HP HP-UX IPQos White Paper

HP-UX IPQoS Performance and Sizing White Paper

November 2005

Abstract.............................................................................................................................................. 2

Introduction......................................................................................................................................... 2

Tool Used for Measuring Performance.................................................................................................... 3

Test Configuration................................................................................................................................ 3

Interpretation of Results......................................................................................................................... 4

Raw Data ........................................................................................................................................... 5

Future Tests ......................................................................................................................................... 7

Abstract

This document outlines the performance impact of HP-UX IPQoS on an HP-UX system. It is divided into five sections. The first section is an introduction to IPQoS; the second section is about the tool used for the tests; the third section gives an overview of test configuration; the fourth section gives an interpretation of the test results; the fifth section contains raw test result data in tabular form; and the last section gives some suggestions for further study.

Introduction

During the past few years, the Internet has evolved from a simple network carrying primarily data traffic, into a complex network handling a variety of traffic, ranging from real time audio and video to web traffic. However, in terms of throughput, delay and packet loss, the “best effort” nature of the current Internet is not sufficient to cope with the requirements of this type of traffic.

HP-UX IPQoS can be used to bring an HP-UX node into conformance with the IETF Differentiated Services (DiffServ) model. The key features of HP-UX IPQoS are:

• Conforms to the IETF Differentiated Services (DiffServ) Model.

• Provides differentiated classes of service on outbound traffic by performing traffic

conditioning actions. Important traffic classes can take bandwidth away from less important classes, up to user-specified limits.

- Classification occurs when traffic classes are defined in filters.

- Marking occurs when marking attributes are set in policies.

- Metering occurs when bandwidth is reserved for defined traffic classes

in policies.

• Allows DSCP and VLAN marking on outbound traffic from the HP-UX server.

- Can assign different DSCP network routing priorities (valid range 0-63).

- Can assigned different VLAN priorities (valid range 0-7).

• Supports traffic classification on broad range of packet attributes.

• Provides provisioned QoS management.

• Supports both IPv4 and IPv6.

HP-UX IPQoS performs the above items by reading the header of an IP packet (plus transport TCP/UDP header) and comparing it to a set of user configured rules. The specific action is taken based on the rules. Because HP-UX IPQoS only reads the IP header, the size of an IP packet does not affect the speed at which HP-UX IPQoS processes the packet.

However, the faster an interface connection is, the more packets per second it can handle. A link capable of 56 Kbps and 1.544Mbps can carry up to 350 and 9650 packets per second, respectively. Faster connections such as 10-baseT (Ethernet) and 100-baseT (Fast Ethernet) can carry up to 62,500 and 625,000 packets respectively.

Because HP-UX IPQoS works the same on all packets, regardless of size, the more packets an interface can handle the more checking HP-UX IPQoS has to do. Therefore, the impact of HP-UX IPQoS on a slower interface is minimal because there are not as many packets for it to check. The performance impact of HP-UX IPQoS on a faster interface is greater, because more packets are passing through, and HP-UX IPQoS must process them all.

Tool Used for Measuring Performance

The tool used for measuring the performance of HP-UX IPQoS is netperf. Developed by HP, this tool is distributed in source form at www.netperf.org The netperf tool has two executables: netperf and netserver. netserver can run as a child process to inetd. In that case, netserver should be added to the /etc/inetd.conf standalone daemon.

Two systems are required to test HP-UX IPQoS. One system runs netserver (the remote system) and the other system runs netperf (the local system). By default, netperf is installed in the /opt/netperf directory (it is possible to install it in a different directory, but changes must be made to the netperf makefiles). The remote system must have an entry for netserver in the /etc/inetd.conf system must have an entry for netperf in the /etc/services port number 12865. After the /etc/inetd.conf the remote system using kill –HUP <PID#>.

netperf provides many scripts for running different tests. See the netperf manpages for a detailed description of these scripts.

The most common use of netperf is measuring bulk data transfer performance. This is also referred to as "stream" or "unidirectional stream" performance. Essentially, these tests will measure how fast one system can send data to another and/or how fast that other system can receive it.

. This site is informally maintained by Hewlett-Packard.

file. netserver can also run as a

file and the local

files identifying that netperf is using TCP

file is edited, make sure to update the inetd daemon on

Test Configuration

The tests were done on two systems; one running HP-UX 11i v1 and the other running HP-UX 11i v2 (update 2). The two systems, connected back to back with a crossover cable, were used for the netperf tests. HP-UX IPQoS A.01.00 was only installed on the HP-UX 11i v2 system (rp3440, 4 CPU). The other system was running HP-UX 11i v1 (rp4440, 6 CPU) without HP-UX IPQoS installed. While testing, care should be taken to make sure that other network traffic does not interfere with the traffic generated by the tests. This can be achieved by performing all the tests in a private LAN.

Tests were done on both 100BaseT and 1000BaseT interfaces with the different HP-UX IPQoS configurations occurring on the local system. As specified, the remote system was running HP-UX 11i v1 without HP-UX IPQoS installed.

On the local system (HPUX 11i v2), the tests were done twice; once with HP-UX IPQoS installed (pushed) but with no filters or policies (rules) configured. In this case HP-UX IPQoS only plays the role of “pass-through”. The second time, HP-UX IPQoS was configured with a simple filter and policy. Note: tests did not specify any bandwidth reservation since the purpose of the tests was to see the performance impact of HP-UX IPQoS on the HP-UX operating system. The simple filter and policy used for the tests was:

HP-UX IPQoS is ENABLED and FILTERING. Name: current Adapter: lan1 (00:0e:7f:4e:bf:88) Type: Ethernet Speed: 100 Mbps Policy: p1 DSCP: 20 Filter: f1 Transport protocol: 6 (tcp) Priority: 0

This filter and policy rewrites the TOS value on every TCP packet. Note: rewriting the TOS value in IP packet involves recalculating and resetting the checksum value for each IP packet.

Interpretation of Results

netperf does not provide per packet processing numbers. Service demand numbers are the closest approximation to the per packet overhead. Service demand indicates the amount of processing done per 1KB of data. For a TCP packet payload of 1460 bytes (MTU of 1500 bytes minus 40 bytes of TCP and IP header), the service demand numbers can be multiplied by 1.46 to get the approximate per packet processing time.

Another important measurement is CPU utilization. In the netperf tests, the CPU utilization numbers show that, for the 1000BaseT interface, more data (and hence more packets) were handled in a given time interval, requiring more processing time. As described in the introduction, this behavior is expected with HP-UX IPQoS.

Typical results from a netperf output indicate an increase in service demand by anywhere from 5 to 45 microseconds in a system with HP-UX IPQoS installed and a simple filter and policy configured, compared to a system without HP-UX IPQoS installed.

CPU utilization increases anywhere from 1% to 15% when HP-UX IPQoS is installed with a simple filter and policy on both 100BaseT and 1000BaseT connections. For faster connections, expect the CPU utilization to be higher.

Throughput utilization decreases anywhere from 1% to 10% when HP-UX IPQoS is installed with the simple filter and policy configured, especially on the 1000BaseT interface since more packets are processed.

Each test was repeated because of varying confidence level in the results reported by netperf. Only one set of data for each type of test is provided here.

Raw Data

This section provides raw data in tabular format to illustrate results described in the previous section.

Without IPQoS (100BaseT interface ) Recv Socket Size 32768 32768 32 60.01 30.29 41.79 456.731 32768 32768 64 60.01 60.11 31.72 162.308 32768 32768 128 60.01 94.07 15.69 54.645 32768 32768 256 60.01 75.38 8.93 38.993 32768 32768 512 60.01 92.45 7.58 26.908 32768 32768 1024 60.01 92.28 5.19 18.418 32768 32768 2920 60.01 94.23 2.53 8.811

With IPQoS no filter, pass through (100BaseT interface) Recv Socket Size 32768 32768 32 60.01 29.79 42.29 496.607 32768 32768 64 60.01 59.59 32.90 182.916 32768 32768 128 60.01 94.01 15.57 54.259 32768 32768 256 60.01 74.94 9.17 40.111 32768 32768 512 60.01 92.34 7.31 25.936 32768 32768 1024 60.01 92.21 5.52 19.598 32768 32768 2920 60.01 94.12 2.60 9.038

With IPQoS simple filter (100BaseT interface ) Recv Socket Size 32768 32768 32 60.01 28.26 47.73 630.270 32768 32768 64 60.01 54.92 34.97 220.909 32768 32768 128 60.01 91.43 18.15 65.111 32768 32768 256 60.01 74.58 10.18 44.456 32768 32768 512 60.01 92.18 8.93 31.700 32768 32768 1024 60.01 92.16 7.04 24.977 32768 32768 2920 60.01 93.96 3.78 13.152

Send Socket Size

Send Msg Size

Elapsed Time (sec)

Throughput 10^6 Bits/s

Utilization Send Local

Service Demand Local

Raw data for GIGE-bit interface:

Without IPQoS (1000BaseT interface ) Recv Socket Size 32768 32768 32 60.01 30.35 38.40 458.873 32768 32768 64 60.01 61.04 36.92 206.787 32768 32768 128 60.01 113.99 39.52 113.733 32768 32768 256 60.01 218.35 38.69 58.206 32768 32768 512 60.01 405.24 36.93 29.981 32768 32768 1024 60.01 754.00 34.87 15.196 32768 32768 2920 60.01 922.74 19.79 7.028

With IPQoS no filter, pass through (1000BaseT interface ) Recv Socket Size 32768 32768 32 60.01 29.49 40.05 445.062 32768 32768 64 60.01 58.31 39.47 218.218 32768 32768 128 60.01 107.97 36.93 114.555 32768 32768 256 60.01 218.01 38.05 56.849 32768 32768 512 60.01 403.76 36.16 29.355 32768 32768 1024 60.01 749.76 34.41 15.061 32768 32768 2920 60.01 921.19 19.96 7.089

With IPQoS simple filter(1000BaseT interface ) Recv Socket Size 32768 32768 32 60.01 28.16 40.47 585.924 32768 32768 64 60.01 54.92 39.97 220.909 32768 32768 128 60.01 105.53 37.93 126.455 32768 32768 256 60.01 209.39 39.22 64.344 32768 32768 512 60.01 361.38 38.78 35.882 32768 32768 1024 60.01 625.90 37.64 19.840 32768 32768 2920 60.01 912.12 28.46 10.224

Send Socket Size

Send Msg Size

Elapsed Time (sec)

Throughput 10^6 Bits/s

Utilization Send Local

Service Demand Local

Future Tests

The following are ideas for tests that may be performed sometime in the future:

• Test to determine the performance impact of HP-UX IPQoS with longer and more complex filters

• Test to determine the performance impact of HP-UX IPQoS on UDP packets

• Test to determine the performance impact of HP-UX IPQoS on TCP request/response

mechanism

• Test to determine the performance impact of HP-UX IPQoS with non-matching filters

• Test to determine the performance impact of HP-UX IPQoS with fragmented IP packet

• Study the impact of HP-UX IPQoS on GIGA-bit interface when sending large packets

© 2005 Hewlett-Packard Company, L.P. The information contained herein is subject to change without notice. The only warranties for HP products and services are set forth in the express warranty statements accompanying such products and services. Nothing herein should be construed as constituting an additional warranty. HP shall not be liable for technical or editorial errors or omissions contained herein.

ipqos_perf, 11/2005

HP HP-UX IPQos White Paper

Specifications and Main Features

Frequently Asked Questions

User Manual