Presentation is loading. Please wait.

Presentation is loading. Please wait.

Modeling Virtualized Environments in Simalytic ® Models by Computing Missing Service Demand Parameters CMG2009 Paper 9103, December 11, 2009 Dr. Tim R.

Published byClinton Park Modified over 9 years ago

Similar presentations

Presentation on theme: "Modeling Virtualized Environments in Simalytic ® Models by Computing Missing Service Demand Parameters CMG2009 Paper 9103, December 11, 2009 Dr. Tim R."— Presentation transcript:

1 Modeling Virtualized Environments in Simalytic ® Models by Computing Missing Service Demand Parameters CMG2009 Paper 9103, December 11, 2009 Dr. Tim R. Norton Simalytic Solutions, LLC 719-635-5825 email: tim.norton@simalytic.com http://www.simalytic.com © 2009 Simalytic Solutions, LLC

2 OR Using Virtual Data to Model Virtual Systems © 2009 Simalytic Solutions, LLC

3 Modeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 3 Agenda u Introduction  Capacity Planning and Virtualization u The Problem  Missing Measurements in Virtualization Guests u A Proposed Solution  Computing Missing Service Demand Parameters u Conclusion Simalytic ® is a registered trademark and registered service mark of Simalytic Solutions, LLC. Simalytic Modeling TM and Simalytic Function TM are trademarks of Simalytic Solutions, LLC. All other trademarked names and terms are the property of their respective owners.

4 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 4 Introduction u Capacity Planning  Capacity is Measured by Business Performance Objectives Making decisions about resource requirements  What do we have to buy and when do we have to buy it to make sure that the business applications perform at the level required to insure the business succeeds?

5 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 5 Introduction u Capacity Planning  Two key aspects: Demand for available resources l What do we have to buy Effective completion of business work l When to buy it?  Requires some predictive technique Usually some form of model l Simple: trend or other statistical techniques l Advanced: simulation or queuing network

6 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 6 Introduction u Virtualization  Increasing parallelization within the host system Increase productive business-related work Increase the usage of resources  Virtualized environment control program Hypervisor l Usually implies a hardware implementation VMM (Virtual Machine Monitor) l VMM often implies a software implementation t Will use VMM for all virtualization control programs

7 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 7 Virtualization Model VM1VM2VM3 VM4VM5VM6 Server 1 Server 2 Shared processors Shared disks

8 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 8 Introduction u Virtualization  From:  To: Sys-ASys-BSys-C Sys-ASys-BSys-C Good system level and process level measurements Only system level measurements

9 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 9 The Problem u Measurements in Virtualization Environments  Guest Operating Systems Most not Virtualization aware l Incorrect accounting for time when guest VM not active t Rate based measurements incorrect t Count based measurements valid  Hypervisor Accurately measures Guest VM active time l Cannot measure processes within the Guest O/S  Affects measurements needed for models VM System Level Service Times good Process Level Service Times incorrect

10 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 10 Capacity Models u What Virtualization does to Models  Reduces accuracy – Measurement issues: System clock Accounting for dispatch time of other VMs Interrupts – delays and re-driven to VM VM vs. process priority Delays – where are they accounted for Virtualization overhead l VMM (Virtual Machine Monitor) and VM context switch Interference from other VMs  Complicates Workload Characterization Shared resources

11 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 11 Where’s the Data u If It’s Not There…  It’s missing! u If It’s Wrong…  It’s missing! u If You Can’t Trust It…  It’s missing!

12 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 12 A Proposed Solution u Use known good measurements  VM utilization from the VMM (Virtual Machine Monitor)  Transaction arrivals from application measurements  Transaction response times from application measurements u Use a Simalytic Model Builds relationships Iterative design Leverages existing tools

13 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 13 Virtualization Metrics CPU OS1 CPU1 OS2OS3 APP1APP2APP3APP4APP5APP6 VM1VM2 VM3 MemoryDisk Hardware Platform Hypervisor Disk1 CPU2Disk2CPU3Disk3 VMM (Known Good) Measurements VM – Guest O/S Measurements Needed to model applications

14 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 14 Model Requirements u Application Transaction Arrivals  Count of transactions over many intervals u Application Response Times  Always needed for validation of model results Now needed for Solver calculations u Service Demand  Menascé technique to compute class service demand from total service demand VMM measurements used to compute VM level measurements that guest O/S cannot provide

15 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 15 Implementation Approach u Collect Measurement Data  Application – counts and response times  VMM resource usage u Compute Service Demand by Application  Validate against measured response times u Build Model of Overall Virtualization Environment  Using a Simalytic Model to express the relationships between applications

16 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 16 Implementation u Measurements  Collect measurements From the applications l Actual response times l Actual transaction counts From the VMM l Total utilization for each resource (CPU, disks, etc.)  Multiple intervals Measurements are needed for many intervals l Variety is important! – but avoid problem areas like low utilization effect t Low to high counts for each application t Low to high resource utilization

17 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 17 Implementation u Compute Service Demand  Use collected measurements to construct multiple Open Multiclass QN formulae Each uses the transaction response time and count for each class (application) along with the total service demand for each resource (CPU, disks, etc.) for one measurement interval.  Solve the non-linear constraint problem Computing Missing Service Demand Parameters for Performance Models. Danny Menascé, CMG 2008. Can be done with Microsoft Excel Solver

18 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 18 Implementation u Model Beyond What was Measured  Measurement data provides historical view Many intervals available l But at lower than expected future traffic volume  Model results provide future view Answer the classic questions: l When does response time become unacceptable? l What resource saturates first?  The following example takes this approach Measure system and applications at lower volumes Predict behavior at higher volumes l Different than example in the paper

19 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 19 Excel Solver Spreadsheet u Solves Multiple Equations  Initial Guess and Solver Results  Known Good Measurements  “Actual” measurements  Computed measurements  Validation Example spreadsheet is available at http://www.simalytic.com/Papers.html

20 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 20 Using Excel Solver u Solver Results – Example Spreadsheets  Start with initial guess Same for both classes – calculated from VMM resource measurements divided sum of trans in both classes.  Computes values for both classes non-linear constraints  Actual Service Demand Data Not real measurements l Validate approach l Generate data for testing

21 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 21 Using Excel Solver u “Actual” Service Demand Values  Used at low transaction volume to create known good response times and device utilizations Used for Solver goals Simulates actual known good measurements l M/M/1 formulae

22 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 22 Using Excel Solver u “Actual” Service Demand Values  Used at high transaction volume to create “actual” (projected) response times and device utilizations Predicts results at higher arrival rates Simulates future actual measurements to validate Solver results

23 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 23 Using Excel Solver u Computed Service Demand Values  Used to create “computed” response times and device utilizations Solver results Validated against “projected” values

24 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 24 Using Excel Solver u Solver Answer Report  Shows details behind solution

25 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 25 Application Performance u Two groups of measurements  “Actuals” are data used to create the models  “Projections” are data used as post-model measurements  Each class (application) shows response time increase as utilizations go up

26 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 26 Application Prediction u How Well do Computed Results Match  Actual Measurements  Results using Computed Service Demands  Some difference but same trend

27 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 27 Usage u How Can Computed Service Demands be Used?  Stand-alone models Same as measured service demands l Adjustments may be needed for VMM overhead and other interference – similar to other model calibrations  Simalytic Models Enhanced Simalytic Function for multi-tier models l Dynamic calculations to t simulate complex usage patterns t account for effects of spikes in other workloads

28 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 28 Conclusion u Future Work  Explore criteria around measurement collection How many intervals needed for minimal effectiveness What improves accuracy  Incorporate solver into modeling tool Possibly as enhanced Simalytic Function

29 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 29 Conclusion u Valid Approach  Works with synthetic test data Some differences l between projected and computed results But trends usable for planning  Refinement needed account for VMM overhead  Model virtual systems with virtual data

30 © 2009 Simalytic Solutions, LLCModeling Virtualized Environments by Computing Missing Service Demand Parameters CMG 2009 - Paper 9103, December 11, 2009 30 Missing Data Doesn’t Stop a Real Modeler! Presentation and spreadsheet will be available at: http://www.simalytic.com/Papers.html

Download ppt "Modeling Virtualized Environments in Simalytic ® Models by Computing Missing Service Demand Parameters CMG2009 Paper 9103, December 11, 2009 Dr. Tim R."

Similar presentations

Topics to be discussed Introduction Performance Factors Methodology Test Process Tools Conclusion Abu Bakr Siddiq.

Topics to be discussed Introduction Performance Factors Methodology Test Process Tools Conclusion Abu Bakr Siddiq.
Capacity Planning in a Virtual Environment

Capacity Planning in a Virtual Environment
Design of Experiments Lecture I

Design of Experiments Lecture I
Hadi Goudarzi and Massoud Pedram

Hadi Goudarzi and Massoud Pedram
Measuring and Modeling Hyper-threaded Processor Performance Ethan Bolker UMass-Boston September 17, 2003.

Measuring and Modeling Hyper-threaded Processor Performance Ethan Bolker UMass-Boston September 17, 2003.
Virtualization in HPC Minesh Joshi CSC 469 Dr. Box Feb 1, 2012.

Virtualization in HPC Minesh Joshi CSC 469 Dr. Box Feb 1, 2012.
Stand Tall, and Carry a Precision Micrometer: Observations on Creating a Measurement Model for Virtual Machines David Boyes Sine Nomine Associates.

Stand Tall, and Carry a Precision Micrometer: Observations on Creating a Measurement Model for Virtual Machines David Boyes Sine Nomine Associates.
Proactive Prediction Models for Web Application Resource Provisioning in the Cloud _______________________________ Samuel A. Ajila & Bankole A. Akindele.

Proactive Prediction Models for Web Application Resource Provisioning in the Cloud _______________________________ Samuel A. Ajila & Bankole A. Akindele.
Efficient Autoscaling in the Cloud using Predictive Models for Workload Forecasting Roy, N., A. Dubey, and A. Gokhale 4th IEEE International Conference.

Efficient Autoscaling in the Cloud using Predictive Models for Workload Forecasting Roy, N., A. Dubey, and A. Gokhale 4th IEEE International Conference.
1 Virtual Machine Resource Monitoring and Networking of Virtual Machines Ananth I. Sundararaj Department of Computer Science Northwestern University July.

1 Virtual Machine Resource Monitoring and Networking of Virtual Machines Ananth I. Sundararaj Department of Computer Science Northwestern University July.
OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.

OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.
Performance Evaluation

Performance Evaluation
ENFORCING PERFORMANCE ISOLATION ACROSS VIRTUAL MACHINES IN XEN Diwaker Gupta, Ludmila Cherkasova, Rob Gardner, Amin Vahdat Middleware '06 Proceedings of.

ENFORCING PERFORMANCE ISOLATION ACROSS VIRTUAL MACHINES IN XEN Diwaker Gupta, Ludmila Cherkasova, Rob Gardner, Amin Vahdat Middleware '06 Proceedings of.
Copyright © 1998 Wanda Kunkle Computer Organization 1 Chapter 2.1 Introduction.

Copyright © 1998 Wanda Kunkle Computer Organization 1 Chapter 2.1 Introduction.
U NIVERSITY OF M ASSACHUSETTS, A MHERST Department of Computer Science Virtualization in Data Centers Prashant Shenoy

U NIVERSITY OF M ASSACHUSETTS, A MHERST Department of Computer Science Virtualization in Data Centers Prashant Shenoy
Measuring Performance Chapter 12 CSE807. Performance Measurement To assist in guaranteeing Service Level Agreements For capacity planning For troubleshooting.

Measuring Performance Chapter 12 CSE807. Performance Measurement To assist in guaranteeing Service Level Agreements For capacity planning For troubleshooting.
Accounting Management IACT 918 April 2005 Glenn Bewsell/Gene Awyzio SITACS University of Wollongong.

Accounting Management IACT 918 April 2005 Glenn Bewsell/Gene Awyzio SITACS University of Wollongong.
OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.

OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.
Failure Avoidance through Fault Prediction Based on Synthetic Transactions Mohammed Shatnawi 1, 2 Matei Ripeanu 2 1 – Microsoft Online Ads, Microsoft Corporation.

Failure Avoidance through Fault Prediction Based on Synthetic Transactions Mohammed Shatnawi 1, 2 Matei Ripeanu 2 1 – Microsoft Online Ads, Microsoft Corporation.
By- Jaideep Moses, Ravi Iyer , Ramesh Illikkal and

By- Jaideep Moses, Ravi Iyer , Ramesh Illikkal and

Similar presentations

Ads by Google