Hp Connect 10 06 08 V5

49
1 Capacity Planning for Itanium Paul O’ Sullivan and Prem S. Sinha, PhD. PerfCap Corporation 76-39A Northeastern Blvd.,, Nashua, NH 03062 www.PerfCap.com; [email protected]; 603-594-0222

description

 

Transcript of Hp Connect 10 06 08 V5

Page 1: Hp Connect 10 06 08 V5

1

Capacity Planning for

Itanium

Paul O’ Sullivan and Prem S. Sinha, PhD.

PerfCap Corporation 76-39A Northeastern Blvd.,, Nashua, NH 03062

www.PerfCap.com; [email protected]; 603-594-0222

Page 2: Hp Connect 10 06 08 V5

2

PerfCap Corporation

• Group Started within Digital/Compaq (now HP) over 21 years ago

• Operating as independent corporation since 2001• Privately Owned, Zero Debt• Currently focused on Performance Monitoring, Capacity

Planning and Asset Management• 20+ Years of Solid Engineering & Development• Worldwide Presence• HP and other resellers continue to sell it world wide• Partnership

– HP, IBM, SUN– Microsoft Certified Partner

Page 3: Hp Connect 10 06 08 V5

3

Some of Current Customers

• Barclays UK• Commerzbank • Deutsche Bank UK• SIAC• Mary Kay • Certegy• Analog Devices• Royal Bank of Scotland

• BNP Paribas (3th Largest Retail Bank in Europe) Enterprise License – Unlimited use (3000+ deployed)

• ISE (Largest Options Stock Exchange)Enterprise License – Unlimited use

• US Postal ServicesMonitoring 450 nodes

• Thomson ReutersUp to 45,000+

• International Papers • Vodafone• British Telecom• MDS Pharmacy• Pfizer• Qwest • Lockheed Martin• Caremark

• Swedish Customs• Netherlands Army• CNS Dubai• UPMC Medical Center• UIC Medical Center• University Hospital, Zurich• US Dept. of Education• SUNY Buffalo Univ.

Page 4: Hp Connect 10 06 08 V5

4

Capacity Planning Endorsement

Adrian Cockcroft winner of A.A. Michelson lifetime achievement award at 2007 CMG, in his personal blog wrote

“The most interesting commercial tool I saw at CMG earlier this month is a capacity monitoring tool called PAWZ from PerfCap Corporation. The key thing they have worked on is taking the human out of the loop as much as possible with sophisticated capacity modeling algorithms and a simple and scalable operational model. ... The core idea is that you care about "headroom" in a service, and anything that limits that headroom is taken into account. Running out of CPU power, network bandwidth, memory, threads etc. will increase response time of the service, so monitor them all, track trends in headroom and calculate the point in time where lack of headroom will impact service response time. At eBay we used to call this the "time to live" for a service. You can easily focus on the services that have the shortest time to live, and proactively make sure that you have a low probability of poor response time.”

Page 5: Hp Connect 10 06 08 V5

5

Challenges

Do More With Less

• Large number of geographically dispersed resources

• Multi-platform

• Automate the process – On a daily basis

– Collect Data

– Consolidate/Analyze Data

– Generate Performance and Capacity Reports

– Send “Need-to-Know” Exception Notification

• Information availability: anytime anywhere

– web access

Page 6: Hp Connect 10 06 08 V5

6

Data ManagementHierarchical Approach

Raw Data

Key Performance Data

Risk Data

: Performance Analysts

: Capacity Planners

Page 7: Hp Connect 10 06 08 V5

7

Desk TopBrowser

Intranet

PAWZFindITServer (NT/W2K)

Networks Storage

Events

Trending

Clusters

Real Time

Applications

Performance

Reports

Daily, Weekly Health Reports

Critical Systems

Asset Location

Change Report

Configuration

Asset

Reports

Windows NT/2000/XP

SUN Solaris

HP-UX

IBM-AIX

OpenVMSCluster

LINUX

Tru64 UNIX

Page 8: Hp Connect 10 06 08 V5

8

PAWZ Components

• PAWZ Agent/Monitor: Resides on each node to be monitored– Collect Performance data 24x7

– Send colleted data to PAWZ Server in real time and/or once a day

• PAWZ Server: Resides on a Windows based server and communicates with hundreds of PAWZ Agents– Receives data from PAWZ Agent

– Processes and produces real time, daily and historical charts and reports

– Produces trend graphs for simple projections

– Runs a queuing network modeler for capacity planning

• PAWZ Browser: Resides on any corporate desktop

– Shows all reports and charts within Internet Browser

– Manage most of PAWZ functions

Page 9: Hp Connect 10 06 08 V5

9

PAWZ Key Functionality

• Collect performance data 24 x 7

• Provide real time and daily alerts based on performance thresholds

• Provide Performance Reports:– Real Time

– Daily

– Historical – for trending

• Performs Saturation Analysis every day for each node for capacity planning

• Performs Risk Analysis to detect systems that could be at Risk.

• Provides consolidated data center configuration report

Page 10: Hp Connect 10 06 08 V5

10

Capacity Planning

Definition: A process to determine how much computing resources are

required to meet business growthOr

How much business can grow before some device will run out of capacity

To answer “What if” questions like:– Can my current configuration handle three times of current workload – when will

my current configuration saturate– What will be impact of a new application on current system performance– What will be impact of upgrading a current server or adding a new server– Can I reduce the number of servers with out violating my “Service Level

Agreement” – a.k.a Server Consolidation

Page 11: Hp Connect 10 06 08 V5

11

Sizing Methods

RulesofThumb

LinearProjec-tions

AnalyticModels

Simula-tionModels

Bench-marks

RealSystem

Cos

t

Accuracy

Page 12: Hp Connect 10 06 08 V5

12

J F M A M J J A S O N D

Capacity Planning via Trending

Time

Per

form

ance

Met

ric

(Av.

or

Pea

k C

PU

Uti

liza

tion

)

• Simple to produce and follow• Issues

• defining right Capacity Limit• single vs composite metric• end user satisfaction

Today

RemainingCapacity

Capacity Limit

Page 13: Hp Connect 10 06 08 V5

13

PAWZ Planner

Workload

Res

pon

se T

ime

Saturation Point

Where do you want to operate?

Current Workload

Headroom

Response Time = {Service Time + Queuing Time}

Page 14: Hp Connect 10 06 08 V5

14

Capacity Planning via Modeling

Steps:

• Data Collection

• Identifying Peak Interval(s)

• Workload Characterization

• Model Validation

• Saturation Analysis

• “What If” Analysis

Page 15: Hp Connect 10 06 08 V5

15

PAWZ Planner

Page 16: Hp Connect 10 06 08 V5

16

Remaining Headroom (Capacity) Trend

Page 17: Hp Connect 10 06 08 V5

17

Headroom Risk Analysis

Time

Hea

dro

om

Headroom threshold

Headroom crosses threshold

Lead time

Amber status – system within lead time of dropping below headroom threshold.

Lead time

Headroom reaches 0

Red status – system within lead time of exhausting capacity.

Current state

Page 18: Hp Connect 10 06 08 V5

18

Risk Analysis

Page 19: Hp Connect 10 06 08 V5

19

Risk Analysis

Page 20: Hp Connect 10 06 08 V5

20

Risk Analysis

Page 21: Hp Connect 10 06 08 V5

21

Page 22: Hp Connect 10 06 08 V5

22

“What if”

PAWZ Planner has a “what-if” Capacity Planning module to forecast:- How much business can grow before some device will run out of

capacity• Can my current configuration handle three times of current

workload – when will my current configuration saturate• What will be impact of a new application on current system

performance• What will be impact of upgrading a current server or adding a

new server

Page 23: Hp Connect 10 06 08 V5

23

“What if”CPU & Disk Upgrade

Before

After

Page 24: Hp Connect 10 06 08 V5

24

Itanium Capacity Study

• Typical Study– Capability to do any platform to any other

platform (Alpha to Integrity)– Hardware:-

• Customer on Integrity Server cluster with HP-UX and Oracle

• RX8620 (4/4/16), 64Gb Memory

• SAN

Page 25: Hp Connect 10 06 08 V5

25

Itanium Capacity Study

• Alternate models considered:-– RX8640 32 Core– P570 32 Core– M8000 32 Core

• 3 or 4 node cluster considered

Page 26: Hp Connect 10 06 08 V5

26

Itanium Capacity Study

• Reason for Study– Expected substantial application growth– System already Peaking at CPU– What alternate configurations would provide

adequate growth of at least 200% current workload?

• HP and non-HP configurations considered

Page 27: Hp Connect 10 06 08 V5

27

Itanium Capacity Study

Page 28: Hp Connect 10 06 08 V5

28

CPU by Image / Disk I/O Rate

Page 29: Hp Connect 10 06 08 V5

29

CPU by Core

Page 30: Hp Connect 10 06 08 V5

30

Memory vs Process Count

Page 31: Hp Connect 10 06 08 V5

31

Total IO Counts

Page 32: Hp Connect 10 06 08 V5

32

IO Rates

Page 33: Hp Connect 10 06 08 V5

33

Disk Response Time

Page 34: Hp Connect 10 06 08 V5

34

Performance Data from Benchmark

CPU Utilization 86.3%

Disk I/O Rate 1514/s

Hard Page Fault Rate 1.2/s

Memory Utilization 73%

Page 35: Hp Connect 10 06 08 V5

35

Current Response Time Curve

Page 36: Hp Connect 10 06 08 V5

36

Where should your system live?

Page 37: Hp Connect 10 06 08 V5

37

At peak sustained load, 9% headroomCPU is primary resource bottleneckPossible solutions:

• Horizontal scaling• Integrity upgrade• Alternate hardware platform

Headroom - Current System

Page 38: Hp Connect 10 06 08 V5

38

Configuration Alternatives(3 or 4 nodes)

HP rx8620 (1.1 GHz, Itanium 2) – current configurationHP rx8640 (1.6 GHz, 24MB L3 cache), 16 coreHP rx8640 (1.6 GHz, 25MB L3 cache), 32 coreIBM p 570 (2.2 GHz, Power 5), 16 coreIBM p 570 (2.2 GHz, Power 5), 32 coreIBM p 570 (4.7 GHz, Power 6), 16 coreSun SPARC Enterprise M8000 (2.4 GHz) , 16 coreSun SPARC Enterprise M8000 (2.4 GHz) , 32 core

Configuration must support 200%

workload growth

Page 39: Hp Connect 10 06 08 V5

39

Response Time vs Workload Growth3-node RAC

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

-100 -50 0 50 100 150 200 250 300 350 400

% Workload Growth from Benchmark

Re

lati

ve

Re

sp

on

se

Tim

e

HP rx8620 (1.1 GHz Itanium 2), 16-core

HP rx8640 (1.6 GHz, 24MB, Itanium 2), 16-core

HP rx8640 (1.6 GHz, 24MB, Itanium 2), 32-core

IBM p570 (2.2 GHz, Power 5), 16-core

IBM p570 (2.2 GHz, Power 5), 32-core

IBM p570 (4.7 GHz, Power 6), 16-core

Sun SPARC Enterprise M8000 (2.4 GHz), 16-core

Sun SPARC Enterprise M8000 (2.4 GHz), 32-core

Page 40: Hp Connect 10 06 08 V5

40

Response Time vs Workload Growth4-node RAC

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

-100 -50 0 50 100 150 200 250 300 350 400

% Workload Growth from Benchmark

Rel

ativ

e R

esp

on

se T

ime

HP rx8620 (1.1 GHz, Itanium 2), 16-core

HP rx8640 (1.6 GHz, 24 MB L3 cache), 16-core

HP rx8640 (1.6 GHz, 24 MB L3 cache), 32-core

IBM p570 (2.2 GHz, Power 5), 16-core

IBM p570 (2.2 GHz, Power 5), 32-core

IBM p570 (4.7 GHz, Power 6), 16-core

Sun SPARC Enterprise M8000 (2.4 GHz), 16-core

Sun SPARC Enterprise M8000 (2.4 GHz), 32-core

Page 41: Hp Connect 10 06 08 V5

41

Projection Conclusions

• CPU is constraining resource• Memory, disk will support 200% growth• 3 configuration platforms support growth:

– HP rx8640 (1.6 GHz, 25MB L3 cache), 32 core

– IBM p 570 (2.2 GHz, Power 5), 32 core

– IBM p 570 (4.7 GHz, Power 6), 16 core

– Sun SPARC Enterprise M8000 (2.4 GHz) , 32 core

• Horizontal scaling to 4 nodes will not change qualifying platforms. However, cores may be adjusted.

Page 42: Hp Connect 10 06 08 V5

42

Minimal Cores, 3-node RAC

0.0

0.2

0.4

0.6

0.8

1.0

1.2

-100 -50 0 50 100 150 200 250 300

% Workload Growth from Benchmark

Re

lati

ve

Re

sp

on

se

Tim

e

Sun SPARC Enterprise M8000 (2.4 GHz), 32 cores

HP rx8640 (1.6 GHz, 25MB L3 cache), 30 cores

IBM p 570 (2.2 GHz, Power 5), 26 cores

IBM p 570 (4.7 GHz, Power 6), 12 cores

Page 43: Hp Connect 10 06 08 V5

43

Minimal Cores, 4-node RAC

0.0

0.2

0.4

0.6

0.8

1.0

1.2

-100 -50 0 50 100 150 200 250 300

% Workload Growth from Benchmark

Re

lati

ve

Re

sp

on

se

Tim

e

Sun SPARC Enterprise M8000 (2.4 GHz), 24 cores

HP rx8640 (1.6 GHz, 25MB L3 cache), 24 cores

IBM p 570 (2.2 GHz, Power 5), 20 cores

IBM p 570 (4.7 GHz, Power 6), 10 cores

Page 44: Hp Connect 10 06 08 V5

44

Mixing 1.1 GHz and 1.6 GHz Itanium Cores

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

-100 -50 0 50 100 150 200 250 300

% Workload Growth from Benchmark

Re

lati

ve

Re

sp

on

se

Tim

e

rx8620 (1.1 GHz, 16 cores)

rx8620 (1.1GHz, 16 cores + 1.6 GHz, 16 cores)

rx8640 (1.6 GHz, 32 cores)

Page 45: Hp Connect 10 06 08 V5

45

Minimal Number of Cores per Node Supporting 200% Growth

Platform 3-node 4-node

Sun SPARC Enterprise M8000 (2.4 GHz) 32 24

HP rx8640 (1.6 GHz, 25MB L3 cache) 30 24

IBM p 570 (2.2 GHz, Power 5) 26 20

IBM p 570 (4.7 GHz, Power 6) 12 10

Page 46: Hp Connect 10 06 08 V5

46

Itanium Capacity Study

• Customer satisfied– Had options

• Reduce Oracle cost by reducing number of cores

• Forecast from real data

• Could approach vendors with confidence

• Today– 90% of this study automated via PAWZ

• Same Graphs

• Same Results

Page 47: Hp Connect 10 06 08 V5

47

Modelling Capability

• Hardware– Alpha to Integrity– Integrity to new models and beyond– Other vendors to Integrity

• Software– Increases in workload– Optimization– Decreases in workload– Virtualization

Page 48: Hp Connect 10 06 08 V5

48

Summary

• PerfCap offers an integrated Performance Management and Capacity Planning Software that is:

– Out-of-the-box (no scripting required)

– Fully automated

– Multi-Platform

– Web based

– Highly scalable

• Pricing – Independent of number and class of CPUs in a server

Page 49: Hp Connect 10 06 08 V5

49

More Information

• Sales– [email protected]

• Web site– www.PerfCap.com

• Hot Line– 603-594-0222