USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated...

26
HOW ADDING PROCESSORS AND ENABLING HYPER-THREADING AFFECT USER CAPACITY OF METAFRAME XP SERVERS By John D’Agati and Gagan Singh Citrix Systems Inc.

Transcript of USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated...

Page 1: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

HOW ADDING PROCESSORS AND ENABLING HYPER-THREADING AFFECT USER CAPACITY OF METAFRAME XP SERVERS

By John D’Agati and Gagan Singh

Citrix Systems Inc.

Page 2: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

NoticeThe information in this publication is subject to change without notice.THIS PUBLICATION IS PROVIDED “AS IS” WITHOUT WARRANTIES OF ANY KIND, EXPRESS

ORIMPLIED, INCLUDING ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A

PARTICULARPURPOSE OR NON-INFRINGEMENT. CITRIX SYSTEMS, INC. (“CITRIX”), SHALL NOT BE

LIABLE FORTECHNICAL OR EDITORIAL ERRORS OR OMISSIONS CONTAINED HEREIN, NOR FOR

DIRECT,INCIDENTAL, CONSEQUENTIAL, OR ANY OTHER DAMAGES RESULTING FROM THE FURNISHING,PERFORMANCE, OR USE OF THIS PUBLICATION, EVEN IF CITRIX HAS BEEN ADVISED OF

THEPOSSIBILITY OF SUCH DAMAGES IN ADVANCE.This publication contains information protected by copyright. Except for internal distribution, no part of this publication may be photocopied or reproduced in any form without prior written consent from Citrix.The exclusive warranty for any Citrix products discussed in this publication, if any, is stated in the product documentation accompanying such products. Citrix does not warrant products other than its own.Product names mentioned herein may be trademarks and/or registered trademarks of their respective companies.© 2003 Citrix Systems, Inc.All rights reserved. Printed in the U.S.A.

Version History

January 2003 John D’Agati Revisions

ii

Page 3: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

March 2003 John D’Agati Revisions

Overview...............................................................................................................................................1

Citrix ICAMark......................................................................................................................................1

Number of CPUs’ Effect on User Capacity................................................................................2

Performance Monitor Statistics for Single processor system:..........................................5Performance Monitor Statistics for Dual processor system:.............................................6Performance Monitor Statistics for Quad processor system:............................................7Hyper-Threading’s Effect on User Capacity.............................................................................9

Dell 2650 Results...............................................................................................................................9

Performance Monitor Statistics for Dell 2650 Hyper-Threading Enabled:..................11

Performance Monitor Statistics for Dell 2650 Hyper-Threading Disabled:................12

Dell 6650 Results.............................................................................................................................14

Performance Monitor Statistics for Dell 6650 Hyper-Threading Enabled:..................16

Performance Monitor Statistics for Dell 6650 Hyper-Threading Disabled:................17

Summary.............................................................................................................................................19

iii

Page 4: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

iv

Page 5: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Overview

The number of users that a server can support depends on several factors including: The MetaFrame server’s hardware specifications The applications that are being run (because of the applications’

CPU and memory requirements) The amount of user input being processed by the applications The maximum desired resource usage on the server, for example,

90% CPU usage or 80% memory usageThis section discusses the increase in user capacity when more CPUs are added, as well as the effect of Hyper-Threading in the processor. First, the Citrix benchmarking test for user capacity, known as ICAMark, is described.

Citrix ICAMark

Citrix ICAMark is an internal tool which is based on the Citrix Server Test Kit (CSTK) and used by Citrix Engineering for benchmarking purposes to quantify the optimal number of simulated client sessions that can be connected to a MetaFrame server with acceptable performance. Extending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact end user experience.The test simulates users constantly typing and performing actions in Microsoft Excel 97, Microsoft Access 97, and Microsoft PowerPoint 97. Other applications can utilize more or less memory and CPU than Microsoft Office 97 and therefore could produce different results. Note, that the simulated users in this test are constantly typing into these applications at 40 words per minute and may be considered more “rigorous” than normal users.In this test, a step size of “number of users” is defined as 5. During the course of the test, after the first 5 users are logged in, ICAMark launches simulated user scripts on all 5 sessions. Each script opens Microsoft Excel and simulates the creation of a spreadsheet, including calculations and charts. Once the Excel phase is complete, Excel is closed and Microsoft Access is opened. The script then simulates the creation of an Access database, including a table, query, and form, with data manipulation. Once the Access phase is complete, a Microsoft PowerPoint presentation is created of 6 slides, including spell checking, font changes, and slide copies and deletions. When a script is finished it remains idle until the scripts on all sessions are complete. The next iteration is then launched, adding 5 more sessions to the test and the process begins again. Based on how long the scripts take to complete, an ICAMark score is calculated. For this test, a score of 80 has been determined as the optimal load for a server. The ICAMark score is calculated by comparing a calibration value of the script time with the time gathered during the iteration. The calibration value was determined by running

1

Page 6: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

the scripts on a calibration machine. This machine is considered to perform at the level we expect from a stand alone workstation. Each script on the calibration machine was run locally, and the data recorded.An ICAMark score of 80 means that the server has enough additional CPU and memory resources to handle spikes in performance. When the test iteration score drops below 80, additional users added to the server consume more resources, producing lower test scores and slower performance.

Number of CPUs’ Effect on User Capacity

The benchmark test was performed with the following:

Server:Dell PowerEdge 6650Quad Processor - 1.6GHz Xeon with 256 KB L2 and 1 MB L3

Cache400MHz Front Side Bus Hyper threading is enabled35 GB HDD with Dell PERC 3/DC Raid Controller3.5 GB RAM4 GB Page FileCitrix MetaFrame XP Feature Release 2/Service Pack 2Microsoft Windows 2000 Advanced Server with Service Pack 2Microsoft Office 97

Clients:Dual Pentium P3 667 w/256 kb Cache256 MB RAM9 GB HDD with Adaptec SCSI ControllerCitrix ICA Program Neighborhood Client version 6.30.1050Microsoft Windows 2000 Service Pack 2

Tests were performed by keeping the hardware static and disabling processors on the server.

Results were collected on the following configurations: Dell 6650 with 1 processor enabled Dell 6650 with 2 processors enabled

2

Page 7: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 6650 with 4 processors enabled

The following results were collected:

# of CPUs

# of Simulated Users

% Performance

Increase

# of Users per Processor

1 70 ± 1 N. A. 70 ± 1

2 126 ± 1 80% 63 ± 1

4 160 ± 1 27% 40 ± 1

3

Page 8: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

User Capacity Benchmark

0

20

40

60

80

100

120

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200

# of Simulated Users

ICA

Mar

k Sc

ore

4 Processors2 Processors1 Processor

The results conclude that the performance of the Dell PowerEdge 6650 with 4 processors enabled and 160 concurrent simulated users, is equivalent to the performance of 2 processors enabled with 126 concurrent simulated users, which is equivalent to the performance of 1 processor enabled with 70 concurrent simulated users. In other words, the user experiences on each system with the number of simulated users shown above are equivalent. As CPUs are added to the server, the increase in performance is not linear thus allowing fewer users per processor as they are added to the system.

At optimal load, the following counters were noted:

# of Processors 1 2 4

% Processor Utilization

65% 65% 60%

Average PQL 70 120 55

Available Memory in KBytes

2,332,816 1,531,280 1,174,824

Current DQL 1 3 1

When the benchmark test ended, the following PTE information was noted:

4

Page 9: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

# of Processors 1 2 4

# of Users at End 100 150 200

Available PTEs 87,554 55,544 23,941

On the single and dual processor systems, the bottleneck is the processor. This is evident by the sustained 65% processor utilization and the high processor queue lengths. On the quad processor system, the bottleneck is believed to be the system bus. By looking at the graphs for the quad system, the processor utilization is 60% and the processor queue length is 55, lower than the single and dual systems at the failure point. As more processors are added to the system the bandwidth of the system bus becomes contentious resulting in the non-linear scalability shown here. A bottleneck in the system bus would cause contention between the processors and threads in memory causing slower script execution time.

Performance Monitor Statistics for Single processor system:

Single Proce s s or Sys te m

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulate d Us e rs

Proc

esso

r Q

ueue

Len

gth

5

Page 10: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Single Proce ss or Syste m

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulate d Us e rs

Proc

esso

r Util

izat

ion

Single Processor System

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

Single Processor System

0

1

2

3

4

5

6

7

8

9

10

11

12

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

6

Page 11: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Performance Monitor Statistics for Dual processor system:

Dual Processor System

0102030405060708090

100110120130140150

# of Sim ulate d Us e rs

Pro

cess

or Q

ueue

Len

gth

Dual Processor System

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulated Use rs

% P

roce

ssor

Tim

e

Dual Processor System

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

7

Page 12: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dual Processor System

0

2

4

6

8

10

12

14

16

18

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

Performance Monitor Statistics for Quad processor system:

Quad Processor System

0

10

20

30

40

50

60

70

80

90

100

# of SIm ulate d Us e rs

Proc

esso

r Que

ue L

engt

h

Quad Processor System

0

10

20

30

40

50

60

70

80

90

100

# of SIm ulate d Use rs

% P

roce

ssor

Tim

e

8

Page 13: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Quad Processor System

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

Quad Processor System

0

5

10

15

20

25

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

Hyper-Threading’s Effect on User Capacity

Hyper-Threading is technology developed by Intel that enables a single physical processor to appear as two logical processors. This technology was introduced in the Pentium IV line of processors. Hyper-Threading allows multi-threaded programs to take advantage of extra execution units on the processor resulting in as much as a 30% performance increase to some applications. Note that the benefit of Hyper-Threading is only seen with applications that are multi-threaded. MetaFrame and the applications it hosts can also benefit from Hyper-Threading. Note that the increase in performance from Hyper-Threading is highly dependent on the type of application that is running on the server. The benchmark test was performed on a Dell 2650 dual processor and a Dell 6650 quad processor with Hyper-Threading capable Pentium IV processors.

9

Page 14: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 2650 Results

This benchmark was performed on the following:

Server Hardware ConfigurationsDell PowerEdge 2650 Dual Processor - 2.2GHz Xeon with 512KB L2 Cache400MHz Front Side Bus16 GB HDD with Dell PERC 3/Di Raid Controller4 GB RAM4 GB Page FileCitrix MetaFrame XP Feature Release 2/Service Pack 2Microsoft Windows 2000 Advanced Server with Service Pack 2Microsoft Office 97

Clients:Dual Pentium P3 667 w/256 kb Cache256 MB RAM9 GB HDD with Adaptec SCSI ControllerCitrix ICA Program Neighborhood Client version 6.30.1050Microsoft Windows 2000 Service Pack 2

Hyper-threading # of Simulated Users

% Performance Increase

Off 116 ± 1 N. A.

On 131 ± 1 11.5%

10

Page 15: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

The results conclude that the performance of the Dell PowerEdge 2650 with Hyper-Threading enabled servicing 131 concurrent simulated users is equivalent to the performance of the Dell PowerEdge 2650 with Hyper-Threading disabled servicing 116 concurrent users. The lower Processor Queue Length on the system with Hyper-Threading enabled allows for faster execution of the scripts and accounts for the 11.5% increase in capacity.

At optimal load, the following counters were noted:

Hyper-Threading On Off

% Processor Utilization

70% 65%

Average PQL 70 115

Available Memory in KBytes

1,876,108 1,680,124

Current DQL 3 1

When the benchmark test ended, the following PTE information was noted:

Hyper-Threading On Off

# of Users at End 150 150

User Capacity Benchmark

0

20

40

60

80

100

120

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160# of Simulated Users

ICAMark Score

Hyper-threading EnabledHyper-threading Disabled

11

Page 16: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Available PTEs 52,201 52,608

Performance Monitor Statistics for Dell 2650 Hyper-Threading Enabled:

Dell 2650 Hyper-Threading On

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulated Us e rs

Proc

esso

r Que

ue L

engt

h

Dell 2650 Hyper-Threading On

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulate d Users

% P

roce

sso

r Tim

e

12

Page 17: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 2650 Hyper-Threading On

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Ava

ilabl

e M

emor

y K

Byt

es

Dell 2650 Hyper-Threading On

0

5

10

15

20

25

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

Performance Monitor Statistics for Dell 2650 Hyper-Threading Disabled:

Dell 2650 Hyper-Threading Off

0

20

40

60

80

100

120

# of Sim ulate d Us e rs

Proc

esso

r Q

ueue

Len

gth

13

Page 18: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 2650 Hyper-Threading Off

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulate d Users

% P

roce

sso

r Tim

e

Dell 2650 Hyper-Threading Off

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

Dell 2650 Hyper-Threading Off

0

5

10

15

20

25

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

14

Page 19: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 6650 Results

This benchmark was performed on the following:

Server Hardware ConfigurationsDell PowerEdge 6650Quad Processor - 1.6GHz Xeon with 256 KB L2 and 1 MB L3

Cache400 MHZ Front Side Bus35 GB HDD with Dell PERC 3/DC Raid Controller3.5 GB RAM4 GB Page FileCitrix MetaFrame XP Feature Release 2/Service Pack 2Microsoft Windows 2000 Advanced Server with Service Pack 2Microsoft Office 97

Clients:Dual Pentium P3 667 w/256 kb Cache256 MB RAM9 GB HDD with Adaptec SCSI ControllerCitrix ICA Program Neighborhood Client version 6.30.1050Microsoft Windows 2000 Service Pack 2

Hyper-threading # of Simulated Users

% Performance Increase

Off 158 ± 1 N. A.

On 160 ± 1 1%

15

Page 20: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

User Capacity Benchmark - Dell 6650

0

20

40

60

80

100

120

0 20 40 60 80 100 120 140 160 180 200# of Simulated Users

ICAM

ark

Scor

eHyper-threading EnabledHyper-threading Disabled

The results conclude that the performance of the Dell PowerEdge 6650 with Hyper-Threading enabled servicing 160 concurrent users is equivalent to the performance of the Dell PowerEdge 6650 with Hyper-Threading disabled servicing 158 concurrent users

At optimal load, the following counters were noted:

Hyper-Threading On Off

% Processor Utilization

60% 65%

Average PQL 55 115

Available Memory in KBytes

1,174,824 1,122,072

Current DQL 1 3

When the benchmark test ended, the following PTE information was noted:

Hyper-Threading On Off

# of Users at End 200 180

Available PTEs 23,941 37,168

16

Page 21: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

With Hyper-Threading enabled on a quad processor system, there are eight logical CPUs. In this instance, the benefit of Hyper-Threading is no longer evident. As the processing capacity of the server increases, less of a bottleneck is placed on the processor’s execution resources and more on the system bus speed, thus resulting in only a 1% increase in performance with Hyper-Threading enabled.

Performance Monitor Statistics for Dell 6650 Hyper-Threading Enabled:

Dell 6650 Hyper-Threading On

0

10

20

30

40

50

60

70

80

90

100

# of SImulated Users

Proc

esso

r Que

ue L

engt

h

Dell 6650 Hyper-Threading On

0

10

20

30

40

50

60

70

80

90

100

# of SImulated Users

% P

roce

ssor

Tim

e

17

Page 22: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 6650 Hyper-Threading On

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

Dell 6650 Hyper-Threading On

0

5

10

15

20

25

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

Performance Monitor Statistics for Dell 6650 Hyper-Threading Disabled:

Dell 6650 Hyper-Threading Off

0

20

40

60

80

100

120

140

# of Sim ulated Use rs

Proc

esso

r Que

ue L

eng

th

18

Page 23: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Dell 6650 Hyper-Threading Off

0

10

20

30

40

50

60

70

80

90

100

# of Sim ulated Use rs

% P

roce

ssor

Tim

e

Dell 6650 Hyper-Threading Off

0

500000

1000000

1500000

2000000

2500000

3000000

3500000

4000000

# of Simulated Users

Mem

ory

Ava

ilabl

e K

Byt

es

Dell 6650 Hyper-Threading Off

0

5

10

15

20

25

# of Simulated Users

Cur

rent

Dis

k Q

ueue

Len

gth

19

Page 24: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

Summary

In conclusion, user capacity does not scale linearly when adding processors to the system. There is an 80% increase in performance when moving from a single to a dual processor system. There is a 27% increase in performance when comparing a dual to a quad processor system. The increase in performance from Hyper-Threading is highly dependent on the type of applications running on the server. The true benefit of Hyper-Threading is seen on the dual processor system where the processor is the bottleneck, thus allowing an additional 11.5% increase in capacity when enabled. Note when sizing MetaFrame XP servers, the number of actual users per server varies based on the type of applications deployed.

20

Page 25: USER CAPACITY FOR METAFRAME SERVERS - Dell · Web viewExtending the number of concurrent simulated users beyond the optimal results will cause a decreased performance and may impact

21