virtual systems - Exemplified by the

tenderlaSoftware and s/w Development

Dec 13, 2013 (3 years and 8 months ago)

48 views

Analysis of resource consumption in
virtual systems
-

Exemplified by the
I/O load of business applications

Robert Wierschke

Master Thesis

in cooperation with

HP SAP Competence Center Waldorf

24.
November 2009

Outline



Introduction



Methods and Tools



Analysis



Summary & Conclusion

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

2

Introduction

Problem statement


Companies need business software



Sizing


Provisioning of required hardware resources


Too high
-
> waste money


Too low
-
> risk high response time / system crash


Historically focused on


CPU


SAP
Application

Performance Standard (SAPS)


Storage capacity



I/O performance (input output operations per second; IOPS)
less important due to storage architecture used to satisfy
capacity needs

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

3

Introduction

Problem statement


Evolution of hard disks
(1)


80% capacity increase per year


10% read/write performance increase per year


Disk paradox
(2)


Sizing for capacity
-
> reduce number of devices


Sizing for performance
-
> increase number of devices



Thus, exact methods to predict I/O performance requirements are
necessary for accurate sizing.



(1)
Computer
Architecture
: A Quantitative Approach; John L. Hennessy, David A. Patterson;
Elsevier
, Kaufmann; 2007

(2)
Statistische Analyse von Speichersystemen in SAP Umgebungen und Entwicklung einer Software
-
Anwendung zur
optimalen Dimensionierung solcher Systeme; Christoph Schelling; Diplomarbeit BA Stuttgart; 2006


Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

4

Introduction

Previous work











Shows correlation between SAPS and IOPS

(see [2])

Physical I/O Rate = 1323,682 + 0,505 * SAPS*CPU%

Physical I/O Rate =
-

620,63 + 0,21 * SAPS*CPU% + 1943,84 *
system_size

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

5

SAPS * CPU% Line Fit Plot
0
2000
4000
6000
8000
10000
0
5000
10000
15000
SAPS * CPU%
Physical IOPS
Physical
IOPS
Predicted
Physical
IOPS
Introduction

Previous work


Tool to estimate I/O load from SAPS capacity











Only small sample size


Only for SAP R/3



Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

6


Extracted from SAPS Meter Portal


> 8 Mio. measurements, > 300 SAP systems


Measurements


CustomerID


HostID



SID



Timestamp


kSAPS


kIOPS



GB


Additional information


Customer industry


System type and usage

Introduction

Data

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

7

Methods

and

Tools

Linear Regression Analysis


Statistical method to analyse the effect of one or more
input
variables

on an
output variable


Derive function to predict output for known input


Procedure


Guess model from scatter plot


Estimate parameters of model


Ordinary least square method


Evaluate fitness of model


Residual plot


Normal distribution


Coefficient of determination


Significance test for input variable(s)

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

8

Methods

and

Tools

Tools


Microsoft Excel


Supports regression analysis and plotting


Data import is complex due to CSV interpretation and number
formatting


Data filtering is cumbersome


R
(http://www.r
-
project.org/)


Software for statistical computing


Supports regression analysis and plotting


Data filtering requires expertise in R Language


Drawbacks can be remedied by using Python (
RPy
)


RPy

has some bugs

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

9

Methods

and

Tools

Tools


Best Solution: Own implementation using Python


SciPy

and
NumPy


Matplotlib


Integrated
Sqlite

database


Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

10


Repeat linear correlation analysis to evaluate Schelling Hypothesis


Results evaluated with thresholds for coefficient of determination:


> 0.8: good fit


0.5


0.8: moderate fit


< 0.5: bad fit



Analysis shows that for the

majority of the systems the

IOPS
-
SAPS consumption
cannot

be described with a linear

correlation
.



Previous result could not be verified

Analysis

Verification of previous study

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

11

Analysis

Scatter plot analysis


Next step: Analyse scatter plots to guess a model


Scatter plots show:


High statistical spread of measurements


Mapping from SAPS to IOPS impossible


Noticeable patterns in some scatter plots


Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

12


Two clear branches visible (upper and lower branch)


Cannot be considered statistical spread due to its regularity


SAPS
-
value cannot be associated with
exactly

one IOPS
-
value


Prediction for upper branch would be sufficient

kIOPS

= 4,325
kSAPS

+ 0,133

separating branches

upper branch

lower branch

Analysis

V Pattern

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

13

Analysis

V Pattern


Branches are not related to specific usage phases











V Pattern was also found in other studies


Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

14

Analysis

Column
/
Trapezium

Pattern


Caused by image scaling









Re
-
scaling reveals a trapezium

whose borders are similar

to a V Pattern



Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

15

Analysis

Column
/
Trapezium

Patten


Filtering maximum values shows an upper border


As well as a considerable amount of statistical spread


An upper limit can be estimated


But is not applicable to


other systems

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

16

Analysis

Patterns and SAP solutions


Patterns cannot be
associated with
a specific SAP solution

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

17

0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
all
V Pattern
Column/Trapezium Patten

Visualizing measurements for specific SAP solutions shows


No mapping from SAPS to IOPS possible


Maximum values do not form a clear upper bound


Highest SAPS
-
values do not correspond with highest IOPS
-
values

Analysis

Upper bounds for SAP solutions

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

18

Analysis

Upper bounds for SAP solutions


Scatter plots show high amount of statistical spread


that depends on the SAPS values


and thus would lead to a regularity in residual plot



Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

19

Summary & Conclusion


Methods for precise prediction of I/O requirements are needed



Measurements of SAP systems form different patterns that


Are not related to specific usage phases


Cannot be associated with specific SAP Solutions



A prediction of IOPS requirements from SAPS capacity of a system
is not possible due to:


High amount of statistical spread of the measurements


High I/O loads do not correlate with high SAPS loads

Analysis of resource consumption in virtual systems | R. Wierschke | 24. November 2009

20