Optimizing SQL Insert Performance - GSE Belux

More documents

Recommendations

Info

Asymmetric Leaf Page Split • Available in NFM and automatic • Design point is to provide performance relief for classic sequential index key problem • Asymmetric index page split will occur depending on an insert pattern when inserting in the middle of key range – Instead of previous 50-50 split prior to DB2 9 – Up to 50% reduction in index split • Asymmetric split information is tracked in the actual pages that t are inserted into, so it is effective across multiple threads across DB2 members • PK62214 introduces changes to the tracking and detection logic, and it should work much better for data sharing – Before: DB2 9 only remembered the last insert position and a counter – Now: DB2 remembers an insert 'range‘ and tolerates entries being slightly out of order • It may still not be effective for large key sizes (hundreds of bytes), or if entries come in very bad order (i.e., they do not look sequential) • But for simple cases like 3, 2, 1, 6, 5, 4, 9, 8, 7, 12, 11, 10 ... DB2 will be able to determine that the inserted entries are ascending 40 © 2010 IBM Corporation
Randomised Index Key • Index contention can be a major problem and a limit for scalability • This problem is more severe in data sharing because of index page P- lock contention • A randomized d index key can reduce lock contention ti • CREATE/ALTER INDEX … column-name RANDOM, instead of ASC or DESC • Careful trade-off required between lock contention relief and additional getpages, g read/write I/Os, and increased number of lock requests • This type of index can provide dramatic improvement or degradation! – e.g. Very bad choice for applications with range predicates • Recommend making randomized indexes only when buffer pool resident 41 © 2010 IBM Corporation
Page 1 and 2: DB2 for z/OS Optimizing Insert Perf
Page 3 and 4: Agenda • Insert Space Search •
Page 5 and 6: Choice: Performance or Space Reuse
Page 7 and 8: Space Search Steps (Classic Partiti
Page 9 and 10: PBG search algorithm for entire tab
Page 11 and 12: Segmented Tablespace • Segmented
Page 13 and 14: Partitioned Tablespace • Use page
Page 15 and 16: Distributed Free Space • Use dist
Page 17 and 18: TRACKMOD NO • Reduces spacemap co
Page 19 and 20: MEMBER CLUSTER … • Rows inserte
Page 21 and 22: MEMBER CLUSTER, PCTFREE = 0 and FRE
Page 23 and 24: Typical Performance Bottlenecks and
Page 25 and 26: Read and Write I/O for Index and Da
Page 27 and 28: Read and Write I/O for Index and Da
Page 29 and 30: Lock/Latch and Service Task Waits
Page 31 and 32: DB2 Latch Contention • Latch Coun
Page 33 and 34: GENERATE_UNIQUE() • Built-in func
Page 35 and 36: DB2 9 Performance Enhancements 35
Page 37 and 38: Large Index Page Size • Available
Page 39: Increased Index Look Aside • Prio
Page 43 and 44: Identifying Unreferenced Indexes
Page 45 and 46: DB2 10 Performance Enhancements 45
Page 47 and 48: V10 Enhancements ... • In-line LO
Page 49 and 50: Summary - Key Points • Decide whe

Optimizing SQL Insert Performance - GSE Belux

Create successful ePaper yourself

Delete template?

Save as template?