Fabricio CataeAdvanced Memory

Agenda: SQL MemoryServer Memory

The concept of Committed Memory

Memory Manager Distribute and Balance the memory among the processes

StrategyUse Page File for pages rarely usedUse RAM for pages frequently used

Windows Memory Manager

Physical Memory

Monitoring (Windows)

ProcessC.exe

ProcessA.exeProcessB.ex

e

ProcessD.exe

Physical Memory

Monitoring (Windows)

ProcessC.exe

ProcessA.exe

ProcessB.exe

ProcessD.exe

Committed Memory = RAM

Working Set(100%)

Working Set (30%)

Memory paged out(Pagefile)

Physical Memory

SQL Server Response

ProcessC.exe

ProcessA.exe

SQLSERVR.EXE

(SQL Server)

ProcessD.exe

Committed Memory = RAM

Working Set(100%)

Working Set (30%)

This is BAD!!!

A significant part of sql server process memory has been paged out. This may result in a performance degradation. Duration: 315 seconds. Working set (KB): 660338, committed (KB): 2201296, memory utilization: 30%.

http://support.microsoft.com/kb/918483

Maximize Throughput for Network Applications

Windows Memory Manager prefers trimming the System Cache prior to the Processes

Lock Pages in MemoryAllows SQL Server to allocate physical memoryEnable AWE – Use direct memory allocationLarge Pages – Use 64-kb block size (TLB cache)Windows cannot use the Page File

Windows Settings

In order to achieve maximum performance, SQL Server memory manager tries to use all the available RAM

SQL Memory UsageGrows if there is “Available Memory”Shrinks if the OS indicates “Low Memory”

Total memory is controlled using a Target

SQL Server Memory Manager

Target

Physical Memory

SQL Server uses all Available MemoryTarget Memory

ProcessC.exe

ProcessA.exeSQL Server

ProcessD.exe

Target

SQL Server

Database Cache

Access Manager

Buffer Manager

Memory Disk

Allocates the memory per NUMA Node

What is NUMA?

Buffer Manager

Hardware Performance BottleneckSymmetric Multiprocessor (SMP)

Node(MemoryNod

e)

Design solution for hardware scaling

MemoryLocalForeign

Non-Uniform Memory Access (NUMA)

Allocates the memory per NUMA Node

SQL Server organizes the buffers and tries to affinitize the tasks with the local memory

Fast = Local MemorySlow = Foreign memory

Foreign access is still faster than Disk times

Buffer Manager

Allocated memory from SPA is called “Stolen”

Provides an internal method for memory allocation (limited to 8Kb)

Example:

Single Page Allocator (SPA)

SELECT type, memory_node_id, single_pages_kbFROM sys.dm_os_memory_clerks

Buffer UsageDatabase cacheSingle Page allocatorMinimum Free pages

Buffer Manager Distribution

DBCC MEMORYSTATUS

Who stole my memory?Server Memory

(Physical Memory)

Single Page Allocator

Multi Page Allocator

Virtual Allocator

Shared Allocator

MemClerkMemClerkMemClerkMemClerkMemClerk

Compilation Procedure Cache

Allocators

MemClerk

MemClerk

MemClerk

MemClerk

SPA MPA VM SMSingle Page Allocator

Multi Page Allocator

Virtual Memory

Sharedsys.dm_os_memory_clerks

Stolen “Memory To Leave” (MTL)

Every Memory Clerk reportsVirtual Memory (VM)Shared Memory (SM) SinglePage allocatorMultiPage allocator

DBCC MEMORYSTATUS

DBCC MEMORYSTATUS

CACHESTORE_SQLCPChanges in Caching Behavior between SQL Server 2000, SQL Server 2005 RTM and SQL Server 2005 SP2http://blogs.msdn.com/b/sqlprogrammability/archive/2007/01/22/3-0-changes-in-caching-behavior-between-sql-server-2000-sql-server-2005-rtm-and-sql-server-2005-sp2.aspx

USERSTORE_TOKENPERMA gradual increase in memory consumption for the USERSTORE_TOKENPERM cache store occurs in SQL Server 2005http://support.microsoft.com/kb/933564

USERSTORE_SCHEMAMANAGERMemory consumption in MEMORYCLERK_SOSNODE and in USERSTORE_SCHEMAMANAGER may cause a performance slowdown of a SQL Server 2008 databasehttp://support.microsoft.com/kb/959767

Known Issues

Balance the memoryCompilation and OptimizationTask ExecutionGeneral Caching mechanismDatabase Cache

All memory allocation is reported into one specific Memory Clerk. This allows finding which resource is consuming the resources.

Memory Contention

Stolen Memory

DBCC MemoryStatus (Compilation)

Limits the number of concurrent compilation

Very LARGE COMPILATION = 1Simple query = 4 queries per CPU

If the number of compilation exceeds the limit, then the following tasks are suspended

RESOURCE_SEMAPHORE_QUERY_COMPILE

Compilation Gateways

DBCC MemoryStatus (Workspace)

Session memory allocated for query execution in order to process

HashSortParallelism

Limits the number of concurrent query execution

RESOURCE_SEMAPHORE

Workspace Memory

TypesGeneric ClerksSpecialized Caches

User StoreCache Store

Object Pools

Examples:MEMORYCLERK_SQLCLRUSERSTORE_OBJPERMCACHESTORE_OBJCPCACHESTORE_SQLOBJECTSTORE_SNI_PACKET

Memory Clerks

Types (sys.dm_os_memory_clerks)Generic ClerksSpecialized Caches

User Store (sys.dm_os_memory_cache_counters)Cache Store (sys.dm_os_memory_cache_hash_tables)

Object Pools (sys.dm_os_memory_pools)

Examples:MEMORYCLERK_SQLCLRUSERSTORE_OBJPERMCACHESTORE_OBJCPCACHESTORE_SQLOBJECTSTORE_SNI_PACKET

Memory Clerks (Caches)

Who stole my memory?Server Memory

(Physical Memory)

Single Page Allocator

Multi Page Allocator

Virtual Allocator

Shared Allocator

Buffer PoolStolen

Single Page Allocator

“MemToLeave”

MemClerkMemClerkMemClerkMemClerkMemClerk

Who stole my memory?Server Memory

(Physical Memory)

Single Page Allocator

Multi Page Allocator

Virtual Allocator

Shared Allocator

MemClerkMemClerkMemClerkMemClerkMemClerk

Procedure Cache SQL CLR

Buffer PoolStolen

Single Page Allocator

“MemToLeave”

AgendaSQL Memory TopicsSQL Latches

SQL OSMemory

Management

Thread Scheduling

Rubén GonzálezLatches

A device for holding a door, gate, or the like, closed, consisting basically of a bar falling or sliding into a catch, groove, hole, etc.

In SQL Server a latch is a flag that protects a memory structure that could be accessed at the same time by other thread.

What is a Latch?

Was introduced in SQL 7.0 as a consecuence of Row Level LockingYou can’t control latches (no latch hint)Latchs are at memory structure level not at row level. There are:

Buffer Latches IO related Not IO related

Transaction Latches Other Latches

Latch facts

Factors under your control can affect

visibility and behavior

Why do I care?You will see latched

in…

Your Application

Your Design and Configuration

Throughput of IO Subsystem

sys.dm_os_wait_stats

sys.dm_os_latch_stats

sys.dm_exec_requests

sys.dm_os_waiting_tasks

XEventsPerformance

CountersError (Latch

Timeout)

OTHER LATCHLATCH_NLLATCH_KPLATCH_SHLATCH_UPLATCH_EXLATCH_DT

Braekdown of Latch WaitsBUFFER LATCH

PAGELATCH_NLPAGELATCH_KPPAGELATCH_SHPAGELATCH_UPPAGELATCH_EXPAGELATCH_DTPAGEIOLATCH_NLPAGEIOLATCH_KPPAGEIOLATCH_SHPAGEIOLATCH_UPPAGEIOLATCH_EXPAGEIOLATCH_DT

TRANSACTION LATCHTRAN_MARKLATCH_NLTRAN_MARKLATCH_KPTRAN_MARKLATCH_SHTRAN_MARKLATCH_UPTRAN_MARKLATCH_EXTRAN_MARKLATCH_DT

Used for synchronization of

commits with marked

transactions

The breakdown of these is in

sys.dm_os_latch_stats

+0

Ok, but what is it really?

+8+16+24 class

countownerwaiters List of Waiters

Task Address of Owner

This is the latch

Latch Class (latch id)

Num KP Stuff Spin Wait DT EX UPNum

SH

0123453563 111236

Count of KP Latches

Count of SH Latches

Superlatch and other stuff

spinlock Waiters exist

DT Latch held

EX Latch held

UP Latch held

x64 C++ Class

Just reading. No ownership tracking

Readers are ok but no writes

It’s all mine, but I’m tracked

Destroy latch; Heaviest of all. Ex. Removing buffer from

cache.

Keep latch; Very light; just keeping reference count

Latch ConcurrencyKP S

HUP

EX DT

KP Y Y Y Y NSH

Y Y Y N N

UP

Y Y N N N

EX Y N N N NDT N N N N N

Y turns N if

waiters

Mode

Not always FIFO

KP

SH

UP

EX

DT

The BUFFER (BUF) latchBUF

SH for Reading

UP and EX for writing

KP for traversing

DT for freeing

PAGELATCH_XX or

PAGEIOLATCH_XX

Buffer latch in AM

PAGE (in cache)

BUF_ONLRU

0x000001

BUF_DIRTY 0x000002BUF_IO 0x000004BUF_HASHED

0x000008

BUF_READ 0x000400BUF_IOERR 0x000800

bpagebpageno

bstat

blatch

blog

m_headerversion

m_page_id

ptr

value

Latch

Class

Why not just reuse the BUF lacth concept?

Shows up as LATCH_XX in DMVs

Wait_resource is the type (class) and mem address of latch

Sys.dm_os_latch_stats to see all 144 of them

Other wait types are not latches (Ex. ASYNC_NETWORK_IO, CX_PACKET, ...)

Let’s talk Non-BUF Latches

Why should you care?

Normally you

wouldn’t because…

Many of the 144

classes you will never

see

But what if thousands

of waits with high avg wait

time

Or the latch shows up as a big waiter or blocker in

your monitoring

Or lot of waits that cumulative

slowing down a query

Nesting Transactions: RCSI and parallel queriesFGCB_ADD_REMOVE: Autogrow of Data FilesLOG_MANAGER: Autogrow of tlogACCESS_METHODS_DATASET_PARENT: Only seen during parallel queriesALLOC_FREESPACE_CACHE: Partition your tableALLOC_EXTENT_CACHE: Partition your tableTRANSACTION_DISTRIBUTED_MARK: Also appears as TRAN_MARKLATCH. Distributed Transactions.

Ones to look out for

Could be an issue when in the seconds rangeNot always a high count for a problem

Are latched causing you problems

PAGEIOLATCH waits = I/O Bottleneck

High BUF latch mean hot pages

Non-BUF latch waits depend on the latch class

Latch Timeouts

What is the avg disk

transfer rate

Unless a bug it is usually allocation pages (and usually tempdb)Could be system table pages for a DDL

Find out the “hot” class and see if one of generally know issues

This is just bad no matter what type (5 minute wait!)

Check sys.dm_db_index_operational_stats

© 2011 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.

The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after

the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.