Performance Testers

From Second Life Wiki
Jump to navigation Jump to search

Metric-based Test Framework

The metric-based test framework can be used to analyze all sorts of data gathered by the viewer at run time. The base implementation offers a variety of services to save and load data, parse LLSD results, compare current performance with a baseline and produce test reports.

The interface is implemented in llcommon/llmetricperformancetester.h. Two abstract classes are available:

  • LLMetricPerformanceTesterBasic
  • LLMetricPerformanceTesterWithSession

LLMetricPerformanceTesterBasic

The abstract class LLMetricPerformanceTesterBasic defines the general metric-based test framework.

This class can be directly inherited from for simple data gathering and provides predefined methods to save, load and compare results of performance sessions.

Below is the detailed doc for this class.

<cpp> /**

* @class LLMetricPerformanceTesterBasic
* @brief Performance Metric Base Class
*/

class LL_COMMON_API LLMetricPerformanceTesterBasic { public:

   /**
    * @brief Creates a basic tester instance.
    * @param[in] name - Unique string identifying this tester instance.
    */

LLMetricPerformanceTesterBasic(std::string name); virtual ~LLMetricPerformanceTesterBasic();

   /**
    * @return Returns true if the instance has been added to the tester map.
    * Need to be tested after creation of a tester instance so to know if the tester is correctly handled.
    * A tester might not be added to the map if another tester with the same name already exists.
    */
   BOOL isValid() const { return mValidInstance; }
   /**
    * @brief Write a set of test results to the log LLSD.
    */

void outputTestResults() ;

   /**
    * @brief Compare the test results.
    * By default, compares the test results against the baseline one by one, item by item, 
    * in the increasing order of the LLSD record counter, starting from the first one.
    */

virtual void analyzePerformance(std::ofstream* os, LLSD* base, LLSD* current) ;

   /**
    * @return Returns the number of the test metrics in this tester instance.
    */

S32 getNumberOfMetrics() const { return mMetricStrings.size() ;}

   /**
    * @return Returns the metric name at index
    * @param[in] index - Index on the list of metrics managed by this tester instance.
    */

std::string getMetricName(S32 index) const { return mMetricStrings[index] ;}

protected:

   /**
    * @return Returns the name of this tester instance.
    */

std::string getTesterName() const { return mName ;}

   /**
    * @brief Insert a new metric to be managed by this tester instance.
    * @param[in] str - Unique string identifying the new metric.
    */

void addMetric(std::string str) ;

   /**
    * @brief Compare test results, provided in 2 flavors: compare integers and compare floats.
    * @param[out] os - Formatted output string holding the compared values.
    * @param[in] metric_string - Name of the metric.
    * @param[in] v_base - Base value of the metric.
    * @param[in] v_current - Current value of the metric.
    */

virtual void compareTestResults(std::ofstream* os, std::string metric_string, S32 v_base, S32 v_current) ; virtual void compareTestResults(std::ofstream* os, std::string metric_string, F32 v_base, F32 v_current) ;

   /**
    * @brief Reset internal record count. Count starts with 1.
    */

void resetCurrentCount() { mCount = 1; }

   /**
    * @brief Increment internal record count.
    */

void incrementCurrentCount() { mCount++; }

   /**
    * @return Returns the label to be used for the current count. It's "TesterName"-"Count".
    */
   std::string getCurrentLabelName() const { return llformat("%s-%d", mName.c_str(), mCount) ;}
   
   /**
    * @brief Write a test record to the LLSD. Implementers need to overload this method.
    * @param[out] sd - The LLSD record to store metric data into.
    */

virtual void outputTestRecord(LLSD* sd) = 0 ;

private: void preOutputTestResults(LLSD* sd) ; void postOutputTestResults(LLSD* sd) ;

std::string mName ; // Name of this tester instance S32 mCount ; // Current record count

   BOOL mValidInstance;                            // TRUE if the instance is managed by the map

std::vector< std::string > mMetricStrings ; // Metrics strings

// Static members managing the collection of testers public:

   // Map of all the tester instances in use

typedef std::map< std::string, LLMetricPerformanceTesterBasic* > name_tester_map_t; static name_tester_map_t sTesterMap ;

   /**
    * @return Returns a pointer to the tester
    * @param[in] name - Name of the tester instance queried.
    */

static LLMetricPerformanceTesterBasic* getTester(std::string name) ;

   /**
    * @return Returns TRUE if there's a tester defined, FALSE otherwise.
    */

static BOOL hasMetricPerformanceTesters() { return !sTesterMap.empty() ;}

   /**
    * @brief Delete all testers and reset the tester map
    */

static void cleanClass() ;

private:

   // Add a tester to the map. Returns false if adding fails.

static BOOL addTester(LLMetricPerformanceTesterBasic* tester) ; }; </cpp>

LLMetricPerformanceTesterWithSession

The abstract class LLMetricPerformanceTesterWithSession is derived from the previous one and provides an additional abstraction that allows the definition of ad-hoc comparison methods for reporting. This class should be used when data need to be collated and analyzed in specific ways.

Below is the detailed doc for this class.

<cpp> /**

* @class LLMetricPerformanceTesterWithSession
* @brief Performance Metric Class with custom session 
*/

class LL_COMMON_API LLMetricPerformanceTesterWithSession : public LLMetricPerformanceTesterBasic { public:

   /**
    * @param[in] name - Unique string identifying this tester instance.
    */

LLMetricPerformanceTesterWithSession(std::string name); virtual ~LLMetricPerformanceTesterWithSession();

   /**
    * @brief Compare the test results.
    * This will be loading the base and current sessions and compare them using the virtual 
    * abstract methods loadTestSession() and compareTestSessions()
    */

virtual void analyzePerformance(std::ofstream* os, LLSD* base, LLSD* current) ;

protected:

   /**
    * @class LLMetricPerformanceTesterWithSession::LLTestSession
    * @brief Defines an interface for the two abstract virtual functions loadTestSession() and compareTestSessions()
    */

class LLTestSession

       {
       public:
           virtual ~LLTestSession() ;
       };
   
   /**
    * @brief Convert an LLSD log into a test session.
    * @param[in] log - The LLSD record
    * @return Returns the record as a test session
    */

virtual LLMetricPerformanceTesterWithSession::LLTestSession* loadTestSession(LLSD* log) = 0;

   /**
    * @brief Compare the base session and the target session. Assumes base and current sessions have been loaded.
    * @param[out] os - The comparison result as a standard stream
    */

virtual void compareTestSessions(std::ofstream* os) = 0;

LLTestSession* mBaseSessionp; LLTestSession* mCurrentSessionp; }; </cpp>

Creating/Adding a Basic Test Metrics

First, you need to create a tester class derived from the LLMetricPerformanceTesterBasic class that will hold your performance data. The key steps are:

  • declare your own tester derived from LLMetricPerformanceTesterBasic
  • in the constructor, declare all metrics you will use in this tester, the declaration order does not matter
  • collect the test data in your own way. The usual way is to define an update() method that gets called and gather the relevant performance data.
  • define the abstract virtual method outputTestRecord(LLSD* sd) to output your test data to the LLSD structure. Everything output to the LLSD in this function will be saved to a log file in the log folder.
  • the final test report contains the following columns: metric_string, baseline_value, target_value, (target_value - baseline_value) and (100 * target_value / baseline_value).

Below is a complete code example implementation.

<cpp> class YourOwnTester : public LLMetricPerformanceTesterBasic { public:

   YourOwnTester() ;
   ~YourOwnTester() ;
   // This will have to get called in code to update your perf data.
   // Note: you can create as many updateXx() variation as your perf system requires
   void update(const S32 d1, const F32 d2) ;

protected:

   // This is required. It tells the class how to pack the data in an LLSD stream
   /*virtual*/ void outputTestRecord(LLSD* sd) ;

private:

   // Define the relevant perf gathering variables. 
   // Note: the default compare method only supports S32 and F32 comparison. You need to overload the compare if you need to carry something else.
   S32 data1;
   F32 data2;
   ...

};

// Note : the tester name "yourtester" is important to remember as it is the name you will use on the command line // when launching the viewer in perf metric gathering mode. YourOwnTester::YourOwnTester() : LLMetricPerformanceTesterBasic("yourtester") {

   // Declare all the metrics used in the tester.
   addMetric("metric-string-1") ;
   addMetric("metric-string-2") ;
   ...
   // Your own initializations
   data1 = 0;
   data2 = 0.0f;
   ...

}

YourOwnTester::~YourOwnTester() {

   // You likely need to invalidate the static pointer holding that test instance
   sYourTester = NULL;

}

void YourOwnTester::outputTestRecord(LLSd *sd) {

   std::string currentLabel = getCurrentLabelName();
   //insert your own code to output test results to sd
   //format like this
   (*sd)[currentLabel]["metric-string-1"] = (LLSD::Integer)data1;
   (*sd)[currentLabel]["metric-string-2"] = (LLSD::Real)data2;
   ...

}

void YourOwnTester::update(const S32 d1, const F32 d2) {

   // Do something with the input data to update your perf data
   data1 += d1;
   data2 += d2;
   ...
   // *Important* You need to call outputTestResults() when some perf gathering condition is met
   // Otherwise your data might not be saved to the log ever.
   if (condition)
   {
       outputTestResults();
   }

}

</cpp>

You may check the class LLImageCompressionTester as an example.

How To Run Metric-based Automated Tests

Performance metric gathering must be done in 2 main steps:

  • Baseline performance data acquisition
  • Target run and analysis

For performance data gathering to be performed, the viewer must be run with special command line arguments. See Client parameters to learn how to do this and for more information on all command line arguments.

Baseline Performance Data Acquisition

During that step, simply run the viewer so to gather initial data. You'll use that run to create a yourtester_baseline.slp file that will be used later to compare new runs with this baseline. If you are for instance on the verge of doing modification to improve the performance of some part of the viewer, this is something you want to do before doing any modification to the code so you can later measure how much your modification improved (or not) the performance measured.

Note that every time you add new performance metric, you need to re-create a baseline. It is therefore advisable to develop performance metric code in a non-modified repository and pull the performance metric code to the modified repository while developing new algorithms. It is actually not uncommon to get ideas of new performance data to gather while doing performance tuning development.

Steps:

  • Make sure the log folder does not contain any previously gathered data, i.e. delete or backup the files yourtester.slp and yourtester_baseline.slp if any
  • Launch the viewer from the command line with --logmetrics yourtester as an argument. It is required to use a tester parameter here. If the argument passed does not correspond to anything, the resulting log yourtester.slp will simply be empty. If you want to gather all performance metrics defined in the code as one, you can use metric as a parameter. Note though that the viewer is extremely slow and hard to use under this circumstance.
  • Perform the test session. It is advisable to repeat a similar scenario between runs though, of course, this really depends about what you want to measure and compare.
  • Quit the viewer: at that point, you should have a new yourtester.slp file in the log folder. Note that quitting might take some time as writing the yourtester.slp file is time consuming. Be patient. Do not force quit the viewer during that stage!
  • Rename the file yourtester.slp to yourtester_baseline.slp and leave it in the log folder
Target Performance Data Acquisition and Analysis

You can perform target runs and analysis using the same baseline over and over. The basic data are saved in a yourtester.slp file and the data comparison is stored in a yourtester_report.csv file that can be open with Excel or Open Office.

Note that the code overwrites old yourtester.slp and yourtester_report.csv files so rename or move your reports if you need to.

Steps:

  • Make sure the log folder contains a yourtester_baseline.slp file. The perf analysis will fail if that file is not present and no yourtester_report.csv file will be created.
  • Launch the viewer from the command line with --logmetrics yourtester --analyzeperformance as arguments
  • Perform the test session
  • Quit the viewer: you can find the test results in the file yourtester_report.csv located in your secondlife log file folder.
More

If one wants to gather data automatically, it's possible to use --autologin and --replaysession on the command line.

Published Testers and Results

TextureTester

To be Completed

ImageCompressionTester

This tester is used to measure compression/decompression performance of JPEG2000 images.

It's triggered specifying the ImageCompressionTester performance tracking on the command line, i.e.:

  • Mac: ./Second\ Life --logmetrics ImageCompressionTester --analyzeperformance
  • Windows: create a shortcut to the executable, edit its properties and add --logmetrics ImageCompressionTester --analyzeperformance to the Target field in the Shortcut tab.
  • Linux: ./secondlife --logmetrics ImageCompressionTester --analyzeperformance

Baseline and Test Runs

For this series of tests:

  • The baseline was created using Kakadu v4.2.1 from the Snowstorm 2.5 216083 build.
  • The openjpeg v1.3.0 data were created using that same build minus the dynamic kdu library (no build required with the old dynamic linking strategy).
  • The Kakadu v6.4.1 data were created using the Snowstorm 2.6 219259 build.

When comparing various decompression implementations, you must create your own baseline on your own machine so the comparison is meaningful.

Results

Windows

Base = Kakadu v4.2.1 vs. Target = Kakadu v6.4.1

<tab class=lltable head=top border=1> Metric Base(B) Target(T) Diff(T-B) Percentage(100*T/B) Time Decompression (s) 32.7087 29.1120 -3.5968 89.0037 Volume In Decompression (kB) 69416.0313 69254.7656 -161.2656 99.7677 Volume Out Decompression (kB) 639489.0625 654995.5000 15506.4375 102.4248 Decompression Ratio (x:1) 9.2124 9.4578 0.2454 102.6633 Perf Decompression (kB/s) 2122.2468 2378.9089 256.6621 112.0939 Time Compression (s) 0.3180 0.3855 0.0675 121.2389 Volume In Compression (kB) 3145.7280 3145.7280 0.0000 100.0000 Volume Out Compression (kB) 393.0400 392.8800 -0.1600 99.9593 Compression Ratio (x:1) 8.0036 8.0068 0.0033 100.0407 Perf Compression (kB/s) 9892.3506 8159.3887 -1732.9619 82.4818 </tab>

Base = Kakadu v4.2.1 vs. Target = openjpeg v1.3.0

<tab class=lltable head=top border=1> Metric Base(B) Target(T) Diff(T-B) Percentage(100*T/B) Time Decompression (s) 31.6939 134.4571 102.7632 424.2364 Volume In Decompression (kB) 67361.5781 67321.7813 -39.7969 99.9409 Volume Out Decompression (kB) 620220.8750 640426.3125 20205.4375 103.2578 Decompression Ratio (x:1) 9.2073 9.5129 0.3056 103.3188 Perf Decompression (kB/s) 2125.3792 500.6934 -1624.6858 23.5578 Time Compression (s) 0.3180 1.5392 1.2212 484.0213 Volume In Compression (kB) 3145.7280 3145.7280 0.0000 100.0000 Volume Out Compression (kB) 393.0400 314.1560 -78.8840 79.9298 Compression Ratio (x:1) 8.0036 10.0133 2.0097 125.1098 Perf Compression (kB/s) 9892.3506 2043.7842 -7848.5664 20.6602 </tab>

Mac

Base = Kakadu v4.2.1 vs. Target = Kakadu v6.4.1

<tab class=lltable head=top border=1> Metric Base(B) Target(T) Diff(T-B) Percentage(100*T/B) Time Decompression (s) 60.0396 47.2449 -12.7947 78.6896 Volume In Decompression (kB) 64200.5156 64484.8047 284.2891 100.4428 Volume Out Decompression (kB) 617027.9375 626571.5000 9543.5625 101.5467 Decompression Ratio (x:1) 9.6109 9.7166 0.1056 101.0990 Perf Decompression (kB/s) 1069.3031 1364.9049 295.6018 127.6443 Time Compression (s) 0.5290 0.3624 -0.1666 68.5154 Volume In Compression (kB) 3932.1599 3145.7280 -786.4319 80.0000 Volume Out Compression (kB) 490.9610 393.0360 -97.9250 80.0544 Compression Ratio (x:1) 8.0091 8.0037 -0.0054 99.9320 Perf Compression (kB/s) 7433.1387 8679.0918 1245.9531 116.7621 </tab>

Base = Kakadu v4.2.1 vs. Target = openjpeg v1.3.0

<tab class=lltable head=top border=1> Metric Base(B) Target(T) Diff(T-B) Percentage(100*T/B) Time Decompression (s) 31.6939 134.4571 102.7632 424.2364 Time Decompression (s) 60.0396 215.3034 155.2638 358.6024 Volume In Decompression (kB) 64200.5156 65215.3086 1014.7930 101.5807 Volume Out Decompression (kB) 617027.9375 671115.9375 54088.0000 108.7659 Decompression Ratio (x:1) 9.6109 10.2908 0.6798 107.0734 Perf Decompression (kB/s) 1069.3031 302.8996 -766.4036 28.3268 Time Compression (s) 0.5290 0.5691 0.0401 107.5776 Volume In Compression (kB) 3932.1599 1572.8640 -2359.2959 40.0000 Volume Out Compression (kB) 490.9610 157.2180 -333.7430 32.0225 Compression Ratio (x:1) 8.0091 10.0044 1.9952 124.9122 Perf Compression (kB/s) 7433.1387 2763.8230 -4669.3154 37.1824 </tab>

Linux

Analysis

Looking at those, it appears that openjpeg v1.3.0 is 4 times slower than kdu v4.2.1 on decompression and 3 times slower on compression. Using kdu as a dynamic library or statically linked has no significant influence on performance.

Links

  • Performance Tracking: Describes the system that can be used to track general rendering performance within the viewer