIMAL(15, 2), [l_discount] DECIMAL(15, 2), [o_orderdate] DATE, [o_shippriority] INTEGER, [o_custkey] BIGINT, [o_orderkey] BIGINT ) WITH ( DISTRIBUTION = HASH([o_custkey]) ); ------------------------------- -- Step 3: SHUFFLE_MOVE -------------------------------- SELECT [l_orderkey], [l_extendedprice], [l_discount], [o_orderdate], [o_shippriority], [o_custkey], [o_orderkey] FROM [dwsys].[dbo].[orders] JOIN [dwsys].[dbo].[lineitem] ON ([l_orderkey] = [o_orderkey]) WHERE ([o_orderdate] < ‘2010-03-05' AND [o_orderdate] >= ‘2010-09-15 00:00:00.000') INTO Q_[TEMP_ID_665]_[PARTITION_ID] SHUFFLE ON (o_custkey); ------------------------------ -- Step 4: PARTITION_MOVE ------------------------------ SELECT [l_orderkey], sum(([l_extendedprice] * (1 - [l_discount]))) AS REVENUE, [o_orderdate], [o_shippriority] FROM [dwsys].[dbo].[customer] JOIN tempdb.Q_[TEMP_ID_665]_[PARTITION_ID] ON ([c_custkey] = [o_custkey]) WHERE [c_mktsegment] = 'BUILDING' GROUP BY [l_orderkey], [o_orderdate], [o_shippriority] INTO Q_[TEMP_ID_664]; ------------------------------ -- Step 5: Drop temp tables at all compute nodes ------------------------------ DROP TABLE tempdb.Q_[TEMP_ID_665]_[PARTITION_ID]; ------------------------------- -- Step 6: RETURN result to client -------------------------------- SELECT TOP 10 [l_orderkey], sum([REVENUE]) AS REVENUE, [o_orderdate], [o_shippriority] FROM tempdb.Q_[TEMP_ID_664] GROUP BY [l_orderkey], [o_orderdate], [o_shippriority] ORDER BY [REVENUE] DESC, [o_orderdate] ; ------------------------------- -- Step 7: Drop temp table at control node -------------------------------- DROP TABLE tempdb.Q_[TEMP_ID_664]; ------------------------------ -- Step 4: PARTITION_MOVE ------------------------------ SELECT [l_orderkey], sum(([l_extendedprice] * (1 - [l_discount]))) AS REVENUE, [o_orderdate], [o_shippriority] FROM [dwsys].[dbo].[customer] JOIN tempdb.Q_[TEMP_ID_665]_[PARTITION_ID] ON ([c_custkey] = [o_custkey]) WHERE [c_mktsegment] = 'BUILDING' GROUP BY [l_orderkey], [o_orderdate], [o_shippriority] INTO Q_[TEMP_ID_664]; ------------------------------ -- Step 5: Drop temp tables at all compute nodes ------------------------------ DROP TABLE tempdb.Q_[TEMP_ID_665]_[PARTITION_ID]; ------------------------------- -- Step 6: RETURN result to client -------------------------------- SELECT TOP 10 [l_orderkey], sum([REVENUE]) AS REVENUE, [o_orderdate], [o_shippriority] FROM tempdb.Q_[TEMP_ID_664] GROUP BY [l_orderkey], [o_orderdate], [o_shippriority] ORDER BY [REVENUE] DESC, [o_orderdate] ; ------------------------------- -- Step 7: Drop temp table at control node -------------------------------- DROP TABLE tempdb.Q_[TEMP_ID_664]; 12

13. Data Movement Operations 3/18/2011JHU DIR March 2011  SHUFFLE_MOVE  Distributed  Distributed data exchange across the appliance  Result is a distributed table hashed on some column  PARTITION_MOVE  Union of distributed partitions across compute nodes into a table in the control node  MASTER_MOVE  Replicate a table from the control node to all compute nodes  BROADCAST_MOVE  Distributed  Replicated data exchange across appliance  Unconditional shuffle to all compute nodes  Combines PARTITION_MOVE and MASTER_MOVE in one step  TRIM_MOVE  Distribute a replicated table by trimming each copy  Since all the nodes have same copy of the replicated tables the idea is that nodes keep the values that belong to the distributions in that node  REPLICATE_MOVE  Moves a replicated table from 1 to N compute nodes. 13

14. Customer Experience 3/18/2011JHU DIR March 2011  Query speed is generally in ballpark with mainstream competition – Sometimes much faster  Mixed workload handling is good – concurrency of multi-queries, loads plus queries  Customers like remote table copy – mechanism to export entire data marts to SMP SQL Server  Fast time to solution – power up, create databases, define tables, load and query. 14

15. Microsoft Column-store Technology VertiPaq and VertiScan In-memory BI (IMBI) Slides by Amir Netz JHU DIR March 20113/18/201115

16. In-Memory BI Technology  Developed by SQL Analysis Services (OLAP) team  Column-based storage and processing  Only touch the columns needed for the query  Compression (VertiPaq)  Columnar data is more compressible than row data  Fast in-memory processing (VertiScan)  Filter, grouping, aggregation, sorting JHU DIR March 20113/18/201116

17. How VertiPaq Compression Works Read Raw Data Dictionary Encoding Value Encoding Bit Packing Run Length Encoding (RLE) Phase I: Encoding Phase II: Compression Convert to uniform representation (Integer Vectors) Encoding is per column Minimize storage space Compression is per 8M row segments 2x – 10x size reduction1x – 2x size reduction 2x – 4x size reduction~100x size reduction Compression Analysis 5%-25% of data75%-95% of data Organize by Columns VertiPaq Compression 3/18/201117 Hybrid RLE JHU DIR March 2011

18. 3/18/2011JHU DIR March 201118 8/20/2015 Star Join Schema 34 rows 436,892,631 rows 41 rows 13,517 rows 118 rows SELECT FE0.RegionName, FL0.FiscalYearName, SUM (A.ActualRevenueAmt) FROM TECSPURSL00 A JOIN SalesDate L ON A.SalesDateID = L.SalesDateID JOIN UpperGeography UG ON A.TRCreditedSubsidiaryId = UG.CreditedSubsidiaryID JOIN Region FE0 ON UG.CreditedRegionID = FE0.RegionID JOIN FiscalYear FL0 ON L.FiscalYearID = FL0.FiscalYearID GROUP BY FE0.RegionName, FL0.FiscalYearName APPX – 1TB

19. Column-Store on APPX 3/18/2011JHU DIR March 201119  Response time < 2s common  Smaller variance in response time  more predictable query performance

20. THANKS! 3/18/2011JHU DIR March 201120