Deep dive deck given at SQL Saturday Manchester 2015 covering Checkpoint File Pairs, Hash Indexign etc.

1. SQL Server 2014
In-Memory OLTP
Deep Dive
tonyrogerson@torver.net
@tonyrogerson
http://dataidol.com/tonyrogerson

2. Agenda
• Memory
• Storage
• MVCC
• Indexes
• Hash
• Range

3. Memory

4. SQL Server Memory Pools
Stable
Storage
(MDF/NDF
Files)
Buffer Pool
Table
Data Data in/out as required
Memory Internal Structures (proc/log
cache etc.)
SQL Server
Memory
Space
Memory Optimised Tables

5. Memory Optimised Tables
SQL Server Memory Pools – MOT aggression
Stable
Storage
(MDF/NDF
Files)
Buffer Pool
Table
Data
Memory Internal Structures (proc/log
cache etc.)
SQL Server
Memory
Space

6. Create Resource Pool
CREATE RESOURCE POOL mem_xtp_pool
WITH ( MAX_MEMORY_PERCENT = 50,
MIN_MEMORY_PERCENT = 50
);
ALTER RESOURCE GOVERNOR RECONFIGURE;
EXEC sp_xtp_bind_db_resource_pool 'xtp_demo', 'mem_xtp_pool';
ALTER DATABASE xtp_demo SET OFFLINE WITH ROLLBACK IMMEDIATE;
ALTER DATABASE xtp_demo SET ONLINE;
Best
Practice

7. Storage

8. CREATE DATABASE
ALTER DATABASE xtp_basics ADD
FILEGROUP xtp_demo_mod
CONTAINS MEMORY_OPTIMIZED_DATA;
ALTER DATABASE xtp_basics ADD
FILE ( NAME = N'xtp_basics_mod1',
FILENAME = N'c:SQLDATAinmemxtp_basics_mod1' ,
MAXSIZE = 4GB),
( NAME = N'xtp_basics_mod2',
FILENAME = N'e:SQLDATAinmemxtp_basics_mod2' ,
MAXSIZE = 4GB),
( NAME = N'xtp_basics_mod3',
FILENAME = N'c:SQLDATAinmemxtp_basics_mod3' ,
MAXSIZE = 4GB)
TO FILEGROUP xtp_demo_mod;
go

9. Multiple Containers – Load Balancing
• Specify an odd number of Files in the File Group
• CFP {Data and Delta files} allocated in round robin
• If only two files – Data will always be on “1” and Delta on “2”

10. Life of a Row
Memory
CFP
(Data / Delta)
CFP
(Data / Delta )
No active rows
3. MERGE
4. GARBAGE
COLLECT
2. CHECKPOINT
1. Write to storage LDF – offline
checkpoint worker writes to
CFP (512MiB written to
log/tran is bigger than CFP
size)
2. Close CFP and mark ACTIVE
(Durable Checkpoint
established)
3. ACTIVE CFP’s with >= 50%
free space can be merged.
4. Files with no active rows can
be deleted

11. Offline Checkpoint Worker
• After 512MiB data written to log (from all DB activity) or Manual
CHECKPOINT
• CFP will become ACTIVE if amount of data written in single transaction
warrants it
• CFP state: UNDER CONSTRUCTION (on recovery, data taken from
transaction log)
• On CHECKPOINT
• UNDER CONSTRUCTION CTP closed, becomes ACTIVE
• Now have a durable checkpoint (otherwise use the log)
• Reiterate the need for Odd containers – read from container ‘A’ and write
new CFP to container ‘B’

12. CFP
UNDER
CONSTRUCTION
LDF
ACTIVE
CHECKPOINT
No new data except
deletes to tombstone
file
ACTIVE
ACTIVE
MERGE TARGET ACTIVE
MERGED
SOURCE
MERGED
SOURCE
Change status
Change status

13. CFP
MERGED
SOURCE
IN TRANSITION
TO TOMBSTONE
TOMBSTONE
Log backup LSN above
file – file no longer required
Waiting on the FILESTREAM
garbage collector

14. DEMO
• CFP.sql
sys.dm_db_xtp_checkpoint_files

15. Multi-Version Concurrency
Control

16. Multi-Version Concurrency Control (MVCC)
• SNAPSHOT isolation
• Update is actually INSERT and Tombstone
• Row Versions are kept in memory
• Compare And Swap replaces Latching
• Each in-memory row has a Start and End timestamp (on row header)
• Each database has:
• xtp_transaction_id counter (increment on BEGIN TRAN, reset on SQL restart)
• Global Transaction timestamp (used on COMMIT)

17. Multi-Version Concurrency Control (MVCC)
• SNAPSHOT isolation
• Update is actually INSERT and Tombstone
• Row Versions are kept in memory
• Compare And Swap replaces Latching
• Memory Garbage Collection cleans up versions no longer required (stale
data rows)
• Versions no longer required determined by active transactions – may be
inter-connection dependencies
• Your in-memory table can double, triple, x 100 etc. in size depending on
access patterns (data mods + query durations)
• Row chains can expand dramatically and cause really poor performance

18. SNAPSHOT isolation
Newest
Oldest
TS_START TS_END Connection
2 NULL 55
7 NULL 60
50 NULL 80

19. SNAPSHOT isolation
Newest
Oldest
TS_START TS_END Connection
2 10 55
7 NULL 60
50 NULL 80
Tombstone

20. SNAPSHOT isolation
Current
TS_START TS_END Connection
2 NULL 55
7 NULL 60
50 NULL 80
Tombstone

21. MVCC Summary
• Scalability achieved because MVCC removes need for latching and
logging (SNAPSHOT isolation)
• Row Versions (BeginTS, EndTS) are the main stay
• For UPDATES’s the biggest issue is because of MVCC
• Row Versions!
• Garbage collection can be slow to clean up stale rows
• In-Memory is not for large UPDATE’s

22. HASH Indexes

23. Hash Indexing
bucket_position = f hash( data )
Hash is deterministic (same value in gives same value out)

24. Hash is used as bucket position in array of
Memory Pointers aka the Hash Index
Bucket
position

25. Hash Collisions
• Example:1 million unique values hashed into 5000 hash values?
• Multiple unique values mush hash to the same hash value
• Multiple key (real data) values hang off same hash bucket
• Termed: Row Chain
• Chain of connected rows
• SQL Server 2014 Hash function is Balanced and follows a Poisson
distribution – 1/3 empty, 1/3 at least one index key and 1/3
containing two index keys

26. BUCKET_COUNT options for HASH index
1024 – 8 kilobytes
2048 – 16 kilobytes
4096 – 32 kilobytes
8192 – 64 kilobytes
16384 – 128 kilobytes
32768 – 256 kilobytes
65536 – 512 kilobytes
131072 – 1 megabytes
262144 – 2 megabytes
524288 – 4 megabytes
1048576 – 8 megabytes
2097152 – 16 megabytes
4194304 – 32 megabytes
8388608 – 64 megabytes
16777216 – 128 megabytes
33554432 – 256 megabytes
67108864 – 512 megabytes
134217728 – 1,024 megabytes
268435456 – 2,048 megabytes
536870912 – 4,096 megabytes
1073741824 – 8,192 megabytes

27. Link from Hash Index to first row in Row
Chain

28. Memory Optimised Table Row Header
Up-to 8 Index
Memory Pointers
• Above is a 3 row – row chain, all three rows hash to bucket #3
• Most recent row is first row in chain
• Rows in stable storage
• accessed using files, pages and extents
• Rows in memory optimised
• accessed using memory pointers
• Timestamp required for MVCC
Memory
Location
Hash Index

29. DEMO
• VISUALISE ROW CHAIN.sql

30. Row Header – Multiple Index Pointers
Data Rows (Index #‘x’ Pointer is next memory pointer in ‘row’ chain)
Memory Pointers
per Hash Bucket
Row chain is
defined
through the
Row Header

31. Hash Index Scan (Table Scan)
• Scan follows Hash Index (8 byte per
bucket)
• Jump to each first row in row chain
• Read row chain (lots of bytes per
row – header + data)

32. Hash Index - thoughts
• Base BUCKET_COUNT on cardinality of the column
• Don’t use where low cardinality – high rows: gives large row chains –
performance will suck
• Equality queries only
• Good candidates are FK on joins

33. What makes a Row Chain grow?
• Hash collisions
• Row versions from updates
• An update is a {delete, insert}
• Late garbage collection because of long running transactions
• Garbage collection can’t keep up because of box load and amount to
do {active threads do garbage collection on completion of work}
• Deleted rows may still be in memory for some time even though no
other connections have a snapshot

34. Affect of Row Chains on Memory
• MEMORY RESOUCES.sql

35. Range Index (Bw Tree ≈ B+Tree)
8KiB
1 row tracks a page
in level below
Root
Leaf
Leaf row per table row
{index-col}{pointer}
B+ Tree

36. Range Index (Bw Tree ≈ B+Tree)
8KiB
1 row tracks a page
in level below
Root
Leaf
Leaf row per table row
{index-col}{pointer}
B+ Tree

37. Range Index (Bw Tree ≈ B+Tree)
Bw Tree
8KiB
Root
Leaf
Leaf row per unique value
{index-col}{memory-pointer}
Row Chain
• Nodes 1 row – 8KiB
• Leaf is pointer to Row
Chain
• Nodes don’t change –
add/merge

38. Range Index (Bw Tree ≈ B+Tree)
8KiB
1 row tracks a page
in level below
Root
Leaf
Leaf row per table row
{index-col}{pointer}
B+Tree
8KiB
Root
Leaf
Leaf row per unique value
{index-col}{memory-pointer}
Row Chain
≈
Bw Tree

39. Index recommendations
• Use HASH when
• Mostly unique values
• Query uses expressions “Equality” i.e. = or !=
• Set BUCKET_COUNT appropriately
• Seriously think if you are considering composite index
• Use RANGE when
• Expressions are not Equality e.g. < > between
• Composite Index
• Low Cardinality
Best
Practice

40. Monitoring/DMV’s
• DMV.sql

41. Q & A / Discuss – uses???
• Feedback Forms