© Demand Technology, Inc. Windows Server 2003: An Update Demand Technology Software 1020 Eighth Avenue South, Suite 6, Naples, FL 34102 phone:

© Demand Technology, Inc.

Windows Server 2003: An Update

Demand Technology Software1020 Eighth Avenue South, Suite 6, Naples, FL 34102

phone: (239) 261-8945 fax: (239) 261-5456 e-mail: [email protected]

http://www.demandtech.com

2©Demand Technology, Inc. Windows 2003 Performance and Tuning: Introduction

Outline

Windows Server 2003 Overview Support for big Iron

64-bit processors ccNUMA architecture multiprocessors

LRU-based memory management WSRM policy-based performance manager New web services application processing

architecture (IIS 6.0 Web Gardens) Assessment


Big Iron support

0

10

20

30

40

50

60

70

80

90

100

10:00 10:30 11:00 11:30 12:00 12:30

perc

ent

Processor Utilization Breakdown (NT)11/19/2003 09:45 - 11/19/2003 12:51

privileged user interrupt DPC


Windows 2003 architecture

No major architectural changes processr.sys support for processor-specific

optimizations ccNUMA support New LRU-based memory management

Identical 32 and 64-bit versions Multiple service host (svchost)

process address spaces For purposes of extended security granularity



Multiple service host (svchost) process address spaces How to identify them

C:\>tasklist /svc /fi "Modules eq srvsvc.dll"



Multiple service host (svchost) process address spaces How to identify them

Module identification

function:


Executive

User

Privileged

User Process

WIN32K.SYS

VideoDriver

ntoskrnl.sys

hal.sys

processr.sys

I/O Manager

DeviceDrivers

NDISDriver

IP

TCP/UDP

NBT

FileSystemDrivers

Services(svchost.

exe)

Server/RedirectorObject

ManagerProcess & Threads

VirtualMemoryManager

SecurityLocal

ProcedureCall

SecuritySubsystem

EncryptionSubsystem

Win32Subsystem

I/O Manager stack

Network protocol stack

HARDWARE

PnP/PowerMgmt

Winlogon

ServiceController


Windows evolution

Common NT code base now in use across all Microsoft platforms Windows XP for workstations Windows Server 2003 Windows CE Embedded Windows xBox???

Which allows .NET Framework programs to execute everywhere


.NET Framework (for applications)

Win32COM+ MSMQ AD WMI

AS

P.N

ET

Form

s

XM

L, S

OA

P

Common Language Runtime (CLR)

Enterprise Services (Clustering)ADO.NET

XML SupportOther services

.NET Languages


Windows 2003 evolution: new in WinXP

New Processor Object Counters

Supports Intel SpeedStep processor that run at different speeds to save power

Disk prefetching

Designed to speedup the boot process and program image file loading

System virtual memory changes

Overcome a variety of “large system” deficiencies, most noticeable with large Terminal Services environments: Registry size limits, accessing very large mapped files, very large device drivers

Volume shadow copy service

Supports file Cache flush and freeze so that volume snapshots can be easily created with integrity


Windows 2003 evolution: new in Win2003

64-bit support IA-64 and AMD-64 (SP1) support; much work remains to port all the server applications

LRU memory management

Page trimming performed based on the age of a working set page

System virtual memory changes

Generic X’ffff ffff’ setting (-1) for PagedPoolSize, NonPagedPoolSize

WSRM Policy-based performance management: dispatching priority, processing, working set

Disk Counters always on

Both Logical and Physical Disk performance counters are always enabled


Windows 2003 evolution: new in Win2003

Processr.sys driver

Hardware-dependent logic moved to special kernel mode driver: IA-32, hyperthreading, power-management, IA-64, AMD-64

Transition Pages Re-purposed/sec

New memory management counter fills a gap in the measurement data

Kernel-mode web services cache

New IIS 6.0 architecture also features Web Gardens support for ASP and ASPX applications

TCP/IP v6 support TCP/IP v4 and v6 can run side by side

Security enhancements


Beyond Windows 2003

Longhorn Preview Production versions no slated to ship until 2006

or 2007 Workstation-oriented UI changes Designed around 64-bit machine requirements New WinFS file system

Comprehensive “flat” view, alongside traditional tree view

New performance monitoring interface gradually will replace Performance Library DLLs


Windows Server 2003 Extended Processor Support

Hardware Abstraction Layer (HAL) Process & Thread Context switching Interrupt processing Synchronization primitives SMP processor signaling Virtual memory addressing

Processr.sys CPU-specific optimizations, e.g.,

ccNUMA Hyperthreaded Processors Power-conserving portables

Ntoskrnl.exe

HALProcessr.sys


Windows Server 2003 Extended Processor Support

32-bit Intel Currently @ 3 GHz Power-conserving Pentium 4 M Both conventional & HT multiprocessors New, improved measurement interface

Support is currently available only in Intel vTune

64-bit Intel ccNUMA support

AMD-64 Expected in Service Pack 1


Intel 686 microarchitecture

Increased execution parallelism; remove some of the instruction sequence

dependencies Instructions are translated into RISC

micro-instructions Pool of 40 GP pseudo-Registers Micro-ops can be executed out of

sequence Executed instructions are retired


Intel 686 micro-

architecture

(single threaded,

superscalar)

Instruction Fetch

Decoded Instruction QueueRegister

AllocationTable

ExecutionTraceQueue

Instruction Decode

Micro-operationReservation Station

Reorder Buffer

Execution Unit

Execution Unit

Execution Unit

Execution Unit

Execution Unit

InO

rder

Ret

irem

ent

TranslationLookaside

Buffer

InstructionCache Data

Cache

Ret

ired

Inst

ruct

ion

s


Pentium 4 Hyperthreading

Two instruction streams scheduled to execute concurrently in the same pipeline.

When one instruction stream stalls, instructions from the other thread can be executed.

Implementation: External interface is replicated Internal resources are

partitioned and/or shared

Instruction Fetch

Decoded Instruction QueueRegister

AllocationTable

ExecutionTraceQueue

Instruction Decode

Micro-operationReservation Station

Reorder Buffer

Execution Unit

Execution Unit

Execution Unit

Execution Unit

Execution Unit

InO

rder

Ret

irem

ent


Buffer

InstructionCache Data

Cache

Ret

ired

Inst

ruct

ion

s

CPU 0 CPU 1

CPU 0 CPU 1

CP

U 0

CP

U 1

CPU 0 CPU 1


Partitioned

Pentium 4 Hyperthreading

Inst

ruct

ion

Fet

ch

RegisterAllocation

Table

ExecutionTraceQueue

Inst

ruct

ion

Dec

ode

Reo

rder

Bu

ffer

Exe

cuti

on

Un

it

Exe

cuti

on

Un

it

Exe

cuti

on

Un

it

Exe

cuti

on

Un

it

Exe

cuti

on

Un

it


Buffer

InstructionCache

DataCache

RetiredInstructions

CP

U 0

CP

U 1

CP

U 0

CP

U 1

CPU 0 CPU 1

CP

U 0

CP

U 1

InOrderRetirement

Mic

ro-o

pera

tion

Res

erva

tion

Sta

tion

Dec

oded

Inst

ruct

ion

Que

ue

Duplicated


Pentium IV Hyperthreading

Two simultaneously executing threads can saturate internal execution pipeline resources!

Windows 2003 support A new Win32 GetLogicalProcessorInformation API call

retrieves information about logical processors and related hardware.

Scheduler support: spread the load across physical processors first, then schedule logical processors when physical processors are all “busy”

HALT the processor in Idle mode processr.sys function


Intel 786 IA-64 architecture

Itanium-2: 1.6 GHz and higher EPIC: Explicitly Parallel Instruction Execution

Explicit parallelism (up to 3 instructions in a bundle) Predication Speculation Massive Resources

Extended Register sets

New instruction set! Requires compilers that can take advantage

of the architecture


AMD-64

AMD’s 64-bit extension to IA-32 8 additional GPRs in

Long Mode Microarchitecture

similar to the IA-32 Native IA-32

instruction execution in Short Mode

Instruction Fetch

8 EntryScheduler


Buffer

InstructionCache

Decode 1

Decode 2

Decode 1

Decode 2

Decode 1

Decode 2

Pack

Decode

Pack

Decode

Pack

Decode

8 EntryScheduler

8 EntryScheduler

36 EntryScheduler

ALU AGU ALU AGU ALU AGU FADD FMUL FMISC

Pick

BranchPrediction

Level 1 Data Cache

Dispatch


Symmetric Multiprocessing (SMP)

RAM

Processor

Cache

Processor

Cache

Processor

Cache

Processor

Cache

Memory BusPeripheral Bus


Multiprocessing scalability

Queued spin lock support During spin lock execution, micro-ops are

discharged into the instruction execution pipeline faster than they can be executed

Leads to resource shortages When the spin lock test finally succeeds,

large portions of the pipeline have to be flushed


Multiprocessing scalability

Queued spin lock support Uses PAUSE instruction

Slows down loop execution slightly to a rate that is synchronized to memory bus access so that

Allows the processor to detect immediately a change in the value of the loop synchronization variable

KeAcquireInStackQueuedSpinLockAtDpcLevel, KeReleaseInStackQueuedSpinLockFromDpcLevel


A ctual vs. Ideal Mult iprocessor Scalability

16,263

24,925

30 ,231

57,015

0

10,000

20,000

30,000

40,000

50,000

60,000

70,000

0 1 2 3 4 5 6 7 8

# of processors

TP

C-C

Tra

nsa

ctio

ns/

sec

Vendor AIdealVendor BIdealLog. (Vendor A )


A ctual vs. Ideal Mult iprocessor Scalability

16,26324,925

30 ,231

57,015

0

30,000

60,000

90,000

120,000

150,000

0 4 8 12 16 20 24 28 32

# of processors

TP

C-C

Tra

nsa

ctio

ns/

sec

Vendor AIdealNonlinear t rendline


ccNUMA Support

Cache-coherent Non-uniform Memory Access Consist of Multiprocessor nodes

Processors (usually 4) Local memory (shared memory bus) Memory controller (for remote access)

Introduces another level of cache coherence Memory accesses are non-uniform when comparing Next Level

Cache hits to main memory references Remote memory latency is 3-5 times slower than local memory

access Overcomes the congestion problems of a single bus

architecture if there is sufficient node affinity in the workload


ccNUMA Support

Threads scheduled on their ideal node

Real memory allocated locally, whenever possible Maintains multiple

Pools & Available Bytes queues

Node 0

CPU 0

Cache

CPU 2

Cache

CPU 1

Cache

CPU 3

Cache

Node 2

CPU 0

Cache

CPU 2

Cache

CPU 1

Cache

CPU 3

Cache

Lo

cal M

emo

ry

Node 1

CPU 0

Cache

CPU 2

Cache

CPU 1

Cache

CPU 3

Cache

Node 3

CPU 0

Cache

CPU 2

Cache

CPU 1

Cache

CPU 3

Cache

Lo

cal M

emo

ry

Remote Memory

Controller

Cro

ssba

rS

witc

h


ccNUMA Support

Traditionally, the drawback to NUMA was that applications had to be re-coded to take full advantage of the architecture

Performance is sensitive to the long latency associated with remote memory access

API Function Description

GetNumaAvailableMemoryNode

Retrieves the amount of memory available in the specified node.

GetNumaHighestNodeNumber

Retrieves the node that currently has the highest number.

GetNumaNodeProcessorMask

Retrieves the processor mask for the specified node.

GetNumaProcessorNode Retrieves the node number for the specified processor.


May be the only way to optimize large, n-way multiprocessors e.g., HP Superdome, Unisys ES7000 16-64

processor machines

But it requires commitment! Understanding your current workload CPU

processing requirements Continuous monitoring of the workload on a per

processor basis Ensure that excessive numbers of threads from

“loved ones” are not in the Ready Queue Periodic review of the partitioning scheme

Multiprocessor partitioning


The workload must be concentrated enough on its dedicated CPUs so that it will benefit from cache warm starts, but not too concentrated that it causes excessive processor queuing.


10 0

90

80

70

60

50

4 0

30

20

10

0CP U 1 CP U 2 CP U n


The interrupt workload should not be so concentrated on its dedicated CPUs that interrupt processing is subject to excessive interrupt pending time delays

Between 5-30% concentration for % Interrupt Time is probably ideal


10 0

90

80

70

60

50

4 0

30

20

10

0CP U 1 CP U 2 CP U n



Reskit Interrupt Filter tool:



WSRM policies:



WSRM policies: Caution:

Not designed to be used with any application that performs its own partitioning

If an app sets a Processor Affinity mask, WSRM will honor it!




Win64 Virtual Memory

Architectural component 64 bit 32 bit

Virtual Memory 16TB 4GB

Paging File Size 512 TB 64 GB

Hyperspace 8 GB 4 MB

Paged Pool 128 GB 470 MB

Non-paged Pool 128 GB 256 MB

System cache 1 TB 1 GB

System PTE (page table entries) 128 GB 660 MB


Page replacement

Windows 2003 uses a form of the popular LRU page replacement algorithm Pages in process working sets are aged using the

access bits maintained by the hardware Older pages without their access bits set are

“trimmed” first Recently trimmed pages are retained in the Standby

list (aka, the page cache) and returned to the process working set via transition faults.

Eventually, older trimmed pages on the Standby List are “re-purposed” and place on the Free List


Page replacement

FreedPages

TransitionFaults/sec Standby

List

FreeList

ZeroList

TrimmedDirty

PagesTrimmedPages

ModifiedList

(Dirty pages)

Modified Page Writer(Pages Output/sec)

Re-purposedTransition Pages

Zero PageThread

Demand ZeroPage Faults/sec

Process Working sets (Working set bytes)

Available Bytes

Working set Page Aging (LRU)

PagesInput/sec


Page replacement

Age of a Page Most Recent: access bit is set Older: Older access bit set last page trimming scan Even older: Not accessed at the time of the last scan Oldest: Recently trimmed page marked in “transition”

that are kept in the Standby List (page cache)

Page trimming working set scans Threshold-driven: the actual rate is not reported Efficiency is very important! Only enough pages to replenish the Standby List are

trimmed


Page replacement

Available Bytes = (free + zero + standby) Windows Server 2003 utilizes RAM much

more efficiently than previous versions of the OS Previously, the only way pages were aged was using

the transition fault mechanism

Applications which attempt to allocate as much real memory as they can get require special consideration MS Exchange MS SQL Server


Page replacement


SQL Server memory tuning

For processes that perform their own working set management: CreateMemoryResourceNotificati

on to create a memory resource notification object

Sends two events to “listening” processes:

LowMemoryResourceNotification

HighMemoryResourceNotification


IIS memory tuning

IIS 6.0 memory caching Kernel cache (new)

http Responses

System File Cache .htm, .jpg, .gif files, etc.

IIS Object cache File handles

Active Server Pages Script engines

Templates

Kernel Inetinfo.exe

w3wp.exe

HTTP.SYS

http GET Request

TCP/IPStack

Response Object Cache

File Cache

ASPASPX

Request

Thread Pool

Thread Pool

Web Gardens


IIS memory tuning

Monitoring IIS 6.0 memory caching Kernel cache (new)

Web Server Cache object

System File Cache Cache object (MDL interface)

IIS Object cache IIS Global object

Active Server Pages Active Server Pages


IIS memory tuning

Tuning parameters for controlling IIS memory caching: MaxCachedFileSize – defaults to 256 KB MemCacheSize – defaults to dynamic sizing ObjectCacheTTL - default is 30 seconds

objects include File handles, Directories .htm, .gif, and .jpg files are cached in the system

file cache

OpenFilesInCache - default is 1000 per 32 MB (obsolete)


IIS ASP memory tuning

Tuning parameters for controlling IIS Active Server Pages memory caching: AspTemplateCache – defaults to 250 files in IIS 5.0

AspScriptEngineCacheMax – defaults to 120 Allow each processing Thread to cache its script engine

AspScriptFileCacheSize – defaults to 500 Allow each processing Thread to cache its precompiled

script engine

Monitor Active Server Pages: ASP Templates Cached, Template Cache Hit Rate, Free Script Engines in Cache


New Win 2003 tools

Resource Kit tools are now supported! Consume Kernrate Poolmon

Windows NT kernel Measurement Interface (Trace) reports


Win 2003 performance monitoring

NTSMF support Configuration file

Analogous to Disable Performance Counters Registry flag Use it with Perflib DLLs that have persistent problems and

pernicious side-effects– .NET Framework perflibs– Lotus Notes Perflib

Logical & Physical Disk counters are always on!

IPv6 and TCPv6 support Discourage use of WMI interface


Priority Queuing/Preemptive Scheduling

Windows timer services Internal processor clock (MHz) is not accessible External clock hardware generates timer interrupts

Clock interrupt every “quantum” GetSystemTimeAdjustment API call

Virtual system clock 100 nanosecond units, SetTimer API Ticks occurs every 10 ms.

High precision timer APIs to access external clock directly

QueryPerformanceCounter and QueryPerformanceFrequency


Processor performance monitoring

Each clock interval, the clock external interval routine determines what thread was executing. CPU Time is sampled and then accumulated in a

Counter at the Thread and Process level Several hundred samples are gathered every second How accurate is this sampling?

A System Idle Thread is dispatched when there are no Ready Threads Processor Busy = 100% - Idle Thread CPU Time Not a true thread: processr.sys hardware-dependent

implementation


What is the relationship between processor utilization and queuing?

“Interpreting NT Queue Length Measurements,” by Yiping, Bolker & Yefim, BMC Software, CMG 2002 Research questions the validity of the

System:Processor Queue Length Counter At low utilization, the Processor Queue

Length can be a relatively higher number than queuing theory predicts


Interpreting NT Queue Length Measurements

Validity of the System:Processor Queue Length Counter At low utilization, the observed Processor Queue Length can

be a relatively high number Instantaneous Counter (i.e., sampled value) Heisenberg Uncertainty Principle: measurement tools can get

in the way of the measurement datae.g., dmperfss collection thread always observed in the Running state Worst case results if you use a 1 second sample interval due to:

– single Timer queue – HAL virtualization of the system clock


Lab exercise. View the

STPSQL02PD.200311060200.sum.smf data file in PerfGal

Determine the relationship between Processor utilization and Processor queuing

Can you make a case for a CPU upgrade?





Plus, examine the Processes in the Ready state to determine if your Loved Ones are impacted…



Support & Resource Kit Tools

Required Documentation Trace collection and formatting tools

(logman, etc.) Debugger Interrupt Affinity Filter Tasklist Kernrate pfmon poolmon NT 4.0 Performance Monitor


Windows 2003 Time Service

Allows you to synchronize your servers to a common machine or to an authoritative external time source using SNTP e.g.,

net time /setsntp:”tock.usno.navy.mil ntp2.usno.navy.mil”net stop w32timew32tm –once

net start w32time

See KB Q216734 for more info.


Resource Kit tools

Miscellaneous measurement tools Pfmon.exe - page fault monitor Consume.exe - CPU & memory stress program Kernrate.exe - detailed CPU profiler Vadump.exe - process virtual address dump

Miscellaneous Counter tools Documentation

Counters.chm Regentry.chm


Kernrate profile

High frequency sampler in the Reskit Records processor utilization by process

virtual address includes instructions executing inside the operating

system kernel, the HAL, device drivers, etc.

e.g.kernrate -n smlogsvc -s 120 -k 100 -v 2 -a -e -z

pdh

See kernrate.doc for more information


Lab exercise. Start Sysmon log

manager gathering process and thread data to a .csv file

Monitor CPU usage using kernrate

Hint: prepare kernrate cmd in advance!


Start smlogsvc


kernrate

Starting to collect profile dataWill collect profile data for 120 seconds===> Finished Collecting Data, Starting to Process Results------------Overall Summary:--------------P0 K 0:00:03.094 ( 2.6%) U 0:01:19.324 (66.1%) I

0:00:37.584 (31.3%) DPC 0:00:00.200 ( 0.2%) Interrupt 0:00:00.991 ( 0.8%) Interrupts= 190397, Interrupt Rate= 1587/sec.


kernrate

Results for Kernel Mode:-----------------------------OutputResults: KernelModuleCount = 619Percentage in the following table is based on the Total Hits for the KernelTime 20585 hits, 19531 events per hit --------Kernel CPU Usage (including idle process) based on the profile interrupt total possible hits

is 33.50% Module Hits Shared msec %Total %Certain Events/Secntoskrnl 15446 0 120002 75 % 75 % 2513923processr 4381 0 120002 21 % 21 % 713032hal 524 0 120002 2 % 2 % 85283win32k 114 0 120002 0 % 0 % 18554Ntfs 42 0 120002 0 % 0 % 6835nv4_disp 22 0 120002 0 % 0 % 3580nv4_mini 12 0 120002 0 % 0 % 1953


kernrate

Results for User Mode Process SMLOGSVC.EXE (PID = 2120)OutputResults: ProcessModuleCount (Including Managed-Code JITs) = 41Percentage in the following table is based on the Total Hits for this ProcessTime 40440 hits, 19531 events per hit --------User-Mode CPU Usage for this Process based on the profile interrupt total possible hits is

65.82% Module Hits Shared msec %Total %Certain Events/Secpdh 31531 0 120002 77 % 77 % 5131847kernel32 6836 0 120002 16 % 16 % 1112597msvcrt 1700 0 120002 4 % 4 % 276684ntdll 362 0 120002 0 % 0 % 58917


kernrate

===> Processing Zoomed Module pdh.dll...----- Zoomed module pdh.dll (Bucket size = 16 bytes, Rounding Down) --------Percentage in the following table is based on the Total Hits for this Zoom ModuleTime 31671 hits, 19531 events per hit -------- (33371 total hits from summing-up the module

components)(51.55% of Total Possible Hits based on the profile interrupt) Module Hits Shared msec %Total %Certain Events/SecStringLengthWorkerA 23334 1 119992 69 % 69 % 3798056IsMatchingInstance 1513 0 119992 4 % 4 % 246269GetInstanceByName 1374 0 119992 4 % 4 % 223644IsMatchingInstance 1332 0 119992 3 % 3 % 216808NextInstance 1305 12 119992 3 % 3 % 212413GetInstanceName 829 0 119992 2 % 2 % 134935PdhiHeapFree 495 428 119992 1 % 0 % 80570GetInstance 469 11 119992 1 % 1 % 76338PdhiHeapReAlloc 428 0 119992 1 % 1 % 69665GetCounterDataPtr 338 338 119992 1 % 0 % 55015NextCounter 246 1 119992 0 % 0 % 40041GetInstanceByName 233 229 119992 0 % 0 % 37925FirstInstance 121 121 119992 0 % 0 % 19695PdhiHeapFree 121 0 119992 0 % 0 % 19695PdhiMakePerfPrimaryLangId 108 108 119992 0 % 0 % 17579


CIMOM CIMOM ((WinMgmt))

Management applicationManagement application

COM interfacesCOM interfaces

ProviderProvider(DLL)(DLL)

ProviderProvider(EXE)(EXE)

CIMCIM repositoryrepository

WMI architecture


ProviderProvider(EXE)(EXE)ProviderProvider(EXE)(EXE)


Management applicationManagement application


PerformancePerformanceExtensionExtension

DLLDLL

SystemSystemPerformancePerformance

DLLDLL

Windows NTPerformance

Monitor

System Monitor

graph control

Custom Performance Tool

RegQueryValueEx()

PerflibWMI

PDH.DLL

Hi-Perf Data Provider Object

Sysmon Log Service

Files

Sysmon log and alert service

SystemSystemPerformancePerformance

DLLDLLSystem

PerformanceDLL

PerformancePerformanceExtensionExtension

DLLDLL

System Monitor architecture

PerformanceExtension

DLL


Windows kernel trace facility

Secure kernel trace facility Architected for high volume

NT Scheduler trace abandoned due to overhead

Designed for application extensions But only Microsoft understands the format

of the Trace data well enough to create reporting applications.


Operating System Traces

Trace Header Process IdParent

Process IdSecurity Id Image Name

Process Start/End

4 bytes 4 bytes Variable length Variable length

Trace Header DiskSignature

TransferSize

Disk I/O

4 bytes 4 bytes

IRP Flags

4 bytes

Trace Header

Page Faults

Response

4 bytes

VirtualAddress

4 bytes

Byte Offset

8 bytes

Trace Header Source IpAddress

Dest IpAddress

TCP/IP

4 bytes 4 bytes

SourcePort

2 bytes

DestPort

2 bytes

Size

4 bytes

Process Id

4 bytes


Windows kernel event trace

Event trace headerField Description Data Type Size

Size Event record size unsigned long 4 bytesTimeStamp Wall clock time (in 100ns) quad integer 8 bytesThreadId Thread responsible for event handle 8 bytesGUID Globally Unique Identifier structure 16 bytesUserTime User-mode CPU time (ticks) unsigned long 4 bytesKernelTime Kernel-mode CPU time unsigned long 4 bytesVersion Version of record unsigned short 2 bytesType Event type unsigned char 1 byteLevel Event level unsigned char 1 byte



Resource Kit tools: Tracelog.exe Tracedmp.exe Reducer.exe

Third party tools ???


Lab exercise

View the Workload.txt Reducer sample



Reducer:

| WINDOWS 2000 Capacity Planning Trace || Version : 2128 || Type : Default |+-----------------------------------------------------------------------------------------------------------------------------------+| || Build : 2195 || Processors: 1 || Start Time: 24 Aug 2001 12:36:06.830 |______________________________________________________________________| || End Time : 24 Aug 2001 12:37:02.786 || Duration : 55 Sec || || Trace Name: NT Kernel Logger || File Name : C:\LogFile.Etl || Start Time: 24 Aug 2001 12:36:06.830 |______________________________________________________________________| || End Time : 24 Aug 2001 12:37:02.786 || Duration : 55 Sec || |+-----------------------------------------------------------------------------------------------------------------------------------+



Windows Event Trace Session Report---------------------------------------------------------- Window Build: 3700 Computer: IISPERFSRV Processors: 8 CPU Speed: 700 MHz Memory: 2048 Mb Trace Name: NT Kernel Logger File: S:\tools\etw\krnl.etl Start Time: Wednesday, October 23, 2002 9:23:53 AM End Time: Wednesday, October 23, 2002 9:24:13 AM Trace Name: IIS Trace File: S:\tools\etw\iis.etl Start Time: Wednesday, October 23, 2002 9:23:58 AM End Time: Wednesday, October 23, 2002 9:24:11 AM


Lab exercise

View the iisdata.xml Report sample





Evaluation: Process End event does provide process end

performance statistics (CPU, I/O, etc.) New per Process Disk I/O Counters satisfy

most capacity planning requirements The I/O trace namespace for capturing I/O

requests by file is formidable captures per process “hard” page faults TCP trace data vs. Network Monitor data


Windows Scheduler

WSRM controls Establish a workload’s

processor utilization target

WSRM will raise and lower process Dispatching Priority based on the management policy

Controls are effective only when there is processor contention!


MS Internet Information Server

Kernel Inetinfo.exe

w3wp.exe

HTTP.SYS

http GET Request

TCP/IPStack

Response Object Cache

File Cache

ASPASPX

Request

Thread Pool

Thread Pool

Web Gardens


Windows 2003: Assessment

Rapidly becoming a corporate de facto standard for mission critical application servers Wintel application servers provide industry best

price/performance Ample, but less than perfect measurement data

Plenty of resource utilization measures Precious few service level measures

Performance Monitor API provides strong foundation

Extended Counter support provided chiefly by Microsoft, but others are getting into the act: Oracle, Veritas, etc.


Where to get more information

“Inside Windows 2000” by David Solomon and Mark Russinovich

Windows 2000 Professional Resource Kit

TechNet or the Microsoft Developer Network (MSDN) CD

“Microsoft Windows NT Resource Kit, Volume 4: Optimizing Windows NT”

by Russ Blake

© Demand Technology, Inc. Windows Server 2003: An Update Demand Technology Software 1020 Eighth Avenue South, Suite 6, Naples, FL 34102 phone:

Documents