RTI Connext DDS Performance Test Getting Started Guide
Contents
1. Auto Throttling enableAutoThrottle This option enables the Auto Throttle fea ture introduced in RTI Connext 5 1 0 When this option is used the publisher automati cally adjusts the writing rate based on the number of unacknowledged samples in the send queue to avoid blocking Note The following values in the DataWriterProtocolQosPolicy are hard coded in the applica tion therefore setting these values in the XML QoS profile will have no effect I rtps_reliable_writer heartbeats_per_max_samples is set to sendQueueSize 10 I rtps_reliable_writer low_watermark is set to sendQueueSize 0 10 Test Parameters 8 3 8 4 8 5 I rtps_reliable_writer high_watermark is set to sendQueueSize 0 90 For more information on the send queue size see the RESOURCE_LIMITS QosPolicy Section 6 5 20 in the RTI Core Libraries and Utilities User s Manual specifically the max_samples field Number of Iterations vs Latency Count When configuring the total number of samples to send during the test numIter lt count gt and the number of samples to send between latency pings latencyCount lt count gt keep these things in mind L Don t send latency pings too often One of the purposes of the test is to measure the throughput that the middleware is able to achieve Although the total throughput is tech nically the total data sent on both the throughput and latency topics for the sake of sim plicity the test measures only2. If scan is specified this value is ignored debug is debug lt true false gt Run in debug mode generates more verbose logging messages which are useful to RTI support personnel Default false durability lt 0 1 2 3 gt durability lt 0 1 2 3 gt Sets the Durability kind e 0 VOLATILE default 1 TRANSIENT LOCAL 2 TRANSIENT 3 PERSISTENT For an introduction to the RTI durability model see the Historical Data design pattern in the RTI Core Libraries and Utilities Getting Started Guide See also Mechanisms for Achieving Information Durability and Persistence Chapter 12 in the RTI Core Libraries and Utilities User s Manual domain lt ID gt domain lt ID gt Domain ID The publisher and subscriber applications must use the same domain ID in order to communicate Default 1 Range 0 99 See Choosing a Domain ID and Creating Multiple Domains Section 8 3 4 in the RTI Core Libraries and Utilities User s Manual Test Parameters Table 8 1 Test Parameters for Publishing and Subscribing Applications Command Line Option Entry in perftest ini Description enableSharedMemory enable shared memory lt true false gt Enable the shared memory transport Default shared memory transport is disabled enable tcp only Disable all the other transports and use only enableTcpOnly TCP transport for communication lt true fa
3. Specify the multicast receive address for receiving user data If unspecified the following default values will be used according to the topic e latency 239 255 1 2 e throughput 239 255 1 1 e announcement 239 255 1 100 nomulticast is multicast lt true false gt Do not use multicast Note Starting in 5 1 0 this option is no longer needed since multicast is disabled by default It exists only to maintain backward compatibility Default do not use multicast Test Parameters Table 8 1 Test Parameters for Publishing and Subscribing Applications Command Line Option Entry in perftest ini Description nic lt ipaddr gt interface lt ip address gt Restrict RTI Connext to sending output through this interface This can be the IP address of any available network interface on the machine By default RTI Connext will attempt to contact all possible subscribing nodes on all available network interfaces Even on a multi NIC machine the performance over one NIC vs another may be different e g Gbit vs 100 Mbit so choosing the correct NIC is critical for a proper test noDirectCommunication direct communication lt true false gt Indicates if the subscribing application will receive samples from the publishing application when RTI Persistence Service is used Only applies when durability lt 011 213 gt is TRANSIENT 2 or PERSISTENT 3 If set to tr
4. The period at which the publishing application will send heartbeats See Reliable Communications Chapter 10 in the RTI Core Libraries and Utilities User s Manual heartbeat period sec Default heartbeat period sec 0 heartbeat Period SACE heartbeat period nanosec 0 lt sec gt lt nanosec gt heartbeat period nanosec meaning use the value as specified lt nanosec gt in the XML QoS Profile which is set to 10 millisec 10 000 000 nanosec See qosprofile lt filename gt Range actual value 1 nanosec to 1 year 31 536 000 sec 12 Test Parameters Table 8 2 Test Parameters Only for Publishing Applications Command Line Option Entry in perftest ini Description fastHeartbeatPeriod lt sec gt lt nanosec gt fast heartbeat period lt sec gt sec fast heartbeat period nanosec lt nanosec gt An alternative heartbeat period used when the publishing application needs to flush unacknowledged samples more quickly See Reliable Communications Chapter 10 in the RTI Core Libraries and Utilities User s Manual Default heartbeat period sec 0 heartbeat period nanosec 0 meaning use the value as specified in the XML QoS Profile which is set to 1 millisec 1 000 000 nanosec See qosprofile lt filename gt Range actual value 1 nanosec to 1 year 31 536 000 sec Must not be slower than heartbeatPeriod lt sec gt lt nanosec gt
5. bin lt arch gt Release perftest_cpp pub noPrint pidMultiPubTest 0 sendQueueSize 32 numSubscribers 2 latencyCount 1000 numIter 1000000 dataLen 200 multicast Subscriber 1 bin lt arch gt Release perftest_cpp sub noPrint dataLen 200 numPublishers 1 sidMultiSubTest 0 multicast Subscriber 2 bin lt arch gt Release perftest_ cpp sub noPrint dataLen 200 numPublishers 1 sidMultiSubTest 1 multicast 9 8 2 to 1 Multicast Reliable UDPv4 Publisher 1 bin lt arch gt Release perftest_cpp pub noPrint pidMultiPubTest 0 sendQueueSize 32 numSubscribers 1 latencyCount 1000 numIter 1000000 dataLen 200 multicast Publisher 2 bin lt arch gt Release perftest_ cpp pub noPrint pidMultiPubTest 1 sendQueueSize 32 numSubscribers 1 latencyCount 1000 numIter 1000000 dataLen 200 multicast Subscriber bin lt arch gt Release perftest_ cpp sub noPrint dataLen 200 numPublishers 2 sidMultiSubTest 0 multicast 10 Example Output The following is an example of the expected output from the performance test Publisher perftest_cpp pub noPrint domain 27 sendQueueSize 50 latencyCount 10000 scan Waiting to discover 1 subscribers Waiting for subscribers announcement Publishing data Length 32 Latency Ave 396 us Std 48 9 us Min 83 us Max 538 us 50 401 us 90 459 us 99 510 us 99 99 538 us Length 64 Latency Ave 399 us Std 53 1 us Min 88 us Max 1062 us 50 403 us 90 461 us 99 537
6. in Functionality of Performance Test s enableSharedMemory Option 1 1 14 Reorganization of QoS Settings ccs cesesssssesessssssesesesesesesesescsssesesesesssessseseeees 2 Supported Platforms ccsscsssscsssssssssesssssssssssesssssssssesesesssssessessssssssesessssessseecsessssescseesesessessessseseeeseees 2 OV OL VI EW iiss hissensesciesuscicsvecevessacvessatescevede cduncascessdsudeecesccencuade veceduecevedessuusevescedbusedeses tussasdses suusaseees suesuseescowccvasesevse 3 Download Tnstru Ct Ons vessteccccsccsvscccasccesscacesiascscdocsecteacecosatesuuteantescastassasscesscceecccaucessucceguesendtanteiesssasessstesasonstoetie 3 Installation Instructions cic cciecceeck cdccaccccocedesccsasecenedcoucdcuseducesedeecsssdeses esutcssocseseetceosetscust sesosdecds suchdscesdscececvosteessec 3 Bele UNIX Based Systems oii anenee eean duces ce sci sccstes seetdtalis Ea AP aeaa E ccscuessstacvestseseescunepasslesenessecuevesesess 3 52 Windows SYSt Miswaiccie eek ction oii r ies ery nace est E ee neil cess St erent ieee onsite 4 Building the Application ccccccsssssecessscsssesesssssssssesesssssssssessssessssesesssssssssesessssssesssssssssesesseessseseseseseees 4 Starting the TeSt orosei ceiien i inskir i E HEET E E ETE E E E T E 5 Test Parameters neien era e Ei E E R AA E ai 7 8 1 o Spinning VsxSleepinrg e eps a aas oaan ETERA oaa EE A Eaa GEA sade Siva ES 15 8 2 Send Queue Size and Queue Full Behavior cccccccesccssessesse
7. keyed types The performance test supports both unkeyed and keyed types Out of the box the makefiles project files use the rtiddsgen utility to generate code for unkeyed types To use the test with keyed types do one of the following _ Define the RTIIDDSGEN_PREPROCESSOR environment variable to be D PERFTEST_KEYED_TYPE Set this environment variable before compiling Or J Uncomment the following line in lt installation directory gt idl test idl define PERFTEST KEYED TYPE 1 For information about keyed and unkeyed types please see Chapter 3 in the RTI Core Libraries and Utilities User s Manual Starting the Test Examples of how to build the application L C on a UNIX based System From a command line terminal cd rtiperftest 5 1 0 perftest_cpp gmake f Makefile lt architecture gt Where lt architecture gt is the architecture that you want to build For example to build on a Red Hat Enterprise Linux 5 32 bit system gmake f Makefile i86Linux2 6gcc4 1 1 Makefiles for some architectures are included in the example If your architecture s makefile is not provided you can create your own makefile based on the shipped exam ple makefiles L C and C on a Windows System 1 Open the file perftest lt architecture gt sln in Visual Studio where lt architecture gt is the architecture that you want to build Projects and solution files for some architectures are included in the example If y
8. latencyCount lt count gt latency count lt count gt Number samples to send before a latency ping packet is sent See Number of Iterations vs Latency Count Section 8 3 Default 1 if latencyTest is not specified automatically adjust to 10 000 if latency Test is specified automatically adjust to 1 Range must be lt numIter latencyTest run latency test lt true false gt Run a latency test consisting of a ping pong The publisher sends a ping then blocks until it receives a pong from the subscriber Can only be used on a publisher whose pidMultiPubTest 0 see pidMultiPubTest lt id gt Default false numiIter lt count gt num inter lt count gt Number of samples to send See Number of Iterations vs Latency Count Section 8 3 and Warming Up Section 8 4 If you set scan true you cannot set this option see scan Default 0 infinite Range latencyCount adjusted value or higher see latencyCount lt count gt numSubscribers lt count gt num subscribers lt count gt Have the publishing application wait for this number of subscribing applications to start Default 1 13 Test Parameters Table 8 2 Test Parameters Only for Publishing Applications Command Line Option Entry in perftest ini Description pidMultiPubTest lt id gt Not applicable in ini file Set the ID of the publisher in a multi pub
9. listener callback or in a separate thread useReadThread via an object called a WaitSet The latter approach can be beneficial in that it decouples the operation of your application from the middleware so that your processing will not interfere with RTI Data Distribution Service s internal activities However it does introduce additional context switches into your data receive path When data is arriving at a high rate these context switches can adversely impact performance when they occur with each data sam ple To improve efficiency the test configuration parameters waitset event count lt count gt and waitset delay usec lt usec gt set in the ini configuration file allow you to process incoming data in groups based on the number of samples and or time rather than individually reducing the number of context switches Experiment with these values to optimize performance for your system For more information see these sections in the RTI Core Libraries and Utilities User s Manual Receive Threads Section 19 3 and Conditions and WaitSets Section 4 6 16 Test Parameters 8 6 8 7 How to Measure Latency for a Given Throughput If you want to measure the minimum latency for a given throughput you have to use the com mand line parameters sleep lt millisec gt spin lt count gt and batchSize lt bytes gt to experimen tally set the throughput level for a given test run For example suppose you want
10. to generate a graph of latency vs throughput for a packet size of 200 bytes and throughput rates of 1000 10K 20K 50K 100K 500K and Max messages per second For throughput rates under 1000 messages per second use sleep lt ms gt to throttle the publish ing application For example sleep 1 will produce a throughput of approximately 1000 mes sages second sleep 2 will produce a throughput of approximately 500 messages second For throughput rates higher than 1000 messages per second use spin lt spin count gt to cause the publishing application to busy wait between sends The lt spin count gt value needed to produce a given throughput must be experimentally determined and is highly dependent on processor performance For example spin 19000 may produce a message rate of 10000 messages second with a slow processor but a rate of 14000 messages second with a faster processor Use batching when you want to measure latency for throughput rates higher than the maximum rates of sending individual messages First determine the maximum throughput rate for the data size under test without batching omit batchSize lt bytes gt For example on a 1 Gigabyte network for a data size of 200 bytes the maximum throughput will be about 70 000 messages sec We will refer to this value as max_no_batch For all throughput rates less than max_no_batch e g 70 000 messages sec do not use batching as this will increase the l
11. IBRARY_PATH or Path for you assuming your NDDSHOME environment vari able is already set For example e Ifyou are using a Windows 32 bit architecture set RTI_PERFTEST_ARCH i86Win32jdk e Ifyou are using a Windows 64 bit architecture set RTI_PERFTEST_ARCH x64Win64jdk e If you are using the Linux i86Linux2 6gcc4 3 4jdk architecture setenv RTI_PERFTEST_ARCH i86Linux2 6gcc4 3 4jdk L Make sure java is in your path before running the java example run script _ The test uses XML and INI configuration files and locates these files based on their paths relative to the directory from which the test is run Therefore to use these configuration files e Change to the top level rtiperftest 5 1 0 directory For example cd opt rti rtiperftest 5 1 0 e Start the test applications You can start the publisher or subscribers first the order does not matter When selecting your optional parameters choose parame ters that allow the test to run for at least 5 seconds to get any kind of meaningful results The longer it runs the more accurate the results will be Ideally you should run the test for at least 1 minute C bin lt architecture gt Release perftest_cpp lt pub sub gt parameters or bin lt architecture gt Debug perftest cpp lt pub sub gt parameters C Test Parameters bin lt architecture gt Release perftest_cs lt pub sub gt parameters or bin lt architecture gt Debug perftest_cs lt pub sub
12. RTI Connext DDS Combined Latency and Throughput Performance Test Getting Started Guide Version 5 1 0 1 f Your systems Working as one 2013 Real Time Innovations Inc All rights reserved Printed in U S A First printing December 2013 Trademarks Real Time Innovations RTI DataBus and Connext are trademarks or registered trademarks of Real Time Innovations Inc All other trademarks used in this document are the property of their respective owners tid and Use Restrictions Dae tet of this publication may be reproduced stored in a retrieval system or transmitted in any form including electronic mechanical photocopy and facsimile without the o written permission of Real Time Innovations Inc The software described in this a is furnished under and subject to the RTI software license agreement The software may be used or copied only under the terms of the license agreement Technical Support Real Time Innovations Inc 232 E Java Drive Sunnyvale CA 94089 Phone 408 990 7444 Email support rti com Website https support rti com Contents 1 a AeA Q N 10 11 Release Not Sivissserescisicieieeiiisrisieereia iini aii i i a e a ae 1 Tele What AOM O a a a a a E a ara ee bn ie 1 LTA Ghangeto Defa lt Domain TD siri iisen isata a eaea aiaiai aa aa Aa 1 1 1 2 Change in Multicast Default Configuration sss sesssssserstsstsrttsssesstessttrttesstestessneestessten tes 1 1 1 3 Change
13. application publishes throughput data it also subscribes to the latency echoes I The subscribing applications subscribe to the throughput data in which the echo requests are embedded they also publish the latency echoes The publisher prints the latency test results the subscriber prints the throughput results Download Instructions Download the test from the RTI Community Portal http community rti com Search for Example Performance Test for RTI Connext If you need help with the download process contact support rti com Installation Instructions UNIX Based Systems Follow the instructions below You do not need administrator privileges All directory locations in this document are meant as examples only adjust them to suit your site 1 Make sure you have GNU s version of the tar utility gtar which handles long file names and the unzip utility gunzip On Linux systems tar generally is gtar and unzip is gun zip On Solaris systems these utilities are separate 2 Use gunzip and gtar to extract the distribution into a directory of your choice For exam ple Building the Application 5 2 gunzip rtiperftestdds510 tar gz gtar xvf rtiperftestdds510 tar The performance test files will be extracted into lt your extraction directory gt rtiperft est 5 1 0 3 Make sure the NDDSHOME variable is set to your RTI Connext installation directory for example opt rti ndds 5 1 0 4 Make sure the appro
14. atency Use batching to test for throughput rates higher than max_no_batch start by setting batchSize to a multiple of the data size For example if the data size is 200 bytes use batchSize 400 this will put 2 messages in each batch batchSize 800 4 per batch etc This will allow you to get throughput latency results for throughputs higher than the max_no_batch throughput rate such as J 100 000 J 200 000 J 500 000 J millions Note For larger data sizes 8000 bytes and higher batching often does not improve throughput at least for 1 Gigabyte networks Auto Tuning and Turbo Mode RTI Connext 5 1 0 includes two features that allow the middleware to auto tune the communications to achieve better performance These features are Auto Throttling and Turbo Mode For more information about both features refer to Sections 10 4 Auto Throttling for DataWriter Performance Experimental Feature and 6 5 2 4 Turbo Mode Automatically Adjusting the Number of Bytes in a Batch Experimental Feature in the RTI Core Libraries and Utilities User s Manual The performance test application includes two command line options to enable these features enableAutoThrottle and enableTurboMode With Auto Throttling the publisher automatically adjusts the writing rate based on the number of unacknowledged samples in the send queue to avoid blocking and provide the best latency throughput tradeoff With Turbo Mode the size of a batch is automatica
15. com kb 811003 for more information 3 Change this entry HKEY_LOCAL_MACHINE SYSTEM CurrentControlSet Services AFD Parameters a Add the DWORD key FastSendDatagramThreshold b Set the value to 65536 0x10000 See http support microsoft com kb 235257 for more information 4 Reboot your machine for the changes to take effect 21
16. cssesseeseceeceeecesceeceeceseessecseeaecssesaeenees 15 8 3 Number of Iterations vs Latency Count ses seessssisstsssterttssssertttsstesttnterstsstesttenteestesntesstenteesten tt 16 84o Warming Upixiicnis tines aiid eaea aane ea a EE na aE Er aE E ta aeina Stereos Ee e ai ta 16 8 5 WaitSet Event Count and Delay cessassion ataks s iS 16 8 6 How to Measure Latency for a Given Throughput ssee ssssessissssssrtssssssstsrtesstestinsneentesteentesnten tes 17 8 7 Auto Tuning arid Turbo Mode 0 iccidecci ces sess scetsisectecscbecbeh statsiasshscudicsberviveusasersesebsucsevtvensvsussescvtervetees 17 Example Command Lines for Running the Performance Test ssssssssssssssssessssessssssessesssssseeeeeeees 18 9 1 1 1 Multicast Best Latency as a Function of Message Size 0 0 cscs cesses eesessssseeeseseseessesees 18 9 2 1 1 Multicast Maximum Throughput as a Function of Message Size with Batching 18 9 3 1 1 Multicast Latency vs Throughput for 200 byte Messages with Batching cscceee 19 94 1 to 1 Multicast Reliable UDPv4 All Sizes ccc cccccssccescesseecssceeesesseeecseceseecsseecaeceesesaecesseeees 19 9 5 1 to 1 Unicast Best Effort UDPV4 1 SiZe oo cccccscccssecssseescessseecseceeeceseeesseceseessecesseceseecseeeaeenses 19 9 6 1 to 1 Multicast Reliable UDPv4 Batching Enabled ccececeseececsssesesesescseesescseseseeseecseees 19 9 7 1 to 2 Multicast Reliable UDPV4 cccccccccccccssccssce
17. d Hat Enterprise Linux 5 0 2 6 kernel x64Linux2 6gcc4 1 1 x64 CentOS 5 4 5 5 2 6 kernel Red Hat Enterprise Linux 5 1 5 2 5 4 5 5 x64Linux2 6gcc4 1 2 2 6 kernel Visual Studio 2005 i86Win32VS2005 Visual Studio 2008 i86Win32VS2008 a Visual Studio 2010 i86Win32VS2010 A ae uae 2012 eon aae Visual Studio 2005 x64Win32VS2005 Visual Studio 2008 x64Win32VS2008 a Visual Studio 2010 x64Win32VS2010 Visual Studio 2012 x64Win32VS2012 Visual Studio 2005 i86Win32dotnet2 0 Visual Studio 2008 i86Win32dotnet2 0 i Visual Studio 2010 i86Win32dotnet4 0 Visual Studio 2012 i86Win32dotnet4 5 C on a Windows System i Visual Studio 2005 x64Win64dotnet2 0 a Visual Studio 2008 x64Win64dotnet2 0 Visual Studio 2010 x64Win64dotnet4 0 Visual Studio 2012 x64Win64dotnet4 5 Overview 5 1 Overview The publishing side of the test writes data as fast as it can Every few samples configured through the command line it sends a special sample requesting an echo from the subscribing side It uses this request gt echo exchange to measure round trip latency echo requests RTI PerfTest M E oa f echo replies Publishers Subscribers As you will see in Section 8 there are several command line options including ones to desig nate whether the application will act as the publisher or subscriber You will start multiple copies of the application typically 1 publisher and 1 or more subscrib ers L The publishing
18. dMultiSubTest lt id gt Not applicable in ini file ID of the subscriber in a multi subscriber test Use a unique value for each subscriber running on the same host that uses the same domain ID Default 0 Range 0 to n 1 inclusive where n is the number of subscribers in a multi subscriber test sub Not applicable in ini file Set test to be a subscriber Default sub 11 Test Parameters Table 8 2 Test Parameters Only for Publishing Applications Command Line Option Entry in perftest ini Description Enables the Auto Throttling feature hats See Auto Tuning and Turbo Mode enableAutoThrottle Not applicable in ini file Section 8 7 Default feature is disabled Enables the Turbo Mode feature See Auto Tuning and Turbo Mode Section 8 7 When turbo mode is enabled batchSize lt bytes gt is ignored Default feature is disabled enableTurboMode Not applicable in ini file Enable batching and set the maximum batched message size Default 0 batching disabled Range 1 to 63 000 For more information on batching data for high throughput see the batchSize lt bytes gt batch size lt bytes gt High Throughput design pattern in the RTI Core Libraries and Utilities Getting Started Guide See also How to Measure Latency for a Given Throughput Section 8 6 and the BATCH QosPolicy Section 6 5 2 in the RTI Core Libraries and Utilities User s Manual
19. do not specify pub or sub sub is assumed For additional information on setting the parameters see I Spinning vs Sleeping Section 8 1 LJ Send Queue Size and Queue Full Behavior Section 8 2 J Number of Iterations vs Latency Count Section 8 3 L Warming Up Section 8 4 LJ WaitSet Event Count and Delay Section 8 5 Test Parameters Table 8 1 Test Parameters for Publishing and Subscribing Applications Command Line Option Entry in perftest ini Description bestEffort is reliable lt true false gt Use best effort communication Default true use reliable communication For an introduction to the RTI reliability model see the Strict Reliability design pattern in the RTI Core Libraries and Utilities Getting Started Guide See also Reliable Communications in the RTI Core Libraries and Utilities User s Manual configFile lt filename gt Not applicable in ini file Path to a configuration file that contains command line options Default perftest ini dataLen lt bytes gt data length lt bytes gt Length of payload in bytes for each send Default 100 bytes Range 28 63 000 bytes The lower limit is the number of overhead bytes in the message i e the timestamp sequence number and other meta data used by the test the upper limit ensures that when the overhead of the wire protocol is added it doesn t overflow the UDP maximum datagram size of 64KB
20. e communication write will block until the subscriber s catch up 2 Slow the writing down by sleeping sleep lt millisec gt This approach is friendlier to the other processes on the host because it does not monopolize the CPU However context switching is expensive enough that you can t actually sleep for amounts of time on the order of microseconds so you could end up sleeping too long and hurting performance Operating systems including Linux and Windows have a minimum resolution for sleeping i e you can only sleep for a period of 1 or 10 ms If you specify a sleep period that is less than that minimum the OS may sleep for its minimum resolution 3 Spin in a tight loop between writes spin lt count gt This approach will add a pause with out giving up the CPU making it easier to sleep for very short periods of time In the test implementation there is a very short loop that just performs some simple math to take up CPU time The argument to spin lt count gt any number gt 0 is the number of times to go through that loop The default is 0 If you specify something else it should be a fairly large number 100 s or 1000 s since spinning the loop just a few times will take negligible time Avoid spinning on a single core machine as the code that would break you out of the spin may not be able to execute in a timely manner 4 Let the publisher automatically adjust the writing rate enableAutoThrottle This o
21. gt parameters Java scripts perftest_java lt pub sub gt parameters where lt architecture gt depends on your architecture such as i86Linux2 6gcc4 3 4 or i86Win32VS2008 l After the publisher recognizes that the specified number of subscribers see the numSubscribers lt count gt option are online and the subscriber recognizes that the specified number of publishers see the numPublishers lt count gt option are online the test begins Test Parameters Several parameters are available you can enter them on the command line or put them in a configuration ini file Entries on the command line take precedence See the configFile lt filename gt and qosprofile lt filename gt options in Table 8 1 for file locations All parameters are optional and case insensitive partial matches are allowed such as h instead of help Some parameters only make sense in the publishing or subscribing application The parameters are presented in the following tables based on whether they may be used in a publishing application a subscribing application or both I Test Parameters for Publishing and Subscribing Applications Table 8 1 on Page 8 I Test Parameters Only for Publishing Applications Table 8 2 on Page 12 L Test Parameters Only for Subscribing Applications Table 8 3 on Page 11 As you will see in the tables the pub parameter specifies a publishing application and sub specifies a subscribing application If you
22. in perftest xml as well as Config uring QoS with XML Chapter 17 in the RTI Core Libraries and Utilities User s Manual 10 Test Parameters Table 8 1 Test Parameters for Publishing and Subscribing Applications Command Line Option Entry in perftest ini Description useReadThread use read thread lt true false gt Use a separate thread instead of a callback to read data See WaitSet Event Count and Delay Section 8 5 Default use callback for subscriber waitsetDelayUsec lt usec gt waitset delay usec lt usec gt Process incoming data in groups based on time rather than individually Only used if the useReadThread option is specified on the subscriber side See WaitSet Event Count and Delay Section 8 5 Default 100 Range gt 0 waitsetEventCount lt count gt waitset event count lt count gt Process incoming data in groups based on the number of samples rather than individually Only used if the useReadThread option is specified on the subscriber side See WaitSet Event Count and Delay Section 8 5 Default 5 Range gt 1 Table 8 3 Test Parameters Only for Subscribing Applications Command Line Option Entry in perftest ini Description The subscribing application will i wait for this number of publishing numPublishers lt count gt num publishers lt count gt applications to start Default 1 si
23. lisher test Use a unique value for each publisher running on the same host that uses the same domain ID Default 0 Range 0 to n 1 inclusive where n is the number of publishers in a multi publisher test Set test to be a publisher sendQueueSize lt number gt send queue size lt number gt pub Not applicable in ini file oo PP t Default sub Run test in scan mode traversing a range of sample data sizes from 32 to 63 000 bytes scan scan lt true false gt If you set scan true you cannot set numlter lt count gt Default false no scan Size of the send queue When batchSize lt bytes gt is used the size is the number of batches See Send Queue Size and Queue Full Behavior Section 8 2 Default 50 Range 1 100 million or 1 indicating an unlimited length sleep lt millisec gt sleep millisec lt millisec gt Time to sleep between each send See Spinning vs Sleeping Section 8 1 Default 0 Range 0 or higher spin lt count gt spin loop count lt count gt Number of times to run in a spin loop between each send See Spinning vs Sleeping Section 8 1 Default 0 Range 0 or higher 14 Test Parameters 8 1 8 2 Spinning vs Sleeping When the publisher is writing as fast as it can sooner or later it is likely to get ahead of the sub scriber There are 4 things you can do in this case 1 Nothing for reliabl
24. lly adjusted to provide the best latency for a given write rate For slow write rates the batch size will be smaller to minimize the latency pen Example Command Lines for Running the Performance Test alty For high write rates the batch size will be bigger to increase throughput When turbo mode is used the command line option batchSize lt bytes gt is ignored To achieve the best latency under maximum throughput conditions use enableAutoThrottle and enableTurboMode in combination 9 Example Command Lines for Running the Performance Test The followings are examples of how to run the performance test for different use cases I The tests below print final results only if you want to see intermediate values remove the noprint argument from the command line LI If you are running on 2 unequal machines i e one machine is faster has better processors than another you will see better performance by running the Publisher on the slower machine l To measure CPU usage while running these tests use top or a similar utility 9 1 1 1 Multicast Best Latency as a Function of Message Size Publisher bin lt arch gt Release perftest_cpp pub noPrint nic lt ipaddr gt domain lt ID gt numIter lt count gt latencyCount 1 dataLen lt length gt latencyTest multicast Subscriber bin lt arch gt Release perftest_cpp sub noPrint nic lt ipaddr gt domain lt ID gt multicast Modify dataLen lt bytes g
25. lse gt Default TCP transport not enabled help Not applicable in ini file Print this usage message and exit instanceHashBuckets lt n gt instance hash buckets lt n gt Number of hash buckets for instances Default 1 means same as the number of instances Range actual value gt 0 instances lt int gt instances lt number of instances gt Set the number of instances to use in the test The publishing and subscribing applications must specify the same number of instances This option only makes sense when testing a keyed data type to do so uncomment the line define PERFTEST_KEYED_TYPE in the IDL type definition file lt install dir gt idl test idl before compiling Alternatively define the RTIDDSGEN_PREPROCESSOR environment variable to D PERFTEST_KEYED_TYPE before compiling Default 1 Range gt 0 keepDurationUsec lt usec gt keep duration usec lt usec gt Minimum duration that a sample is queued for ACK disabled readers Only used if the noPositiveAcks option is specified on the publisher side See Disabling Positive Acknowledgements Section 6 5 3 3 in the RTI Core Libraries and Utilities User s Manual Default 1 000 psec 1 millisec Range gt 0 multicast is multicast lt true false gt Use multicast to receive data Default do not use multicast multicastAddress lt address gt Currently not supported in perftest ini
26. net to yourself The test is designed to measure latency under loaded network condi Starting the Test tions it will produce those loads itself there is no need to produce them externally and your throughput results may not be meaningful if you do L The C and C executables are in these directories lt installation directory gt bin lt architecture gt Release lt installation directory gt bin lt architecture gt Debug where lt architecture gt depends on your architecture such as i86Linux2 6gcc4 3 4 or i86Win32VS2008 On Windows systems As described in Section 5 2 set your Path environment variable to the appropriate NDDSHOME lib lt architecture gt directory this allows a shared library to be picked up when you run the executable Or copy the DLLs from NDDSHOME lib lt architecture gt into the directory where the perftest_ lt language gt executable is located once you ve built the test The Java script that is used to run the Java example is in the lt installation directory gt scripts directory As described in Installation Instructions Section 5 you should have already set your LD_LIBRARY_PATH for UNIX based systems or Path for Windows systems so that the application can pick up the native shared libraries on which RTI Data Distribution Ser vice depends If you have not you can set the environment variable RTI_PERFTEST_ARCH to your specific architecture and let the run script set LD_L
27. number gt spin lt count gt multicast Subscriber bin lt arch gt Release perftest_cpp sub noPrint nic lt ipaddr gt multicast I Set numlter lt count gt to be in the millions for high throughput tests reduce as needed for lower throughputs otherwise the tests will take longer and longer to complete I To adjust throughput experiment with the value of spin lt count gt For example to get a rate of 10 000 messages sec use spin 20000 to see the resulting rate then adjust up or down as needed 1 to 1 Multicast Reliable UDPv4 All Sizes Publisher bin lt arch gt Release perftest_cpp pub noPrint sendQueueSize 32 latencyCount 10000 scan multicast Subscriber bin lt arch gt Release perftest cpp sub noPrint multicast 1 to 1 Unicast Best Effort UDPv4 1 Size Publisher bin lt arch gt Release perftest_cpp pub noPrint sendQueueSize 32 latencyCount 1000 numIter 1000000 dataLen 1024 bestEffort Subscriber bin lt arch gt Release perftest_ cpp sub noPrint dataLen 1024 bestEffort 1 to 1 Multicast Reliable UDPv4 Batching Enabled Publisher bin lt arch gt Release perftest_cpp pub noPrint sendQueueSize 32 latencyCount 1000 numIter 1000000 dataLen 200 batchSize 6400 multicast Subscriber bin lt arch gt Release perftest_cpp sub noPrint dataLen 200 batchSize 6400 multicast 19 Example Output 9 7 1 to 2 Multicast Reliable UDPv4 Publisher
28. our architecture s file is not provided you can create your own project and solution files based on the shipped example projects and solution files 2 For better performance select the Mixed Platforms Release configuration in the Standard toolbar in Visual Studio 3 From the Build menu select Build Solution I Java on UNIX Based or Windows Systems You will need A build tool called Ant which is available free from http ant apache org e JDK 1 7 The latest JDK version can be found here http java sun com javase downloads index jsp e The JAVA_HOME environment variable must be set to the JDK installation direc tory because Ant uses this variable Enter cd rtiperftest 5 1 0 perftest_java ant propertyfile resource properties dds_release properties Starting the Test The test is provided in C C and Java The list below identifies how to run the executables once you have built them and how to pass configuration parameters to them For detailed descriptions of the test parameters see Test Parameters Section 8 For example test configura tions see Example Command Lines for Running the Performance Test Section 9 When running the test keep in mind that a throughput test will necessarily place a heavy load on your network and potentially on your CPU s as well For the most accurate results and the fewest complaints from your coworkers and IT department run the test when and where you have a sub
29. priate NDDSHOME lib lt architecture gt directory is in the LD_LIBRARY_PATH environment variable to pick up the shared library The test is provided in source code only after installation you will need to build the applica tion see Section 6 Windows Systems Follow the instructions below You do not need administrator privileges All directory locations in this document are meant as examples only adjust them to suit your site 1 Right click the distribution file and extract its contents into a directory of your choice The performance test files will be extracted into lt your extraction directory gt rtiperftest 5 1 0 2 Make sure the NDDSHOME variable is set to your RTI Connext installation directory for example C Program Files RTI ndds 5 1 0 3 Make sure the appropriate NDDSHOME lib lt architecture gt directory is in the Path environment variable so that it picks up the shared library Or you can copy the DLLs from NDDSHOME lib lt architecture gt into the directory where the perftest_ lt language gt executable is once you ve built the test see Building the Applica tion Section 6 The test is provided in source code only after installation you will need to build the applica tion see Building the Application Section 6 Building the Application First as described above make sure the environment variable NDDSHOME is set to the direc tory containing your RTI Connext installation To use
30. ption enables the Auto Throttle feature introduced in RTI Connext 5 1 0 and its usage is pre ferred over spin lt count gt because the amount of spin is automatically determined by the publisher based on the number of unacknowledged samples in the send queue See also Send Queue Size and Queue Full Behavior Section 8 2 Send Queue Size and Queue Full Behavior In many distributed systems a data producer will often outperform data consumers That means that if the communications are to be reliable the producer must be throttled in some way to allow the consumers to keep up In some situations this may not be a problem because data may simply not be ready for publication at a rate sufficient to overwhelm the subscribers If you re not so lucky your publisher s queue of unacknowledged data will eventually fill up When that happens if data is not to be lost the publication will have to block until space becomes available Blocking can cost you in terms of latency To avoid the cost of blocking consider the following L Enlarge your publisher s queue sendQueueSize lt number gt Doing so will mean your publisher has to block less often However it may also let the publisher get even further ahead of slower subscribers increasing the number of dropped and resent packets hurt ing throughput Experimenting with the send queue size is one of the easy things you can do to squeeze a little more throughput from your system LJ Enable
31. rom 59 to 1 Change in Multicast Default Configuration In previous releases multicast was enabled by default and could be disabled using the command line option nomulticast Starting with release 5 1 0 multicast is disabled by default and can be enabled using the command line option multicast The command line option nomulticast is kept for backward compatibility RTI Issue ID CORE 5854 Change in Functionality of Performance Test s enableSharedMemory Option In this release the effect of the command line option enableSharedMemory has been changed Now it forces the application to only use the Shared Memory transport In previous releases this option allowed the application to use Shared Memory as well as the default transport UDPv4 Supported Platforms RTI Issue ID COREPLG 5867 Reorganization of QoS Settings In this release the QoS settings that are not dependent on the command line options have been moved to the perftest xml file in order to ease QoS modification RTI Issue ID COREPLG 5919 Supported Platforms The performance test application provides makefiles projects for the following architectures Red Hat Enterprise Linux 4 0 2 6 kernel i86Linux2 6gcc3 4 3 xo Red Hat Enterprise Linux 5 0 2 6 kernel i86Linux2 6gcc4 1 1 Red Hat Enterprise Linux 4 0 2 6 kernel x64Linux2 6gcc3 4 5 C on a UNIX based System Re
32. ssccssscesseceseecesseessceseecssseesseceseecsscesaeceeeecesesseesess 20 9 8 2 to 1 Multicast Reliable UDPV4 oo cccccccscccsscessscssesesseceseecsseeessceseecesseesaeceseecssceeaeceeeeceseseesees 20 Example Outpt ere eeose onneen ree Seea o ie nasaan ra Neri EAE Eea E EEEE ARSEN Eins 20 Optimizing Your OS for Network PerformancCe ssessessesreeseesessesrcoseesessesscoseesessroncoreeoressessoseesressessessesee 21 ILT Optimizing Linux System S siiiesissni iatis a aa aaa aia a 21 11 2 Optimizing Windows Systems ssns Berosa oiik e ias iar ieiet nais ap custecebse tistebseseabins 21 Testing Performance This document describes how to run a combined latency and throughput test application for RTI Connext The test measures what is sometimes called loaded latency latency at a given throughput level It can help you answer questions such as l For a given publishing configuration e g queue size batching settings sample size and subscribing configuration e g queue size Listener vs WaitSet what is the through put of my network L When my network is heavily loaded what latency can I expect l For a given configuration what is the best case latency with no other traffic on the net work For examples of specific configurations see Example Command Lines for Running the Perfor mance Test Section 9 Release Notes What s New in 5 1 0 Change to Default Domain ID The default domain ID has been changed f
33. t to see latencies for different data sizes I Set numiter lt count gt to be gt 100 for statistically better results 9 2 1 1 Multicast Maximum Throughput as a Function of Message Size with Batching Publisher bin lt arch gt Release perftest_cpp pub noPrint nic lt ipaddr gt numIter lt count gt dataLen lt length gt batchSize lt bytes gt sendQueueSize lt number gt multicast Subscriber bin lt arch gt Release perftest_cpp sub noprint nic lt ipaddr gt multicast I Set numlIter lt count gt to be millions for small message sizes lt 1k reduce as needed for larger sizes otherwise the tests will take longer and longer to complete L To achieve maximum throughput start by setting batchSize lt bytes gt to 6400 then increase the size to see if you get better throughput The largest valid batch size is 63000 bytes L For maximum throughput start by setting sendQueueSize lt number gt to 30 the best value will usually be between 30 50 18 Example Command Lines for Running the Performance Test 9 3 9 4 9 5 9 6 Note For larger data sizes 8000 bytes and higher batching often does not improve throughput at least for 1 Gig networks 1 1 Multicast Latency vs Throughput for 200 byte Messages with Batching Publisher bin lt arch gt Release perftest_cpp pub noPrint nic lt ipaddr gt numIter lt count gt dataLen 200 batchSize lt bytes gt sendQueueSize lt
34. the former The implicit assumption is that the latter is negligible by comparison If you violate this assumption your throughput test results will not be meaningful l Keep the number of iterations large enough to send many latency pings over the course of the test run That is keep numIter lt count gt small compared to latencyCount lt count gt Your latency measurements and the spread between them will be of higher quality if you are able to measure more data points L When selecting numiter lt count gt choose a value that allows the test to run for at least a minute to get accurate results Set numIter lt count gt to be millions for small message sizes lt 1k reduce as needed for larger sizes otherwise the tests will take longer and lon ger to complete Warming Up When running the performance test in Java and to a lesser extent C you may observe that throughput slowly increases through the first few incremental measurements and then levels off This improvement reflects the background activity of the just in time JIT compiler and optimizer on these platforms For the best indication of steady state performance be sure to run the test for a number of samples numlter lt count gt sufficient to smooth out this start up arti fact WaitSet Event Count and Delay RTI Connext and by extension this performance test gives you the option to either process received data in the middleware s receive thread via a
35. ue the default the subscribing application gets samples from the publishing application and RTI Persistence Service This mode provides low latency between endpoints If set to false the subscribing application only gets samples from RTI Persistence Service This brokered communication pattern provides a way to guarantee eventual consistency Default true direct communication noPositiveAcks use positive acks lt true false gt Disable use of positive ACKs in the reliable protocol Default true use positive ACKs See qosprofile lt filename gt option for more information noPrintIntervals print intervals lt true false gt Prevent printing of statistics at intervals during the test By default statistics are printed every second in the subscribing application and after receiving every latency echo in the publishing application qosprofile lt filename gt qos profile file lt filename gt Path to the XML file containing DDS QoS profiles Default perftest xml The default file contains these QoS profiles The ThroughputQos LatencyQos and AnnouncementQos profiles are used by default The NoAckThroughputQos and NoAck LatencyQos profiles are used if you spec ify noPositiveAcks Note some QoS values are hard coded in the application therefore setting them in the XML file has no effect see the Note on Page 15 See comments
36. us 99 99 1062 us Subscriber bin i86Linux2 6gcc3 4 3 Release perftest_cpp sub noPrint domain 27 Waiting to discover 1 publishers Waiting for data Length 32 Packets 10000000 Packets s ave 47913 Mbps ave 12 3 Lost 0 Length 64 Packets 10000000 Packets s ave 47580 Mbps ave 24 4 Lost 0 20 Optimizing Your OS for Network Performance 11 Optimizing Your OS for Network Performance The network stacks of popular operating systems are not always tuned for maximum perfor mance out of the box RTI has found that the following configuration changes frequently improve performance for a broad set of demanding applications Consider testing your network performance with and without these changes to learn if they can benefit your system 11 1 Optimizing Linux Systems Edit the file etc sysctl conf and add the following net core wmem_max 16777216 net core wmem_default 16777216 net core rmem_max 16777216 net core rmem_default 16777216 net ipv4 tcp_rmem 4096 16777216 33554432 net ipv4 tcp_wmem 4096 16777216 33554432 net ipv4 tcp mem 4096 16777216 33554432 run sbin sysctl p 11 2 Optimizing Windows Systems 1 From the Start button select Run then enter regedit 2 Change this entry HKEY_LOCAL_MACHINE SYSTEM CurrentControlSet Services Tcpip Parameters a Add the DWORD key MaximumReassemblyHeaders b Set the value to Oxffff this is the max value See http support microsoft
Download Pdf Manuals
Related Search