1 / 45

SQL Server 2000 Performance Improvements Walter Wong Technology Specialist Microsoft Hong Kong Limited

SQL Server 2000 Performance Improvements Walter Wong Technology Specialist Microsoft Hong Kong Limited. Current Customer Requirements And Future Performance Trends. Increasing user populations Increasing transaction rates Increasing database disk I/O Increasing data and query complexity

Gideon
Télécharger la présentation

SQL Server 2000 Performance Improvements Walter Wong Technology Specialist Microsoft Hong Kong Limited

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. SQL Server 2000 Performance Improvements Walter Wong Technology Specialist Microsoft Hong Kong Limited

  2. Current Customer Requirements And Future Performance Trends • Increasing user populations • Increasing transaction rates • Increasing database disk I/O • Increasing data and query complexity • Increasing application network I/O • Stricter response time requirements • Need to “Scale-Up” on SMP platforms • Need to “Scale-Out” on data tier

  3. Performance ChallengeIncreasing Disk I/O • Drive larger/faster I/O subsystems • More efficient use of large memory • Support for TB+ size databases • Optimize for OLTP I/O profiles • Optimize for DSS I/O profiles • Handle text/image data efficiently

  4. SQL Server 2000 Solutions • Improved parallel scan processing • Merry-go-round scans • More scalable Readahead • Dynamic asynchronous I/O • Large memory support (AWE) • Scalable checkpoint processing • Efficient “Text in Row” format

  5. CPU1 CPU2 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K 8K Readahead I/O Readahead I/O Readahead I/O Readahead I/O Improved Parallel Scan • Less time spent in parallel page supplier critical section • Readahead I/O now issued outside the critical section • Architecture feeds 16 pages at a time (fewer calls to PPS) SDES SDES Parallel Page Supplier CritSec Parallel Page Supplier CritSec

  6. User 1: 25% Scanned User 1: 50% Scanned User 2: 25% Scanned With Merry-go-round scans: User 1: 25% Scanned User 2: Starts Scanning User 1: Scan Complete User 2: 75% Scanned User 2: Remaining 25% Scanned Merry-Go-Round Scans Without Merry-go-round scans: Can result in disk thrashing during large table scans! • Reduces disk thrashing • Improves cache behavior • Affects unordered scans • No synchronization

  7. = Mapped = Unmapped < 3GB > 3 GB • AWE allows 36bit physical address space via PAE (up to 64gb) • Pages mapped/unmapped between virtual and physical • Maps are magnitude less costly than copies • Enabled via “awe enabled” parameter (static non-paging) Large Memory Support • Virtual address space still limited to 3GB

  8. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 Chkpt starts 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Chkpt continues 1 1 1 1 …while Lazywriter flushes pages 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Chkpt completes Scalable Checkpoint Results in Fewer Checkpoint I/Os on Busy Systems!

  9. SQL 7.0 Text Format: • Text stored “off row” in tree • Text pointer jump to root Col1 Col2 Col3 Tree Root Text Pages • Text stored “in row” if it fits Text In Row Format: • Otherwise, root stored “in row” Col1 Col2 Col3 Text • Reduced I/O for text retrieval • Enabled/disabled by table option Col1 Col2 Col3 • Optional “inline limit” (256 default) • Increased I/O for non-text scans Text Pages Text In Row Format

  10. Performance ChallengeProcessor and SMP Scale-Up • More efficient use of CPU resources • Reduction of critical code paths • Elimination of unnecessary processing • Improve overall SMP scaling • Improve processor affinity • Reduce cross processor cache traffic/contention

  11. SQL Server 2000 Solutions • Server side cursor caching • Parameter caching • Enhanced intra-query parallelism • Full parallel index creation • Partitioned free buffer lists

  12. 1. Server Side Cursor Caching • Cursor compile plan • Saves allocation of object • Cursor execution plan • Representation of cursor is cached • Benefits: • Workloads dominated by cursor operations • 10-15% reduction in CPU consumption for cursor operations • No application changes necessary!

  13. 2. Parameter Caching • Type checks and default parameter values cached on first execution • Subsequent executions use cached information and default values • Re-cached if plan recompile/schema change • Benefits: • Heavy stored procedures use with many parameters/defaults • 6% improvement over SQL Server 7.0 on SAP SD workload • No application changes necessary (if using prepare/execute to call stored procedure)

  14. 3. Intra-Query Parallelism • Integrated bitmap filtering • Multi-threaded access to spools • Reduced parallel sort stalls • Elapsed time costing • Insert/Update/Delete integration • Results: • Improved TPC-H DSS workload by 40% over SQL Server 7.0

  15. Index • DOP Determined on Execution • Partition Functions Defined Per CPU • Per CPU Table Scans Initiated • Filter Applied to Scans • Filtered Results Merged CPU 1 CPU 2 CPU 3 CPU 4 Partition Function Filter Partition Function Filter Partition Function Filter Partition Function Filter CPU 1 Scan CPU 2 Scan CPU 3 Scan CPU 4 Scan Table: 4. Parallel Index Creation Near Linear Scaling!

  16. CPU1 CPU2 CPU3 CPU4 Hash Table (PageId, DbId) Free List Free List Free List • Results in Free List per CPU • Enhances cache locality of free list descriptors and buffers • Distributed Lazywriting balances population of free lists • Improves Scale-Up on SMP platforms • Architecture in place for further NUMAish enhancements 5. Partitioned Free Buffer Lists Free List Buffer Array (contains buff descriptors which point to 8k pages)

  17. Performance ChallengeIncreasing Network I/O • Larger and more complex app architectures (2-tier, N-tier, hybrid) • App tier network traffic increasing (eCommerce driven) • Data tier network traffic increasing • Stricter response time requirements • Increasing CPU consumption for network I/O processing

  18. SQL Server 2000 Solutions • Prepare/unprepare caching to reduce network roundtrips • Metadata caching to reduce data on wire • Support for high-speed System Area Networks (SAN)

  19. 1. Prepare/Unprepare Caching • Prepare/unprepare cached on client • Prepare piggybacked on first exec • Unprepare piggybacked on next prepare • Removes N+2 roundtrip scenario • Stored procedure help • Overall benefit inversely related to increase in execs of prepared plan • Must bind before exec in application to see benefit

  20. Event class Description Exec Prepared SQL Indicates when a prepared SQL statement or statements have been executed by ODBC, OLEDB, or DB-Library. Prepare SQL Indicates when an SQL statement or statements have been prepared for use by ODBC, OLEDB, or DB-Library. SQL:BatchCompleted Transact-SQL batch has completed. SQL:BatchStarting Transact-SQL batch has started. SQL:StmtCompleted Transact-SQL statement has completed. SQL:StmtStarting Transact-SQL statement has started. Unprepare SQL Indicates when a prepared SQL statement or statements have been unprepared by ODBC, OLEDB, or DB-Library. Prepare/Unprepare event class

  21. 2. Metadata Caching • Metadata cached on client between subsequent calls on connection • Coordinated between client/server • Metadata only sent if schema change • Data reduction inversely related to # of columns (~10 bytes per column) • Bonus: Saves CPU required to process metadata on both client and server! • Reduces cost of select with > 50 columns by over 10% • No application changes necessary!

  22. 3. SAN Support • Socket Netlib supports WSD (DC only) • Native VI Netlibs (Servernet and Giganet) • High bandwidth, low latency, reliable • Bonus: • Low system CPU overhead • Benefits: • Network intensive client/server apps • Scale-Out data tier interconnects • Reduced kernel time by 1/3 in SAP workload - 20%+ increase in users!

  23. Performance ChallengeData and Query Complexity • Efficient query processing of ever increasing data set sizes • Increasing query complexity for business intelligence and data warehousing systems • Understanding of the “nature of the data” • Efficient execution of multi-user workloads with mixed query complexities

  24. SQL Server 2000 Solutions • Improved statistics collection and representation • Even more efficient plans • Advanced index enhancements • Improved multi-user robustness

  25. 1. Improved Statistics • Similar Bucket Merge Algorithm (SBM) builds maxdiff histogram for statistics • Dynamic bucket boundaries based on variances • Effectively handles high low and high data skew • Captures density and distribution in one pass • Improves cardinality estimates by creating more granular statistics • Optimizer still understands SQL 7.0 statistics • Necessary in case where database is upgraded • Will update to new format automatically over time if “auto update statistics” enabled • To update all stats in db use sp_updatestats

  26. 2. More Efficient Plans • More granular and accurate statistics means improved cardinality estimates • Better understanding of “nature of data” • Higher quality input for optimizer • Modular and flexible QP design has allowed “fine tuning” for unique and complex queries • Even more “trivial plan” cases supported

  27. 3. Indexed Views • Index on view may contain join, aggregation, or combination • Once created, contents of view persist • Optimizer may use index even if not directly referenced by query • Benefits: • Decision Support Systems • Data Warehouses / DataMarts • Business Intellegence Systems • Integration with OLAP Services • Consider cost of index maintenance!

  28. 4. Multi-User Robustness • Complex or long running queries require resources (memory/CPU) • SQL Server 2000 implements more intelligent scheduling of queries • Prior to execution, QP will examine the “cost” of query relative to other current query activity • QP may choose to delay execution momentarily until adequate resources are available

  29. Scale-Up Results!SAP SD Industry Standard Benchmark 50% More Users on Same 8way Platform!

  30. Performance ChallengeData Tier Scale-Out • Eliminate single server bottleneck • Provide scalable performance/capacity • Application transparency to tables • Intelligent distributed query • Distributed transaction coordination • Integration with cluster services for high availability

  31. SQL Server 2000 SolutionsBuilding blocks for Scale-Out • Distributed Transaction Coordinator • OLEDB/Distributed Query • Distributed Partioned Views • Instead Of Triggers • System Area Network support • Windows Cluster Services • IIS/ASP/ADO/COM+ • Windows 2000 Cluster Services

  32. Distributed Partitioned View • UNION ALL view over local table and one or more remote “member” tables • Member tables have identical structure with unique partitioning column • Disjoint partitioning intervals enforced with CHECK constraints • Participating member servers use OLEDB linked server support • DTC maintains transactional integrity • QP optimizes distributed plans

  33. 227,079 tpmC @ $19.12 per tpmC 12 Compaq Proliant 8500 Servers 96 CPUs, 48GB memory, 18.7TB Disk Scale-Out Results! “Enough throughput to handle all the e-commerce orders placed on the web last year in two days!”

  34. SQL Server 2000 features for high-end servers • More memory • Faster networking • DSS workloads • Multi-instance SQL Server • Distributed Partitioning

  35. = Mapped = Unmapped < 3GB > 3GB • AWE/PAE allows 36bit physical address space (up to 64gb) • Pages mapped/unmapped between virtual and physical • Maps are magnitude less CPU cycles than I/Os • Mapping latency is nearly zero 1. More Memory - AWE • Virtual address space still limited to 3GB

  36. 1. More Memory – AWE cont’d • More memory means… • Reduced disk I/O rate  fewer disks • Less I/O  fewer I/O stack CPU cycles • Faster latency  reduced lock contention • Considerations • Enable with sp_configure “awe enabled” • AWE memory is truly non pageable • Use “max server memory” as appropriate • Multi-instance/clusters need care

  37. 2. Faster networking - SAN • 3 tier apps change net equation • App servers make short, frequent calls to SQL Server • High network traffic  CPU cycles • Reduces CPU available for “real” work • New network technology… • Apps can interact directly with NIC • TCP/IP software stack bypassed • Vast savings in CPU cycle cost • More CPU for “real” work (SQL!)

  38. 3. DSS workloads • Parallel disk scans improvements • Reduce internal friction • Aggressive, self-tuning I/O scheduling • “max async I/O”  GONE in SQL 2000 • Higher utilization of more disks • Parallel index build improvements • Partitioning of build across all CPUs • Merry-go-round I/O efficiencies • Higher utilization of more CPUs • Faster index creation times!

  39. 4. Multi-instance SQL Server • Server consolidation solution • Mix differing apps on single system • Take advantage of more CPUs • Take advantage of more memory (PAE) • Tools to make life simple… • Automatic memory governor • sp_configure “max server memory” • sp_configure “affinity mask”

  40. 5. Distributed partitioned views • Server scale-out solution • Transparency provided by SQL • Shared-nothing  very linear scaling • Take advantage of more hardware • Works best when … • User requests partition readily • Usually routes to “home” partition • Reliable, hi-speed network between nodes • Not for complex queries that span cluster

  41. Tips for SQL 2000 • Data placement • Spread all your data everywhere • Let file groups/proportional fill do the work • Except the log – separate device, cached • Use fiber scheduling feature • User mode context switching • Best when CPU usage is high (80-90%) • Especially on 4 or more way SMP • Not for heterogeneous queries or extended SP environments

  42. Where can I get MSDN? • Visit MSDN Online at msdn.microsoft.com • Register for the MSDN Flashemail newsletter at msdn.microsoft.com/resources/msdnflash.asp • Join the MSDN Online Membership at msdn.microsoft.com/community • Become an MSDN CD Subscriber at msdn.microsoft.com/subscriptions • Attend more MSDN events!

  43. More Resources www.microsoft.com/dna www.microsoft.com/sql

  44. Questions?

More Related